 P e t e r H o e f s l o o t , A m o r I n e s , J o s v a n D a m , G r e g o r y D u v e i l l e r , F r a n c o i s K a y i t a k i r e a n d J a m e s H a n s e n 2 0 1 2                                                                           Combining  Crop  Models  and  Remote  Sensing  for   Yield  Prediction:  Concepts,  Applications  and   Challenges  for  Heterogeneous  Smallholder   Environments     Report  of  Joint  CCFAS-­‐JRC  Workshop   Venue:  Joint  Research  Centre  (JRC),  Ispra,  Italy   Date:  June  13-­‐14,  2012     Peter  Hoefsloot,  Hoefsloot  Spatial  Solutions,  The  Netherlands   Amor  Ines,  IRI-­‐Columbia  University,  NY,  USA   Jos  van  Dam,  Wageningen  University,  The  Netherlands   Gregory  Duveiller,  Joint  Research  Centre,  Ispra,  Italy   Francois  Kayitakire,  Joint  Research  Centre,  Ispra,  Italy   James  Hansen,  CCAFS  &  IRI-­‐Columbia  University,  NY,  USA       November  2012   2         3       Contents   1   Executive  summary  ...............................................................................................................................  5   1.1   Context  ..........................................................................................................................................  5   1.2   Workshop  rationale  and  objectives  ..............................................................................................  5   1.3   Workshop  salient  findings.............................................................................................................  6   2   Summary  of  presentations  and  discussions  ..........................................................................................  9   2.1   Data  assimilation  techniques  ........................................................................................................  9   2.2   Crops  researched  ........................................................................................................................  10   2.3   Use  of  Crop  Models  .....................................................................................................................  11   2.4   Use  of  Remote  Sensing  parameters  as  proxies  for  biomass  production  ....................................  13   2.5   Use  of  satellite  sensors  ...............................................................................................................  15   2.6   Research  locations  ......................................................................................................................  16   2.7   Spatial  scales  ...............................................................................................................................  17   2.8   Heterogeneity  .............................................................................................................................  17   2.9   Crop  masks  ..................................................................................................................................  18   2.10   Crop  management  factors  ..........................................................................................................  18   2.11   Uncertainty  of  predictions  ..........................................................................................................  19   2.12   Linkage  with  other  sources  of  information  .................................................................................  19   3   Presentation  Abstracts  ........................................................................................................................  21   3.1   The  challenges  of  an  operational  crop  yield  forecasting  system  in  Sub-­‐Saharan  Africa  ............  21   3.2   Crop  Forecasting  within  the  CCAFS  Program  ..............................................................................  23   3.3   Integration  of  agro-­‐hydrological  modelling,  remote  sensing  and  geographical  information  ....  24   3.4   Assimilating  remote  sensing  data  into  crop  models  improves  predictive  performance  for   spatial  application  ...................................................................................................................................  25   3.5   Regional  Crop  Simulation  Modelling  for  Yield  Prediction  Using  Remote  Sensing  and  GIS:  Indian   Experiences  .............................................................................................................................................  26   3.6   Integration  of  MODIS  products  and  a  crop  simulation  model  for  crop  yield  estimation  ...........  27   3.7   Exploring  the  Response  of  the  Central  US  Agro-­‐Ecosystem  to  Climate  Change  .........................  28   3.8   Crop  Yield  Forecasting  Over  Various  Scales  Combining  Models  and  Remote  Sensing  ...............  29   3.9   On  the  Assimilation  of  Remote  Sensing  Data  with  Crop  Models  for  Crop  Yield  Forecasting  .....  30   4     3.10   Simultaneous  Estimation  of  Model  State  Variables  and  Observation  and  Forecast  Biases  using  a   Two-­‐Stage  Hybrid  Kalman  Filter  .............................................................................................................  31   3.11   Satellite  image  simulations  for  data  assimilation  at  multiple  scales  ..........................................  32   3.12   MARS  operational  crop  monitoring  and  yield  forecasting  activities  in  Europe  ..........................  33   3.13   Experiences  with  data  assimilation  for  regional  crop  yield  forecasting  .....................................  34   3.14   Crop  Monitoring  and  Early  Warning  Service  in  Africa  ................................................................  35   3.15   Data  Assimilation  based  on  the  Integration  of  Satellite  Data  and  Field  Sensor  Data  for  Drought   Monitoring  ..............................................................................................................................................  36   3.16   Data  assimilation  for  the  carbon  cycle  in  Sudan  savannah  smallholder  communities  ...............  37   3.17   Soil-­‐water-­‐crop  modelling  for  decision  support  in  Sub-­‐Saharan  west  Africa:  experiences  from   Niger  and  Benin  .......................................................................................................................................  38   3.18   Wheat  yield  modelling  in  a  stochastic  framework  within  and  post  season  yield  estimation  in   Tunisia  39   4   References  ..........................................................................................................................................  41   5   Acronyms  and  Abbreviations  ..............................................................................................................  43   6   Workshop  Program  .............................................................................................................................  45   7   Participants  .........................................................................................................................................  47   8   Sponsors  ..............................................................................................................................................  48         5     1 Executive  summary   1.1 Context     dŚĞƌĞĂƌĞŵĂŶLJĨĂĐƚŽƌƐĐŽŶƚƌŝďƵƚŝŶŐƚŽƚŚĞƐƚƌĂŝŶŽŶƚŚĞǁŽƌůĚ͛ƐĨŽŽĚƐƵƉƉůLJ͕ƌĂŶŐŝŶŐĨƌŽŵŝŶƐƵĨĨŝĐŝĞŶƚ investment  in  the  agricultural  sector,  a  lack  of  access  to  markets,  climate  change  and  climate  variability,   water  shortages  and  drought,  to  simple  increased  demand  for  food  and  changes  in  diet.     It  is  a  major  challenge  for  the  world  to  feed  its  growing  population.  It  can  easily  be  seen  that  agriculture   is  a  key  to  this  challenge.  Given  the  world͛s  environmental  problems,  simply  growing  more  food  is  not   enough.  Agriculture  will  have  to  be  developed  with  sustainability  built-­‐in  from  the  start.   One  of   the   important  subjects   in  agriculture   is  crop  yield   forecasting.     /ƚ͛ƐĂĚǀĞŶƚďĞŐĂŶ ŝŶ ƚŚĞϭϵϳϬƐ.   Crop  yield  forecasting   is  key  for  government  structures  at  all   levels,   including  E'K͛ƐĂŶĚŝŶƚĞƌŶĂƚŝŽŶĂů organization   such   as   the   United   Nations   as   well   as   companies   that   are   dependent   on   agricultural   produce   as   an   input.   Research   in   crop   yield   forecasting   has   seen   advancements   in   response   to   widespread   famine   in   the   Sahel,   Ethiopia   and   other   countries   in   the   seventies   and   eighties.   The   operational  knowledge  gained  helped  predict  and  partly  avoid  food  shortages  in  the  years  after.   The  target  environments  for  crop  yield  forecasting  have  always  been  two-­‐fold.  In  countries  with  highly   mechanised  large-­‐scale  agriculture,  such  as  the  USA,  Europe  and  Russia,  crop  yield  forecasting  provides   data   to   governmental   structures,   companies   and   farmers.  Good  yield   and  price   predictions   provide   a   clear  strategic  advantage.  Governments  and  supra-­‐national  bodies  (such  as  the  EU)  use  these  data  for   rationalisation  of  policy  adjustments.   The  other  crop  forecasting  arena  is  formed  by  developing  countries,  where  low  staple  food  production   can   have   disastrous   effects.   Predicting   food   shortages   in   developing   countries   early   has   been   the   mandate  of  the  crop  forecasting  units  of  UN  organisations  such  as  FAO  and  WFP,  as  well  as  FEWSNET,   JRC  and  a  number  of  others.   The  technical  methodology  supporting  the  two  operational  sectors   is   largely  comparable,  although  an   important   difference   lies   in   the   type   of   agriculture   studied,   distinguishing   homogenous   large-­‐scale   production  environments  from  heterogeneous,  smallholder  environments.     1.2 Workshop  rationale  and  objectives     Predictions  of   crop  yields  within   the  growing  season  are  critical   inputs   for   a   range  of  agricultural  and   food  security  decisions.    For  example,  management  of  agricultural  input  and  credit  supplies,  agricultural   trade,  food  security  safety  net  and  relief  programs,  agricultural  insurance,  and  recommendations  about   crop  varieties   and  production   technologies  depend  on  or  benefit   from   the  best  possible  estimates  of   6     crop  production.    They  differ  primarily   in   the  timing  of  key  actions  and  hence  the  required   lead-­‐time.     Agricultural  and  food  security  management  can  generally  benefit  from  improvements  in  accuracy  (at  a   given  lead-­‐time)  and  lead-­‐time  (at  a  given  threshold  of  accuracy).       Both  simple  water  balance  and  process-­‐based  crop  models  are  often  used  to  estimate  yields  within  the   growing   season.     In   some   cases,   they   are   coupled   with   seasonal   climate   forecasts   to   reduce   the   uncertainty   associated   with   climate.   The   uncertainties   associated   with   crop   models,   input   data   and   modelling  assumptions  ʹ  collectively  referred  to  as  model  error  ʹ  also  contribute  to  the  uncertainty  of   crop  yield   forecasts.  One  way   to   correct   crop  model  errors   is  by  data  assimilation.    Data  assimilation   involves   using   observed   data   to   update   simulated   model   state   variables   or   to   estimate   model   parameters.  Evidence  in  the  literature  suggests  that  data  assimilation  can  improve  model  performance.   Remote  sensing  (RS)  by  satellites  offers  several  options  for  reducing  crop  forecasting  errors,  particularly   in  data-­‐sparse  regions.  Biophysical  variables  retrieved  from  remote  sensing  data,  such  as  Leaf  Area  Index   (LAI),   soil  moisture  and  ET,  obtained  at  adequate   spatial  and   temporal   resolutions,   can  potentially  be   coupled  with  crop  models  to  provide  valuable   information   for  crop  yield   forecasting  at  various  scales.   However,  heterogeneous,  smallholder  farming  environments  present  significant  challenges  for  the  use   of   remote   sensing   data   assimilation   for   crop   yield   forecasting,   as   field   size   within   these   highly   fragmented   landscapes   is  often  smaller   than   the  pixel   size  of   remote  sensing  products   that  are   freely   available.     JRC  and  CCAFS  jointly  sponsored  the  workshop  on  June  13-­‐14,  2012,  at  the  JRC  in  Ispra,  Italy,  to  identify   avenues  for  exploiting  remote  sensing  information  to  improving  crop  forecasting  in  smallholder  farming   environments.  The  ǁŽƌŬƐŚŽƉ͛Ɛ  objectives  were:   x To  advance  the  state-­‐of-­‐knowledge  of  data  assimilation  for  crop  yield  forecasting;   x To  address  challenges  and  needs  for  successful  applications  of  data  assimilation  in  forecasting  crop   yields  in  heterogeneous,  smallholder  environments;  and     x To  enhance  collaboration  and  exchange  of  knowledge  among  data  assimilation  and  crop  forecasting   groups.     The   workshop   succeeded   in   bringing   together   scientists   from   around   the   world.   This   has   enabled   discussions  on  research  and  results  and  has  greatly  enhanced  collaboration  and  exchange  of  knowledge,   especially  about  data  assimilation  and  crop  forecasting.     1.3 Workshop  salient  findings     This   workshop   was   organized   to   exchange   knowledge   on   crop  models   and   remote   sensing   for   yield   prediction,   especially   for   heterogeneous,   smallholder   environments.   Organisations   such   as   JRC   and   various   UN   organisations   are   interested   in   progress   in   crop   modelling,   as   it   helps   to   improve   their   operational   yield   forecasting.   From   an   operational   viewpoint   Francois   Kayitakire   of   the   EU   Joint   Research  Centre  sets  the  most  pressing  challenges  as  follows:   7     x Advanced   remote   sensing   and   modelling   techniques   have   not   yet   reached   operational   real-­‐time   crop  forecasting.   x So  far,  the  spatial  resolution  of  models  and  feasible  remote  sensing  is  hardly  adequate  for  most  of   cropping  systems  in  Africa.   x About  the  timing  of  crop  yield  forecasting:  for  operational  circumstances   it  would  be  best  to  have   good  crop  forecasts  about  two  months  before  harvest,  although  it  might  be  more  realistic  to  have  it   one  month  before.   x In  smallholder  environments,  it  is  still  unknown  which  crops  are  grown  and  when.   The   workshop   shows   that   there   have   been   clear   advances   in   crop   yield   forecasting.   Important   innovations  were  made  in  the  use  of  remote  sensing-­‐crop  model  integration  through  data  assimilation.   In  essence  data  assimilation   is   the  technique  whereby  remote  sensing  data  are  used  as   inputs   in  crop   models,  to  adjust  or  reset  state  variables  in  crop  models.  Several  techniques  exist  to  do  this  of  which  the   Ensemble  Kalman  Filter  is  applied  most.   The  most  noticeable  advances  have  been  made  in  homogenous  environments.  Good  examples  of  these   cropping   environments  were   presented   for   the  mid-­‐western   states   of   the  USA   and   Russia.   For   these   environments,  scientists  showed  that  the  solution  lies  in  the  use  of  high-­‐resolution  remote  sensing  data   integrated  with  advanced  crop  models.  Some  of  this  research  has  reached  practical  applicability.  As  an   example,  grain  yields  can  be  forecasted  using  high  resolution  remote  sensing  fed  into  a  crop  model  and   subsequently  checked  against  combine  harvester  data.   This   is   not   (yet)   feasible   in   an  African   setting.   For   these   environments,   low-­‐cost  moderate   resolution   imagery   is   more   feasible,   combined   with   increased   knowledge   on   extracting   signatures   for   targeted   crops   and   cropping   systems.   In   the   workshop   in-­‐depth   research   has   been   presented   on   smallholder   environments  in  Africa  and  Asia  based  on  the  study  of  carbon,  water  and  energy  cycles.  It  was  showed   that   the  heterogeneous,   smallholder   cropping  environment   is   slowly  being  understood   in   satisfactory   detail.   Incorporation  of  other  data   (e.g.  socio-­‐economic  data)  proved  to  be  needed  to  understand  the   crop  production  to  its  full  extent.   For   smallholder   environments,   some   participants   advocated   the   use   of   high-­‐resolution   techniques,   coupled  with  an  in-­‐depth  knowledge  of  the  area  of  study.  Promising  field  experiments  are  being  set  up   in  Mali,  Niger,  India  and  other  countries  to  study  the  heterogeneous,  smallholder  environments.  Others   felt   that   (for   country   of   continental   scale   predictions)   low-­‐resolution   techniques   (remote   sensing,   models  and  data)  are  the  way  to  go  forward.     CCAFS  theme  2  main  goals  are  to  build  resilient  rural  livelihoods,  ensure  food  delivery,  trade,  and  crisis   response   and   enhanced   climate   information   and   services.   Assisting   scientists   in   the   field   of   crop   forecasting  is  one  of  the  ways  to  achieve  these  goals.  During  the  workshop  Jim  Hansen  (Theme  2  leader)   of  CCAFS  led  the  discussion  on  how  to  address  the  challenges  for  applying  RS  data  assimilation  for  crop   forecasting  in  heterogeneous,  smallholder  environments.  With  respect  to  data,  high  resolution  remote   sensing  was  offered  (to  the  m  scale)  but  seems  to  be  unfeasible  for  operational  use  in  Africa  because  of   scale   and   cost.  Moderate   resolution   remote   sensing   combined  with   downscaling   techniques   e.g.,   un-­‐ 8     mixing   vegetation   signature   seems   to   be   interesting,   like  what   is   being   pursued   by   IRI/JPL.   Fusion   of   moderate   (shorter   return  period)   and  high   resolution   (longer   return   period)   remote   sensing  was   also   discussed   with   some   reluctance   from   the   group.     In   terms   of   data   integration,   the   state-­‐parameter   simultaneous  update  within  the  Ensemble  Kalman  Filter  was  discouraged  especially  when  using  LAI  for   data  assimilation.  A   framework  was  proposed   in  which   crop  model  parameters   first   are  estimated  by   inverse  modelling,  and  then  the  calibrated  model  can  be  linked  with  the  Ensemble  Kalman  Filter  for  the   assimilation  of  LAI  for  forecasting  yield.  Proof  of  concept  study  was  discussed  using  data  from  India  and   Mali.       9     2 Summary  of  presentations  and  discussions     During  the  workshop,  crop  production  has  been  highlighted  from  many  sides,  using  a  variety  of  models,   satellite   parameters   and   field   data.   Subjects   range   from   field   to   continent   level,   from   small   scale   to   large-­‐scale   crop  production,   from   tropical   to   temperate   regions,   from  maize   to  millet.  With   the   large   variety  of  presented  subjects,  it  is  difficult  to  honour  each  and  every  subject  in  this  summary  report.   However,  trends  in  crop  yield  forecasting  for  heterogeneous,  smallholder  environments  can  certainly  be   observed.  Some  of  the  trends  and  observations  discussed  in  the  following  pages:   x Data  assimilation  techniques   x Crops  researched   x Use  of  Crop  Models   x Use  of  Remote  Sensing  parameters  as  proxies  for  biomass  production   x Use  of  satellite  sensors   x Research  locations   x Spatial  scales  (from  field  to  continent)   x Heterogeneity   x Crop  masks   x Crop  management  factors   x Uncertainty  of  predictions   x Linkage  with  other  sources  of  information   Although  the  presenters  covered  a  wide  range  of  subjects,  the  analysis  of  the  presentations  has  led  to   some  conclusions  that  are  summarized  in  the  following  sections.     2.1 Data  assimilation  techniques     During   the   workshop   it   has   been   shown   that   data   assimilation   can   be   applied   successfully   in   crop   modelling   studies.   In   general   two   types   of   assimilation   techniques   were   demonstrated   in   the   presentations:   1. A  recalibration  strategy  where  some  uncertain  model  parameters  (for  example  the  emergence  date)   are  optimized  by  minimizing  the  difference  between  the  model  and  the  observations  available.     2. A   sequential   updating   strategy  where  model   states   or   parameters   are   updated  during   the  model   run.  A  prerequisite   for   this   technique   is   that   the  model   allows  adjusting   the   states  or  parameters   during  the  model  run.  Essentially  in  data  assimilation,  model  parameters  in  a  time-­‐step  are  re-­‐set  or   corrected  by  observations  from  the  real  world.     10     For  crop  modelling  a  logical  source  of  these  observations  is  remote  sensing.  Remotely  sensed  data  are   typically  sensor  data  gained  from  platforms  such  as  satellites,  aircraft  and  surface-­‐bound  sensors.     In  data  assimilation,  one  could  simply  replace  the  model  results  by  observations.  In  practice  this  is  not  a   good  idea,  because:   x Both  the  external  data  and  the  model  results  contain  errors;   x Often  a  proxy  of  the  state  variables  is  assimilated;   x Almost  always  one  needs  to  update  many  (unobserved)  state  variables  using  only  one  or  a  couple  of   observations;  and   x Continuity  of  the  observations  is  not  guaranteed  (cloud  cover,  satellite  failure).   Therefore  methods   like   the  Ensembles  Kalman   filter  have   to  be  applied   (Pauwels)  or   re-­‐calibration  of   model  parameters  are  better  options  for  data  assimilation.   Presenters  argued  that  data  assimilation  in  crop  growth  related  models  has  its  challenges:   x The  studied  processes  and  models  have  biases  that  are  not  taken  care  of  in  some  of  the  algorithms   that  support  data  assimilation.  Working  with  biases  is  often  possible  by  applying  corrections  for  bias   to  the  original  algorithms  (Pauwels).   x A  parallel  process   to  biomass  production,  or   crop  growth,   is   crop  development   (phenology  of   the   plant).  When  trying  to  assimilate  a  remote-­‐sensing  estimation  of  biomass   in  a  crop  growth  model,   ŽŶĞŵĂLJďĞĨĂĐĞĚďLJǁŚĂƚƐŽŵĞĂƵƚŚŽƌƐĐĂůůĂ͞phenological  shift͟;ƐĞĞƵƌŶĞůĞƚĂů͘ϮϬϭϭͿ͘ĂƐŝĐĂůůLJ͕ a  given  amount  of  green  biomass  may  be  attained  at  different  stages  of  the  crop  season,  e.g.  in  the   increasing  or   the  decreasing  part  of   the  curve,  and  forcing   this  biomass  value   in  a  model  without   knowing  the  phenology  can  result  in  dramatically  wrong  results.     x It   is   still   unclear  which   combination   of   satellite   data   and   crop  modelling   (input   data;   calibration;   assimilation)   is  most   effective.   It   needs   to   be   studied  which   crop  data  are  most   suitable   for   data   assimilation  at  available  temporal  and  spatial  scales  of  satellite  images  (van  Dam)   x Interaction/dependency  between  parameters  may  lead  to  errors  in  their  estimation  (Guerif)     2.2 Crops  researched     The  table  below  shows  that  most  of  the  research  in  this  workshop  is  done  on  Maize  and  Wheat.  Both   crops  are  among   the  most  cultivated   in   the  world   (grown   in   tropical  as  well  as  areas  with   temperate   climates),  which  could  explain  part  of  the  popularity.  Furthermore,  these  crops  are  often  grown  in  large   fields  on  large  farms  with  advanced  crop  management  practices  like  precision  farming  (Bach),  leading  to   a  nicely  homogenous  crop.   Of  the  pure  tropical  crops,  rice,  sorghum  and  millet  are  mentioned  most  (Table  1).  Especially  millet  and   sorghum  are  grown  in  the  heterogeneous,  smallholder  environments  that  are  subject  of  this  workshop.   11     Millet  and  sorghum  are  often  grown  as  landraces.  Even  within  a  cropped  field  a  large  variety  of  heights,   phenology  and  production  exists.   For   farmers   this  might  have  an  advantage,   as   they   seem  to  aim   for   minimizing   risks   rather   than  maximizing  production   (Akponikpe).   It   does,  however,  make   research   for   these  corps  challenging.   Table  1.  Names  of  crops  mentioned  in  the  presentations   Crop   No.  of  Presentations   Total  Occurrences   Maize   11   56   Wheat   10   46   Rice   5   61   Soybean   5   34   Sorghum   5   17   Sugar  beet   3   22   Millet   3   13   Cereals   2   3   Pulses   1   1   Tubers   1   1     A  distinction  has  been  made  between  C4  crops  like  maize  and  C3  crops  like  wheat  and  rice.  As  the  C4   crops  have  a  slightly  different  photosynthetic  cycle,  their  reaction  on  radiation,  CO2  content  and  other   environmental  parameters  proved  to  be  different  (Drewry).     2.3 Use  of  Crop  Models     It  is  interesting  to  see  which  crop  models  are  used  most  often  in  research  presented.  Both  statistical  and   dynamic/mechanistic  crop  modelling  has  been  used  in  the  workshop.   The  models  SWAP/WOFOST  and  DSSAT  were  most  popular  among  scientists  presenting  in  this  workshop   (Table  2).     The  SWAP/WOFOST  model   (Soil,  Water,  Atmosphere  and  Plant)   simulates  vertical   transport  of  water,   solutes   and   heat   in   unsaturated/saturated   soils.   The   program   is   designed   to   simulate   the   transport   processes   at   field   scale   level   and   during   entire   growing   seasons.   SWAP   is   open-­‐source   and   can   be   downloaded  here:  http://www.swap.alterra.nl/.  SWAP  incorporates  WOFOST,  which  is  also  used  stand-­‐ alone.     A   standalone   version   of   WOFOST   and   derived   models   can   be   downloaded   from   http://www.wageningenur.nl/wofost   other   Wageningen   models   can   be   downloaded   from   http://models.pps.wur.nl.       The  Decision   Support   System   for   Agrotechnology   Transfer   (DSSAT)   is   a   software   application   program   that  comprises  crop  simulation  models  for  over  28  crops.  The  crop  simulation  models  in  DSSAT  simulate   12     growth,  development  and  yield  as  a  function  of  the  soil-­‐plant-­‐atmosphere  dynamics.  Although  DSSAT  is   not  open  source,  its  source  code  and  executables  can  be  requested  for  free  from  http://www.dssat.net/   Some   models   are   developed   by   the   presenting   scientists   themselves   and   not   distributed   to   other   groups.   These   models   are   often   used   in   precision   agriculture   for   direct   advice   to   farmers   (APSIM,   http://www.apsim.info).   Table2.  Use  of  models  in  workshop  presentations   Models   No.  of  Presentations   Total  Occurrences   SWAP   3   18   DSSAT   3   14   CSM   3   14   WOFOST   3   8   PROMET   1   27   MM5   1   11   WTGROWS   1   9   STICS   1   6   MODFLOW   1   3   SUCROS   1   3   ORYZA1   1   2   LINGRA   1   1   WARM   1   1   PROSAIL   1   1   AGROMETSHELL   1   1   APSIM   1   1   MLCan   1   1     The  models  described  in  the  workshop  describe  crop  biomass  production  roughly  through  the  study  of  3   processes:   water   cycle   (water   balance   models),   energy   cycle   (radiative   transfer   models)   and   carbon   cycle.  Many  models  take  two  or  more  of  these  processes  into  account.   Plant  growth  models  are  relatively  good  in  simulating  the  potential  growth,  as  affected  by  climate  and   crop  characteristics  (Figure  1).     Also   the   growth   inhibiting   effects   of   water   shortage,   oxygen   shortage,   salinity   excess   and   nutrient   shortage  can  be  simulated  quite  well  with  current  crop  growth  models.  However,  the  growth  reduction   due   to  weeds,  pests,  diseases  and  pollutants   is   still  difficult   to   simulate.  Satellites  measure   the  actual   growth   conditions,   which   includes   the   total   effect   of   all   growth   reducing   factors.   This   may   cause   a   mismatch  between  crop  growth  simulations  and  measured  crop  growth  by  satellites.   ^ĐŝĞŶƚŝƐƚƐŝŶĐƌĞĂƐŝŶŐůLJƵƐĞ͞ŵŽĚĞůŝŶǀĞƌƐŝŽŶ͟ǁŚĞƌĞďLJƚŚĞŵŽĚĞůŝƐĨĞĚǁŝƚŚŽƵƚƉƵƚƉĂƌĂŵĞƚĞƌƐƚŽŐĞƚĂ better  understanding  of  the  driving  input  variables/properties  (Honda,  Guerif  and  Sehgal).   13       Figure  1.  Plant  growth  simulation  is  affected  by  defining  climate  and  crop  characteristics  (potential   growth),  limiting  factors  and  reducing  factors.  All  factors  together  result  in  actual  growth.     2.4 Use  of  Remote  Sensing  parameters  as  proxies  for  biomass  production     Proxies   for   yield   and   biomass   production   have   been   developed   over   the   years   from   remote   sensing   derived   spectral   measurements.   The   products   involve   different   spectral   bands,   various   retrieval   algorithms  and  corrections.  The  most  popular  products  (in  terms  of  occurrence  in  the  presentations  of   this  workshop)  are  mentioned  in  Table  3.   Table  3.  Occurrence  of  parameters  and  proxies  in  the  presentations     Parameters/Proxy   No.  of  presentations   Total  Occurrences   LAI  &  GAI   (Leaf  Area  Index  &  Green  Area  Index)   11   177   NDVI  (Normalized  Difference   Vegetation  Index)   9   83   Evapotranspiration   8   26   Precipitation  derived  from  RS   5   8   fAPAR  (fraction  of  Absorbed   Photosynthetically  Active  Radiation)   4   31   IR  (infrared)   2   3   EVI  (Enhanced  vegetation  Index)   1   8   Global  Radiation   1   3     Potential Actual defining factors‡ CO2‡ radiation‡ temperature‡ crop characteristics9 physiology, phenology9 canopy architecture defining factors+limiting factors‡ water shortage‡ oxygen shortage‡ salinity excess‡ nutrient shortage defining factors+limiting factors+reducing factors‡ weeds‡ pests‡ diseases‡ pollutants 14     The  parameters  above  can  be  extracted  from  a  variety  of  satellite  platforms.  In  practice,  MODIS,  SPOT,   NOAA-­‐AVHRR   and   MSG   are   often   used.   The   parameters   have   been   used   at   low,   medium   and   high   resolutions  at  various  scales.   The  most   frequently  used  parameter   is   LAI   (Leaf  Area   Index).   This  parameter  has  been  developed  50   years  ago  for  field  experiments.      /ƚ͛s  defined  as  half  the  total  developed  area  of  green  leaves  per  unit  of   ground   horizontal   area   (Chen   &   Black,   1992).   The   satellite-­‐based   LAI   products   are   generally   not   the   same  variables  as  the  LAI  in  crop  growth  models  or  the  LAI  measured  in  a  field.  A  main  reason  for  this   discrepancy   is   that   available   satellite   LAI   are   produced   from   reflectance   obtained   from   coarse   spatial   resolution  pixels,  in  which  various  different  types  of  vegetation  covers  are  present.  For  the  same  reason,   several  scientists  have  proven  that   the  satellite  based  LAI  can  differ  considerably   from  field  measured   LAI   (Honda).   Sometimes   LAI   is   referred   to   as   GAI   (for   Green   Area   Index).   For   several   crops   in  which   various  part  of  the  plant  photosynthesis  (e.g.  cereals),  it  is  actually  more  appropriate  to  use  this  term  to   refer   to   the   biophysical   variable   retrieved   from   remote   sensing   since   the   radiance  measured   by   the   instrument  is  made  of  electromagnetic  radiation  reflected  from  all  plant  organs  (Duveiller  et  al.,  2011a).   A   biophysical   variable   that   is   generally   as   widely   available   as   LAI   is   the   fraction   of   Absorbed   Photosynthetically  Active  Radiation  (fAPAR).  This  variable   is  actually  more  closely  related  to  yield  than   LAI.  For  diverse  reasons  (one  being  that  fAPAR  is  generally  not  a  state  variable  in  the  current  generation   of   simulation   models)   it   seems   to   be   much   less   popular   for   data   assimilation   in   crop   models,   even   though   it   probably   avoids   some   of   the   problems/uncertainties   encountered  with   LAI.   This   point  was   raised  in  the  workshop  and  proposed  as  a  justified  research  direction.   The   NDVI   (Normalized   Difference   Vegetation   Index)   has   been   used   widely.   This   parameter   has   been   around  for  quite  some  time  and  long  historical  records  exist.  Many  derivatives/refinements  of  NDVI  are   now   in   use   such   as   DVI   (Difference   Vegetation   Index)   and   EVI   (Enhanced   Vegetation   Index;   used   by   Hoogenboom).   An   estimate   of   actual   evapotranspiration   can   be   based   on   satellite   signals   only.   Crop   models   often   calculate  actual  evapotranspiration  as  output.  While  the  first  method  is  based  on  evapotranspiration  of   the  entire  vegetation  by  pixel,  the  second  approach  makes   it  possible  to  be  crop-­‐specific.  Examples  of   both  approaches  were  shown.   An  issue  that  returned  various  times  in  the  discussions  was  which  model  variables  should  be  updated  at   satellite  overpass.  For  instance,  if  LAI  is  measured,  not  only  the  LAI  but  also  many  other  model  variables   that  are  related  to  leaf  area  index  (such  as  plant  biomass,  green  area  index,  development  stage)  should   be  updated.  The  plant  model  update  should  be  consistent.  Various  groups  use  different  methods.   Satellite  derived  precipitation  estimates  are  used   in  crop  forecasting,  but   it  has  been  proven  that   this   parameter  is  related  poorly  to  yields  when  applied  as  cumulative  over  the  crop  period  (Irénikatché).  As   input  to  crop  models  at  a  daily  or  dekadal  time-­‐step  it  has  however  proven  its  usefulness.       15     2.5 Use  of  satellite  sensors     For  data  assimilation,  satellite  based  parameters  are  widely  used  in  combination  with  crop  models.  See   below  a  table  of  the  satellites  mentioned  by  the  presenters  in  the  workshop  where  satellite  names  and   sensor  names  are  mixed.  Of   the  satellites/sensors   listed  below,  data   from  Landsat,  NOAA  AVHRR,  EO,   Terra  (Aster  and  MODIS),  Aqua  (MODIS)  and  Envisat  (MERIS  and  ASAR)  are  available  free  of  charge  (van   Dam).     Table  4.  The  use  of  satellites  and  sensors   Platform   Sensor   No.  of  Presentations   Total  Occurrences   Terra  and  Aqua   MODIS   9   30   SPOT   VEGETATION   5   10   SPOT   HRG/HRV/HRVIR   2   4   NOAA     AVHRR   5   6   LANDSAT   TM/ETM   4   7   MSG  (METEOSAT)     2   6   Sentinel  (still  to  be   launched)   OLCI   2   2   RapidEye     1   2   Envisat/MERIS   MERIS   1   2   Quickbird     1   2   TRMM     1   2     The   MODIS   sensors   are   mentioned   most   frequently.   MODIS   (Moderate   Resolution   Imaging   Spectroradiometer)   is   an   instrument   aboard   the   Terra   (EOS  AM)   and  Aqua   (EOS  PM)   satellites.   Terra   MODIS  and  Aqua  MODIS  are  viewing  the  entire  Earth's  surface  every  1  to  2  days,  acquiring  data   in  36   spectral  bands.  MODIS  is  widely  used  because  its  products  are  free,  easily  available  for  download,  and   some   more   elaborated   products   such   as   LAI   and   FAPAR   are   distributed   along   the   usual   spectral   reflectances  and  indices.       Although  with  the  higher  level  products  such  as  LAI  a  wide  range  of  corrections  have  been  applied,  some   researchers   report   that   these   products   have   to   be   used  with   care   and   do   not   always   align  with   the   situation   on   the   ground   (Honda;   van   Dam).   This   is   in   part   due   to   lack   of   adequacy   between   the   observation  support   (i.e.  where   satellite  data  was  collected)  and   the   field   size  which   is  visited  on   the   ground.   A   tentative  movement   away   from  optical   sensors   to   radar   sensors  has  been  noted.  Radar  penetrates   clouds   and   is   therefore   less   susceptible   to   atmospheric   disturbances   (Bakary).   However,   the   passive   radar  sensors  generally  have  a  low  resolution,  and  in  general  radar  signals  are  still  a  challenge  to  use.   16     Many   low-­‐resolution   satellite   data   are   available   at   high   frequency,   while   high   resolution   data   are   available  at   low  frequency.  Various  algorithms  exist  to  combine  low  and  high-­‐resolution  data  to  derive   the  optimal  amount  of  information  (Ines;  Honda).   Besides  satellite  sensors,  some  scientists  use  earth-­‐bound  sensors  on  poles  as  well  as  small,  unmanned   airplanes  (Drewry,  Honda).     2.6 Research  locations     Most  of  the  research  presented  has  been  conducted  in  Africa,  with  the  country  of  Niger  at  the  top  of  the   list   (Table   5).   Niger   occurs   33   times   in   4   presentations   (Akponikpe,   Traoré,   Bakary,   Hansen).   Other   African  countries  the  presenters  mentioned  were  Senegal,  Mali,  Sudan  and  Burkina.  Little  research  from   English  speaking  African  countries  has  been  presented  with  the  exception  of  Kenya  and  Ghana.   Table  5.  Names  of  countries,  regions  and  states  in  all  presentations   Country   No.  of  Presentations   Total  Occurrences   Niger   4   33   Senegal   3   15   Mali   3   10   Europe   3   8   Sudan   3   7   USA   3   7   Belgium   3   6   Netherlands   3   6   Burkina   3   5   Ethiopia   3   4   Nepal   2   4   Ghana   2   4   Kenya   2   2   France   1   9   Thailand   1   4   Egypt   1   3   Japan   1   3   Tunisia   1   3   Armenia   1   2   Iberian   1   1   Tanzania   1   1   Uganda   1   1     17     Quite  some  research  findings  were  presented  on  European  countries,  mainly  The  Netherlands,  Belgium,   Germany,  the  Iberian  Peninsula  and  Russia.  The  mid-­‐western  states  of  the  United  States  were  frequently   used  as  research  locations.  These  states  have  an  advantage  over  other  study  areas  due  to  their  relatively   ŚŽŵŽŐĞŶŽƵƐĐƌŽƉĐŽǀĞƌƐĚƵƌŝŶŐ ƚŚĞĐƌŽƉƉŝŶŐ ƐĞĂƐŽŶ͘dŚŝƐĞŶĂďůĞƐ ƚŚĞ ƐĐŝĞŶƚŝƐƚƐ ƚŽ ĨŝŶĚ͞ĂůŵŽƐƚƉƵƌĞ ƉŝdžĞůƐ͟ŝŶƌĞŵŽƚĞƐĞŶƐŝŶŐŝŵĂŐĞƌLJ͘     2.7 Spatial  scales     The   spatial   scale  of   the   research  matters   for   the  methods   and  data   that   can   be   applied   successfully.   Studies  were  presented  at  a  wide  range  of  spatial  scales,  ranging  from  field  to  continent.  A  somewhat   arbitrary  list  of  scales  mentioned:   x Field  level  (van  Dam  and  Bach,  Drewry,  Akponikpe)   x Village  level  (Traoré,  Akponikpe)   x District  level  (Seghal,  Bakary,  Guerif)   x Country  level  (Marinho,  Meroni)   x Sub-­‐continent  and  continent  level  (Duveiller  and  Terink)   Some  debate  was  noticeable  among   the  scientist  on   the  question  whether  methods  at   the   finer   level   (e.g.  field)  can  successfully  be  scaled  up  to  any  level  above.  While  some  argued  that  it  is  just  a  matter  of   computing   power,   others   insisted   that   different  models   and   datasets   have   to   be   applied   at   different   spatial  scales.     In   general,   it   became  apparent   that   research  at   the   field   level  helps   to  understand   complex   cropping   systems  and  leads  to  better  inputs  and  management  techniques  on  farm  level  while  research  on  district   ĂŶĚŚŝŐŚĞƌƐĐĂůĞƐŚĞůƉƐƉŽůŝĐLJŵĂŬĞƌƐŝŶŐŽǀĞƌŶŵĞŶƚƐ͕E'K͛ƐĂŶĚŝŶƚĞƌŶĂƚŝŽŶĂůŽƌŐĂŶŝƐĂƚŝŽŶƐ͘ Ideally  a   methodology  should  be  developed  which  addresses  both  field  and  regional  scale,  as  for  instance  shown   by  Bach.     2.8 Heterogeneity     One  of  the  most  challenging  aspects  of  the  use  of  remote  sensing  proved  to  be  the  heterogeneity  of  the   crop/vegetation   in  one  pixel.   This   is  most   apparent   in   low-­‐resolution   imagery   (e.g.   >   1   km  pixel   size).   ͞WƵƌĞ ƉŝdžĞůƐ͟ ĨŽƌ low-­‐resolution   imagery   can   be   found   in   the   USA   and   Russia,   but   are   almost   non-­‐ existent   in   Africa   and   Europe   minus   Russia.   Some   recent   research   has   shown,   however,   that   pure   enough   pixels   can   be   obtained   in   highly   fragmented   landscapes   in   Europe   in   order   to   have   a   crop   specific  signal  (de  Wit)  if  medium  spatial  resolution  imagery  such  as  MODIS  (250m)  is  employed  and  the   spatial  response  of  the  instrument  is  carefully  taken  into  account  (Duveiller  et  al.  2011b).  This  approach   18     allows  an  alternative   solution   to  un-­‐mixing   coarse  pixels,   but  on   the  other  hand   still   requires   some  a   priori  information  of  where  the  crops  are  located  beforehand.   High-­‐resolution   imagery  proved   to  be  helpful   to  detect   in-­‐field  variability  on   large-­‐scale   farms   (Bach).   This  kind  of  high  spatial   resolution   imagery   is   typically  available  only   for  a   limited  geographic  extend,   and  with  a  temporal  revisit  capacity  which   is   lower  than  desired  for  agricultural  monitoring.  Although,   ĨƵƚƵƌĞ ƐĂƚĞůůŝƚĞ ĐŽŶƐƚĞůůĂƚŝŽŶƐ ;ƐƵĐŚ ĂƐ ƚŚĞ ƵƌŽƉĞĂŶ ^ƉĂƚŝĂů ŐĞŶĐLJ͛Ɛ ^ĞŶƚŝŶĞů-­‐2)   aim   at  making   high   spatial  resolution  imagery  operationally  available  worldwide,  there  remains  the  challenge  of  managing   this  exorbitant  amount  of  data  and  extract  from  it  a  clear  and  reliable  information  than  can  be  used  for   assessing  crop  status.   In  heterogeneous,  smallholder  environments,  even  high  resolution  imagery  had  to  be  complemented  by   extensive  field  research  to  successfully  describe  the  heterogeneity  of  fields  and  crops  (Traoré).     2.9 Crop  masks     Several  researchers  noted  the  lack  of  good  crop  masks  (Marinho,  Kayitakire).  Unfortunately,  land  cover   maps   just   specify   agricultural   practices   (arable   land,   rangeland   etc.),   and   rarely   go   down   to   the   crop   level.  For  many  areas,  such  crop  masks  should  ideally  be  done  on  a  yearly  basis  to  reflect  the  changes   that   occur   due   to   crop   rotation   or   expansion/regression   of   crop   extends.   Crop   rotation   is   the  main   limitation   in   Europe   that   forces   the  operational  MARS   crop   yield   forecasting   system  of   the  European   Commission  from  using  crop  specific  time  series  (Duveiller).   Another   challenge   is   that   crop  masks   cannot   be   considered   constant   as   different   crops   are   grown   in   different  years.  Even  percentage-­‐wise  pixel  estimates  (for  example  20%  wheat,  30%  maize  etc.)  are  only   available  for  some  well-­‐researched  areas.   Researchers   generally   put   quite   some   work   into   crop   masks,   before   the   actual   research   topic   was   investigated  (Traoré,  Hoogenboom).     2.10 Crop  management  factors     Crop  yields   are   to  a  high  degree  determined  by   the  management  practices   applied   to   it   (Sehgal).   For   crop   yield   forecasting   the  most   important   ones   are   sowing   dates,   irrigation   and   nutrient   application.   Crop  model  outcome  is  to  a  high  degree  dependent  on  sowing  date  (Traoré).   Participants  showed  several  methods  to  estimate  sowing  dates:   x Simulated  sowing  date,  based  on  external  parameters  (Akponikpe);   19     x Estimated  sowing  dates  extracted  from  remote  sensing  time  series  (Guerif);   x Establishing  sowing  dates  through  field  work  (Sehgal)  or  local  sensors  in  fields  (Honda).   Obviously  the  scale  of  the  study  (from  field  to  continent)  determines  the  possibilities.  At  higher  scales   (country,  continent),  fieldwork  is  not  a  workable  solution  to  determine  management  factors  applied.     2.11 Uncertainty  of  predictions     Uncertainty  in  crop  yield  predictions  remains  a  problem.  This  is  particularly  the  case  early  in  the  season.   Generally  the  uncertainty  declines  towards  the  end  of  the  season.  Uncertainty  during  the  season  can  be   lowered  through  seasonal  climate  forecasts  (Hansen).     Model  uncertainty  can  partly  be  addressed  by  data  assimilation   techniques,  while   climate  uncertainty   can  be  addressed  by  seasonal  forecasts  (Ines).     2.12 Linkage  with  other  sources  of  information     It   has   been   advocated   during   the  workshop   that   scientists   look   at   linkages  with   information   sources   outside   the   traditional   soil-­‐water-­‐plant   system.   Social   economic   databases   and   other   sources   that   explain  small-­‐scale  farmers  livelihoods  from  a  different  angle  are  to  be  integrated  with  crop  models  for  a   better   understanding   of   crop   production   systems.   Potentially   this   could   go   further   than   establishing   simple  correlations.  Models   integrating   for  example  socio  economic   information  with  crop  production   systems  are  yet  to  be  developed  (Guerif).   The   recent   AgMIP   project   combines   climate,   crops   and   economics   (Traoré).     Within   AgMIP   a   large   number  of   crop  and  agronomy  modelling   groups   cooperate   to   compare  modelling   results   for   existing   crop  datasets  and  for  future  conditions,  including  climate  change.       20         21     3 Presentation  Abstracts     3.1 The  challenges  of  an  operational  crop  yield  forecasting  system  in  Sub-­‐ Saharan  Africa   Francois  Kayitakire,  JRC,  MARS  Unit,  FOODSEC  Action,  Ispra,  Italy   The  Food  Security  Assessment  (FOODSEC)  Action  of  the  EC-­‐JRC  supports  the  implementation  of  EU  Food   Security   and   Food  Assistance   policies   by   providing   scientific   advice   and  objective   assessment   of   food   security  situation.  It  has  been  developing  pieces  of  an  early  warning  system  to  monitor  crop  and  pasture   production,  with  a  focus  on  most  food  insecure  areas,  mainly  in  Sub-­‐Saharan  Africa.  The  system  was  by   ůĂƌŐĞĐŽŶĐĞŝǀĞĚĂƐĂŶĞdžƚĞŶƐŝŽŶŽĨƚŚĞ͞DŽŶŝƚŽƌŝŶŐŐƌŝĐƵůƚƵƌĞǁŝƚŚZĞŵŽƚĞ^ĞŶƐŝŶŐ͟;DZ^ͿƉƌŽũĞĐƚƚŽ regions   outside   the   European  Union.   Thus,   it   relies  mainly   on   remote   sensing   solutions   and   to   some   extent  on  crop  modelling.  Low-­‐spatial  satellite  imagery  is  extensively  used  to  derive  the  crop  conditions   in   agricultural   areas   and   pasture   availability   in   pastoral   areas.   This   approach   proved   effective   for   qualitative  assessment  of  proxies  of   food  production.   In  a   few  cases,   tentative   to   link   remote  sensing   derived   indicators   to   crop   yield   or   production   has   been   done.   Those   indicators   are   usually   analysed   together  with  those  derived  from  meteorological  data,  and  they  make  the  basis  of  the  MARS  crop  and   food  security  monitoring  reports  (http://mars.jrc.ec.europa.eu/mars/Bulletins-­‐Publications).   Crop  modelling  has  up-­‐to  now  played  a  minor  role  in  the  system  for  several  reasons.  The  main  constraint   has   been   the  model   calibration   and   the   availability   of   historical   yield   (and   production)   statistics.   The   area  of  interest  of  the  FOODSEC  Action  is  actually  very  large,  with  many  different  ecological  conditions   and  agricultural  systems  that  are  poorly  understood  and  mapped.  Moreover,  yield  statistics  that  are  a   key  component  in  any  crop  forecasting  solution  are  rarely  available  at  the  appropriate  spatial  resolution   and   temporal   coverage.   Therefore,   JRC   opted   for   a   simple   crop   model   (AgrometShell)   that   was   developed  by  FAO.  For  instance,   it  has  been  used  to  forecast  maize  production  in  Kenya  by  regressing   yield   to   two   variables   derived   from   the   AgrometShell   model:   actual   evapotranspiration   (ETA)   and   water   requirement  satisfaction  index  (WRSI)  (see  graph,  Rojas   2007).     Building  from  this  experience,  JRC,  in  collaboration  with   Alterra   (Netherlands)   implemented   the   core   of   the   AgrometShell   while   customizing   some   modules   and   introducing   a   number   of   improvements   in   the   input   data  (soil  data,  crop  masks,  etc.)  to  build  an  application   that  will  help  to  easily  provide  analysts  with  the  WRSI  at   the   global   level.   That   is   the   Global  Water   Satisfaction   Index   (GWSI)   application   (available   online   through   the   MARS  Viewer:  http://www.marsop.info/marsop3/).     22     ,ŽǁĞǀĞƌ͕ ƚŚĞƌĞ͛Ɛ Ă ŶĞĞĚŽĨ ĂŶĞĨĨĞĐƚŝǀĞ ƋƵĂŶƚŝƚĂƚŝǀĞ Đrop   yield   forecasting   solution.   Crop   forecasting   only  makes  sense  when  the  conclusions  can  be  published  in  time.  In  an  ideal  case,  the  forecast  of  crop   production  is  released  2  months  before  harvest.  It  is  more  realistic  to  expect  estimates  1  month  before   harvest,   but   also   an   analysis   that   comes   in   at   harvest   time   is   still   practical.   The   forecasting  method   should  also  be  able   to  correctly  capture  the   inter-­‐annual  variability  of  yield  because  such  variability   is   the  most  critical  for  food  security  of  vulnerable  households.   A  crop  forecasting  system  based  on  crop  modelling  and  remote  sensing  faces  a  number  of  challenges:     x the  availability  of  yield  data  at  sub-­‐national  levels;   x the  calibration  and  validation  of  models;   x the  availability  of  long  time  series  in  input  data;   x the  course  spatial  resolution  of  input  data,  such  as  remote  sensing.  This  spatial  resolution  is  hardly   adequate  for  most  of  cropping  systems  in  Africa  (mixture  of  crop  fields  and  other  land  cover  types);   x the  necessity  to  know  where  crops  are  grown  (crop  masks).     To  address   these  challenges  will   require   long-­‐ƚĞƌŵƌĞƐĞĂƌĐŚĂŶĚĚĞǀĞůŽƉŵĞŶƚƐ͘Ƶƚ ƚŚĞƌĞ͛Ɛ  perhaps  a   room   for   searching   for   simpler   solutions   with   a   reasonable   accuracy.   This   workshop   provided   some   directions  to  such  solutions.     23     3.2 Crop  Forecasting  within  the  CCAFS  Program   James  Hansen,  Theme  2  Leader  of  the  Climate  Change,  Agriculture  and  Food  Security  research   program  of  the  CGIAR.   The   CGIAR   research   program   on   Climate   Change,   Agriculture   and   Food   Security   (CCAFS)   is   a   major   research   initiative   that   aims   to:     identify   and   develop   pro-­‐poor   adaptation   and   mitigation   practices,   technologies   and   policies   for   agriculture   and   food   systems;   and   support   the   inclusion   of   agricultural   issues  in  climate  change  policies,  and  of  climate  issues  in  agricultural  policies,  at  all  levels.    CCAFS  work  is   organized  in  4  research  themes:   x Theme  1:  Adaptation  to  Progressive  Climate  Change   x Theme  2:  Adaptation  through  Managing  Climate  Risk  (led  by  James  Hansen)   x Theme  3:  Pro-­‐poor  Climate  Change  Mitigation   x Theme  4:  Integration  for  Decision  Making   Theme  2  seeks   to  enhance  the  resilience  of   rural   livelihoods  and  food  systems  to  climate-­‐related  risk.     Improving  climate-­‐related  information  for  risk  management,  across  multiple  scales,  is  an  important  part   ŽĨ ƚŚĞ dŚĞŵĞ͛Ɛ ĐŽŶƚƌŝďƵƚŝŽŶ ƚŽǁĂƌĚ ĐůŝŵĂƚĞ-­‐resilience.   CCAFS   research   currently   focuses   on   East   and   West  Africa  and  South  Asia.   A  number  of  agricultural  and  food  security  decisions   depend   on   the   best   possible   estimates   of   the   impacts  of  climate  fluctuations  on  crops.    While  the   decision   calendar   influences   the   timing   of   information   needed,   most   climate-­‐sensitive   decisions  can  benefit   from   increasing  accuracy   (at  a   given   lead   time)   or   lead   time   (at   a   given   accuracy   threshold).   The   uncertainty   of   a   crop   forecast   consists   of   climate   uncertainty   and   model   uncertainty   (encompassing   all   non-­‐climatic   uncertainties).    Total  uncertainty  diminishes,  and  the   contribution   of  model   uncertainty   increases,   as   the   season   progresses   (see   graph).   Climate   uncertainty   in  weather  can  be  reduced  by  seasonal  forecasts.    Typically  the  greatest  positive  impact  on  uncertainty   occurs   early   in   the   season   (Hansen   et   al.,   2006).   Options   for   reducing   model   uncertainty   include   improving   models,   improving   input   data   and   parameters,   and   data   assimilation   techniques.     These   techniques  show  the  greatest  benefit  later  in  the  season.     CCAFS   contributions   to   crop   forecasting   methodology   and   capacity   include:   reconstructing   historic   meteorological   inputs,   integrating   seasonal  climate   forecasts   into  crop   forecasts,   remote  sensing  data   assimilation,   and   software   platform   development.     However,   understanding   and   fostering   the   use   of   that  information  for  decision-­‐making  is  a  particular  emphasis.     24     3.3 Integration  of  agro-­‐hydrological  modelling,  remote  sensing  and   geographical  information     Jos  van  Dam,  Department  of  Environmental  Science,  Wageningen  University,  The  Netherlands   For  many   years  Wageningen   University   has   been   in   the   forefront   of   crop  modelling   leading   to  well-­‐ known  crop  models  as  WOFOST,  SUCROS  and  LINTUL.  Many  of  these  models  can  be  downloaded  from   http://models.pps.wur.nl.  These  models  have  been  developed  from  a  thorough  understanding  of  crop   production,   down   to   the   role   of   leaf   stomata.   The   agrohydrological   model   SWAP   (Soil   Water   Atmosphere  Plant)  combines  the  crop  growth  model  WOFOST  with  a  detailed  soil  transport  model.  The   graph  below  visualizes  the  processes  modelled  by  SWAP.   Wageningen   University   has   conducted   several   research   projects   in   India   (Sirsa)   and   Iran   (Esfahan)   with   local   partners   with   the   aim   to   gain   knowledge   of   local   cropping   systems,   study   the   water   cycle   and   look   for   ways   to   aggregate   results   from   field   to   region.   The   projects   started   with   data   collection   (both   field   data   and   remote   sensing   data).   The   data   have   been   input   to   the   crop   model   SWAP   and   WOFOST.   A   comparison   is   made   between   the   crop   models   run   with   and   without   input   of   remote   sensing   data   through  data  assimilation.     In   the   uncorrected   SWAP  model,   the   simulated   LAI  was   larger   than   satellite  measured   LAI.   The  main   reasons  are  the  difference  in  scale  between  model  and  satellite  as  well  as  the  fixed  harvesting  data  in   the   model.   The   model   also   showed   larger   fluctuations   than   the   satellite   data,   which   was   also   contributed  to  a  spatial  and  temporal  scale  effect.   As  a  second  track,  remote  sensing  parameters  have  been  used  to  reset  state  variables  in  the  model.  The   assimilation  of  satellite-­‐based  LAI  measurements  was  most  effective.  This  significantly  reduced  the  bias   percentages  for  predictions  one  month  in  advance  of  harvest.  However,  bias  percentages  for  predictions   two  months  ahead  of  harvest  were  not  influenced  positively  by  assimilation  with  LAI  (Vazifedoust  et  al.,   2009).  In  the  near  future,  Wageningen  University  intends  to  apply  these  methodologies  at  common  sites   in  Mali  and  India.       25     3.4 Assimilating  remote  sensing  data  into  crop  models  improves  predictive   performance  for  spatial  application   Martine  Guerif,  UMR  EMMAH  INRA  UAPV,  Avignon,  France   Crop   models   are   powerful   tools   for   dealing   with   agro-­‐environmental   issues   such   as   the   impact   of   agriculture  on  soil  and  water,   the   impact  of  agriculture  on  climate  change,   the  evaluation  of  cropping   systems.  Models  can  help  with  strategic  and  tactic  decisions  for  sustainable  cropping  systems.  Models   can  be  applied  at  different  scales:  field/farm,  region,  country  and  continent.   For   the   use   in   data   assimilation   the   following   parameters  can  be  obtained  from  three  sensor  types   (solar,  TIT  and  µ-­‐wave).  The  number  of  plusses  on  the   figure  indicates  the  applicability  level.   INRA   uses   the   crop   model   STICS,   which   has   a   daily   time-­‐step.   Its  main   aim   is   to   simulate   the   effects   of   the  physical  medium  and  crop  management  schedule   variations  on  crop  production  and  environment  at  the   field  scale.     With   STICS,   three   types   of   data   assimilation   techniques  have  been  applied:   x Forcing   of   observed   variables   into   the   model   when   the   model   doesŶ͛ƚ ƐŝŵƵůĂƚĞ ƚŚŝƐ ǀĂƌŝĂďůĞ   (or  when  the  model  accepts  to  replace  the  simulated  value  by  a  prescribed  one);   x Sequential   correction  of  model  predictions.    Observations  and  predictions  at   the  preceding   times   are  considered  in  the  model  to  produce  the  prediction  for  the  next  time  step;   x Model   inversion.  Observations  are  used   in  order   to  estimate  parameters  and/or   initial   conditions   considered  as  difficult  to  estimate  and  sensitive.   These  algorithms  have  been  applied  in  a  number  of  experiments  in  France  and  India  on  sugar  beet  and   wheat.   The   conclusions   are   that   remote   sensing   data   assimilation   into   crop   models   can   improve   significantly  the  predictions  of  crop  models  (Varella  et  al.,  2010).     Further   research   work   is   needed   to   determine   the   most   favourable   observation   configurations   to   estimate   parameters   as  well   as   finding   the   right   combinations   of   different   types   of   RS   observations.   INRA   has   great   expectations   for   data   availability   in   the   future   with   a   better   temporal   and   spatial   resolution  (e.g.  Sentinel  constellation).         Sola r TIR µ-w ave Leaf area index +++ + ++fAPAR +++Cover fraction +++ + +chlorophyll content +++water content ++temperature +++moisture + ++roughness + ++organic matter ++residues ++ Canopy structure Leaf characteristics Soil characteristics BIOPHYSICAL VARIABLES 26     3.5 Regional  Crop  Simulation  Modelling  for  Yield  Prediction  Using  Remote   Sensing  and  GIS:  Indian  Experiences   Vinay  Sehgal,  Indian  Agricultural  Research  Institute,  New  Delhi,  India   IARI  is  a  106  years  old  national  institute  in  agricultural  research  &  teaching  in  India,  instrumental  in  the   ͞'ƌĞĞŶƌĞǀŽůƵƚŝŽŶ͟ĂƐǁĞůůĂƐƚŚĞĨŝƌƐƚŶĂƚŝŽŶĂůŝŶƐƚŝƚƵƚĞƚŽŝŶƚƌŽĚƵĐĞZĞŵŽƚĞ^ĞŶƐŝŶŐŝŶĐŽƵƌƐĞƐ͘ZĞŵŽƚĞ Sensing  (RS)  can  be  of  use  for  crop  simulation  models  at  regional  scale.  It  may  provide  inputs  parameters   and/or  initial  conditions  as  well  as  improve  the  accuracy  of  the  model  results.  This  positive  effect  of  data   assimilation  techniques  is  due  to  correction  of  errors  in  the  structure  of  the  model  and  the  correction  of   the  growth  affecting   factors   like  pests,  diseases,  salinity  etc.  This  can  be  done  by  making  models  self-­‐ correcting  as  well  as  inserting  RS  measurements  directly  into  the  model  as  state  variables.   Data  assimilation  algorithms  used  at  IARI:   x Direct  use  of  the  driving  RS  variable  in  model;   x Forcing:  updating  of  a  state  variable  derived  from  RS  (e.g.  LAI);   x Re-­‐initialization:   adjustment   of   an   initial   condition   to   obtain   a   simulation   in   agreement   with   RS   derived  observations;   x Re-­‐calibration:  adjustment  of  the  model  parameters  to  obtain  a  simulation  in  line  with  RS;   x Corrective  method:  error  between  simulated  and  RS  derived  variable  to  correct  yield  values.   In  the  study  area  of  Thanjavur  MODIS  LAI  images  are  used  in  the  ORYZA1  model  (a  model  for  irrigated   rice  production)  to  get  an  estimate  of  the  phenological  stage  of  different  rice  classes.   In   another   study   the  WTGROWS  model   is   applied  with   the   forcing   and   re-­‐initialization   techniques   to   estimate  wheat  grain  yields  at   farmer   field   level.  Grain  yield  estimations   improved  considerably  using   this  approach.     In  India,  conventionally,  crop  forecasting  is  done  by  Crop  Cutting  Experiments  (CCE)  following  a  sampling   plan  that  varies  from  state  to  state,  and  the  results  are  then  aggregated  at  higher  administrative  units.   Recent  research  initiatives  have  improved  crop  forecasting  in  India,  namely:   x FASAL   (Forecasting   Agricultural   Output   Using   Space,   Agrometeorology   and   Land-­‐based   Observations).  FASAL   forecasts  different  crops  using  different   remote  sensing  data  combined  with   field  data.   x NADAMS  (National  Agricultural  Drought  Assessment  &  Monitoring  System).  NADAMS  uses  seasonal   NDVI  profiles  integrated  with  ground  information  to  estimate  crop  conditions.   Main  findings  (Sehgal  et  al.,  2011):   x Crop  models  need  to  describe  at  least  three  interdependent  systems:  canopy,  root  and  soil  system.   x Crop  management  practices  are  the  main  determinant  of  variability  in  crop  yields  at  small  scale.   x Remote   sensing   derived   crop   phenology   (sowing),   LAI   and   soil  moisture   assimilations   at  multiple   time  in  crop  season  are  a  possible  way  forward.     27     3.6 Integration  of  MODIS  products  and  a  crop  simulation  model  for  crop  yield   estimation   Gerrit  Hoogenboom,  Washington  State  University,  Prosser,  WA,  USA,  in  collaboration  with   Hongliang  Fang,  Institute  of  Geographic  Sciences  and  Natural  Resources  Research,  Chinese   Academy  of  Sciences,  Beijing,  China,  and  Shunlin  Liang,  University  of  Maryland,  College  Park   Maryland,  USA.   Washington  State  University,   in  cooperation  with  various  international  partners  has  developed  various   methodologies   for   crop  yield  estimation  with   remotely   sensed  data.   The  data  assimilation   techniques   used  were:   x Direct  input:  the  model  uses  remotely  sensed  data  as  inputs   ƒ Sequential    assimilation:  replace  state  variables  in  the  model  with  remotely  sensed  data   ƒ Variational   assimilation:   minimize   the   difference   between   variables   estimated   by   remote   sensing  and  crop  model  simulations   Crop  models  used  were:   ƒ DSSAT  (Decision  Support  System  for  Agro-­‐technology  Transfer)   ƒ CSM-­‐CERES-­‐Maize  and  CSM-­‐CERES-­‐Wheat  for  maize/corn  and  wheat   ƒ CSM-­‐CROPGRO-­‐Soybean  for  soybean   In  the  scenarios  5  different  remotely  sensed  datasets  were  tested  for  data  assimilation  into  the  models   above:  LAI  ,  EVI  (enhanced  vegetation  index),  NDVI,  EVI  &  LAI  and  NDVI  &  LAI,  all  derived  from  MODIS   imagery.  In  the  various  models  these  data  were  used  to  adjust  planting  date,  planting  population,  row   spacing  and  fertilization  date  and  amount.  These  methods  were  applied  on  a  regional  scale  for  the  state   of   Indiana   in  the  United  States  (Fang  et  al.,  2011).  These  studies   showed  that  regional  crop  yields  can   successfully   be   estimated  with   the   data   assimilation   approach,   whereby   the   combination   of   LAI   and   NDVI   performed   best   against   measured   USD   NASS   corn   yield   data.   Furthermore,   a   method   was   developed  to  aggregate  the  individual  maize  yields  to  county  level.  See  maps  below.   The   conclusions   are   that  MODIS   products   are   useful  for  crop  yield  estimation  at  the  regional   scale.  Field  hydrological  conditions  can  also  be   simulated  successfully  with  this  approach.  The   methodology  could  be  improved  if  new  remote   sensing  products  (e.g.,  crop  percentage  at  250   meter  grid  size)  would  become  available.     28     3.7 Exploring  the  Response  of  the  Central  US  Agro-­‐Ecosystem  to  Climate   Change   Darren  Drewry,  NASA  Jet  Propulsion  Laboratory  /  Caltech  Climate  Physics  Group,  Pasadena,  CA,   USA   Rapid  changes   in   the  earth  atmosphere  have  been  detected   in   the   last  hundred  years.  This  applies  to   carbon  dioxide,  methane,  nitrous  oxide  and  sulphur.  At  the  same  time  scale,  global  warming  has  been   detected  with  an  average  increase  of  about  1  degree  Celsius.  A  relation  between  the  two  phenomena  is   plausible.   An   assumption   is   often   made   that   higher   CO2   levels   lead   to   higher   plant   production,   as   one   of   the   growth  factors   is  available  at  a  higher  rates.  However,  some  scientists  doubt   this,  as  plants  appear   to   close   the   stomata   under   influence   of   elevated   CO2   levels.   A   Free   Air   Carbon   Enrichment   (FACE)   experiment  has  been  conducted  to  find  evidence   of  this  effect  (see  photo).   Field   experiments   have   been   done   with   maize   (C4)  and  soybean   (C3)  whereby  CO2   levels   in   the   crop   were   elevated   with   a   ring   tube   providing   extra   CO2   directly   in   the   field.   For   these   experiments,   a   multi-­‐layer   canopy-­‐root-­‐soil   system   model   (MLCan)   capable   of   accurately   predicting   canopy-­‐atmosphere   exchange   of   CO2   has  been  set  up.  PAR,  NIR,  LW,  U,  Ta  and  Ea  were   measured  from  a  tower  construction  in  the  field.   The  CO2  application  causes  higher   temperatures   inside   the  maize  canopy.  Furthermore,   in   soybean,  a   10%   higher   LAI   was   observed   related   to   an   increase   in   photosynthetic   substrate.   In   soybean,   the   stomata  were  recorded  to  have  a  more  closed  state.   In  maize  these  effects  were  not   found,  although   some  ecophysiological  acclimation  was  recorded.  Regarding  the  canopy  responses  to  elevated  CO2  the   following  conclusions  were  drawn  (Drewry  et  al.,  2010):   x Modelled   gas   exchange   and   leaf   state   responses   to   elevated   CO2   are   in   good   agreement   with   SoyFACE  observations  for  both  C3  and  C4  crops  (soybean  and  maize);   x Net   canopy   CO2,   latent   energy   and   sensible   heat   were   increased   by   24%,   decreased   by   5%   and   increased  by  37%  at  mid-­‐day  for  soybean.    Structural  acclimation  &  increase  in  substrate  availability   offset  much  of  the  effect  that  would  have  otherwise  occurred;     x Net   canopy  CO2,   latent  energy  and   sensible  heat  were   increased   by  1.5%,  decreased  by  16%  and   increased  by  47%  at  mid-­‐day  for  corn;       x Structural   acclimation   results   in   greater   shortwave   energy   absorption   in   light-­‐saturated   upper-­‐ canopy;   x In  both  crops  there  is  a  negligible  impact  of  carbon  enrichment  on  carbon  uptake.   29     3.8 Crop  Yield  Forecasting  Over  Various  Scales  Combining  Models  and   Remote  Sensing   Wilco  Terink,  FutureWater,  Wageningen,  Netherlands   Crop  growth  models   can  be  applied  at  various   scales   ranging   from   field   to  basin  and  country   to  even   continent.   Some  models   are  more   suitable   to   be   applied  at  field  scale,  while  other  models  are  more   suitable   to  be  applied  at   the  basin  or   continental   scale.   FutureWater   uses   (agro)hydrological   simulation   models   that   are   in   the   opensource   domain.   For   each   project,   FutureWater   determines   which   model   is   most   suitable,   given   the   spatial   and   temporal   scales   at   hand   (see   graphic).     An   important  aspect   is  formed  by  the  trade-­‐off  between  the  physical  model  detail   (model  complexity)   and   the   availability   of   the   required   data.   As   expected,   field  scale  models  have  generally  more   physical   detail   than   continental   models,   where   input   data   are   usually   scarce   and   more   assumptions  have  to  be  made.  Similarly,  in  remote   sensing,   stratification   is   possible,   based   on   temporal   frequency   and   spatial   resolution   (see   graphic).     FutureWater  conducts  projects  at  various  scales  as   mentioned   above.   In   Egypt   it   evaluated   the   (agro)hydrological   performance   of   an   farm-­‐level   (field   scale)   irrigation   improvement   project..   In   India   the   company   calibrated   at   the   basin   scale   a   hydrological  model  with  remotely  sensed  evapotranspiration   (Immerzeel  and  Droogers,  2008).  For   the   22   countries   in   the  Middle   East   and   North   Africa   (MENA   region),   FutureWater   performed   a   water  supply  and  demand  analysis  for  the  period  2010-­‐2050.   In  every  project  choices  have  to  be  made  again.  The  selection  of  models  and  data  is  determined  by:   x The  trade-­‐off  between  required  resolution,  available  resolution,  and  costs   x High  resolution  inputs  are  almost  always  needed  for  calibration  and  correcting   Regarding  remote  sensing  data  there  is  a  movement  from  optical  to  radar  based  imagery  in  order  to  be   able  to  look  through  clouds.     30     3.9 On  the  Assimilation  of  Remote  Sensing  Data  with  Crop  Models  for  Crop   Yield  Forecasting   Amor  Ines,  International  Research  Institute  for  Climate  and  Society  (IRI),  Palisades,  NY,  USA   Uncertainty   is   one   of   the   largest   challenges   when   predicting   crop   yields.   This   uncertainty   can   be   attributed  to  both  models  and  climate  data.  Uncertainty  is  highest  early  in  the  crop  cycle  and  generally   diminishes   towards   the   harvest.   Model   uncertainty   can   be   lowered   partly   by   data   assimilation   techniques,   while   climate   uncertainty   can   be   decreased   by   using   seasonal   climate   forecasts,   (e.g.   as   produced  by  IRI).  Although  seasonal  forecasts  have  a  low  temporal  and  spatial  scale,  they  have  proven   to  be  useful  in  projects  like  the  WFP  Africa  Risk  Capacity  project  to  reduce  the  uncertainty  a  few  months   before   harvest.   See   example   of   the   scale   of   the   seasonal   forecasts  in  map.   Generally   crop   models   work   best   in   large-­‐scale,   homogeneous   agro-­‐ecosystems.   However,   for   complex,   heterogeneous   agricultural   systems   in   the   context   of   smallholder   agriculture   in   developing   countries,   the     fractions  of  target  crops  grown  are  usually  small.     Innovations  are  especially  needed  in:   x Un-­‐mixing   RS   vegetation   signature,   which   should   result   in  signals  for  different  crops  rather  than  a  mixture  of   crops  and  natural  vegetation;   x Promising  RS  soil  moisture  data  are  available  (based   on   radar),  but   the   scale   is   still   too  coarse   (both  spatial  and   temporal)  for  most  modelling  applications.     Crop   monitoring   and   yield   forecasting   have   been   investigated   for   the   continental   United   States   (especially   Georgia  and   Iowa)  using  DSSAT-­‐CSM  crop  models   in  combination  with   soil  moisture  and  LAI  products   from   remote   sensing.   Using   AMSR-­‐E   soil   moisture   data   with   the   Kalman   filter,   did   not   lead   to   significantly  better   results,  while   the  use  of  MODIS  LAI  did  have  a  positive  effect  on  accuracy  of  yield   data  against  USDA  yield  figures.   Conclusions  of  research  are  (Ines  et  al.,  2012):   x Regarding  the  Ensembles  Kalman  filter  performance  applied  to  DSSAT-­‐CSM  crop  models,  the  value   of  data  assimilation  with  climate  forecasts  is  more  evident  later  in  the  growing  season.   x The  skills  of  climate  forecasts  is  most  important  in  the  early  part  of  the  growing  season.   x The  availability  of  downscaled   remotely   sensed  soil  moisture  and  LAI  data  would  make  modelling   considerably  more  accurate.   x Using  both  soil  moisture  +  LAI  gave  better  results  compared  to  using  them  independently  in  the  data   assimilation  -­‐  possibly  due  to  the  interaction  of  the  two  in  the  simulations.     31     3.10 Simultaneous  Estimation  of  Model  State  Variables  and  Observation  and   Forecast  Biases  using  a  Two-­‐Stage  Hybrid  Kalman  Filter   Valentijn  R.N.  Pauwels,  Laboratory  of  Hydrology  and  Water  Management,  Ghent  University,   Ghent,  Belgium   In  Earth  sciences  data  assimilation  is  defined  as  the  updating  of  modelled  state  variables  using  external   datasets.  Therefore,  in  theory,  one  could  simply  replace  model  results  by  observations.  In  practice  this  is   not  a  good  approach  because:   x Both  the  external  data  and  the  model  results  contain  errors.   x Many  times  a  proxy  of  the  state  variables  is  assimilated,  and  not  the  state  variable  itself   x Almost  always  one  needs  to  update  many  (unobserved)  state  variables  using  only  one  or  a  couple  of   observations.   Therefore  more  complicated  methods  for  data  assimilation  are  developed  of  which  the  most  popular  is   the  Kalman  filter.  In  the  original  Kalman  Filter  (1960),  the  state  variables  and  observations  are  assumed   to  be  unbiased.   It  uses  a  model  state-­‐space  representation  of   the  system  whereby  the  state  variables   are  mapped  onto  observation  space.   In   an   example   with   soil   moisture,   the   modelled   soil   moisture   is   expressed   as   a   volumetric   fraction   (between  0  and  1).  The  fact  that  the  observations  are  in  percentages  shows  that  state  variables  can  be   ƵƉĚĂƚĞĚ ƵƐŝŶŐ ĂŶLJ ŽďƐĞƌǀĂƚŝŽŶ ƚŚĞLJ ĂƌĞ ƌĞůĂƚĞĚ ƚŽ͘ /Ŷ ƚŚŝƐ ĂƉƉƌŽĂĐŚ͕ ƚŚĞ ͞<ĂůŵĂŶ 'ĂŝŶ͟ ďĞĐŽŵĞƐ Ă weighting  factor  between  the  observation  error  and  the  model  error.   The   Kalman   filter   has   been   designed   for   linear,   un-­‐biased   systems.   Unfortunately   many   data   sets   (especially   remote   sensing   data)   assimilated   into   hydrologic  models   contain   a   significant   bias.   Many   studies   remove   the   bias   before   the   assimilation   by   removing   the   long-­‐term   difference   between   the   model  and  the  external  data.  Since  models  contain  bias  as  well,  this  may  not  be  the  optimal  strategy.   Therefore  a  refinement  of  the  Kalman  filter  has  been  developed  by  Evensen  (1994),  which  is  called  the   Ensemble   Kalman   Filter.   It   enables   the   assimilation  of   external   data   into   nonlinear   biased   systems.   It   essentially   estimates   forecast   and   observation   biases   together   with   the   model   state.   An   essential   assumption  is  that  the  observation  and  forecast  bias  errors  are  independent  of  each  other  and  are  also   independent  of  the  system  state  errors.   Crop   models   tend   to   very   complicated   (many   processes   combined   in   one  model),   and   need   a   wide   variety  of  data  sets,  model  parameters,  and  meteorological   forcing.  This  can   lead  to  both  random  and   systematic  errors  in  the  model  results.     Unfortunately,   straightforward   application   of   data   assimilation   techniques   leads   to   both   random   and   systematic  errors.  Therefore,  if  bias  is  apparent  in  observations  and/or  model,  this  has  to  be  taken  into   account  to  get  meaningful  model  outputs  (Pauwels  and  De  Lannoy,  2009).       32     3.11 Satellite  image  simulations  for  data  assimilation  at  multiple  scales   Heike  Bach,  VISTA  Remote  Sensing  in  Geosciences,  Munich,  Germany   VISTA  Remote  Sensing  is  a  private  company  (SME)  situated  in  Munich,  Germany  (www.vista-­‐geo.de).  Its   main   expertise   is   in   remote   sensing   applications   in   hydrology   and   agriculture.   Vista   works   in   close   connection  with  the  University  of  Munich  with  the  aim  to  bring  (crop)  science  to  practice.  For  farmers,   Vista  develops  satellite  techniques  for  precision  farming  in  Europe  and  Russia  (www.talkingfields.de).   PROMET  (Mauser  &  Bach  2009)  and  SLC  (Verhoef  &  Bach  2012)  are  land  surface  models  that  couple  a   crop   growth   model   with   a   radiative   transfer   model   offering   simulated   satellite   images   that   can   be   compared   to   real  ones   for  data  assimilation  purposes.   SLC  uses   structural,   spectral   and  observational   input   data.   PROMET   is   raster-­‐based,   and   produces   a   completely   closed   water   and   energy   balance.   Management  practices  such  as  sowing  date  and  harvest  date  can  be  fed  onto  the  system.     On   a   field   scale,   PROMET/SLC   has   been   used   to   predict   wheat   yields   in   a   large-­‐scale   farm   in   Germany,   producing   high   resolution   (20   m)   output,   which   fit   very   well   with   measurements   from   combine   harvester   recordings   (see   graphic   from  Hank  et  al.,  2012).   In   a   meso-­‐scale   study   encompassing   the   Upper   Danube   Watershed   (76000   km²)   surface   temperatures  were  calculated  with  PROMET   that   correlated   very   well   to   similar   NOAA-­‐AVHRR   extracted  temperature  data  at  a  resolution  of  1  by   1  km.       In   a   macro-­‐scale   study   for   Central   Europe   (1.36   million   km²)   the  MM5  model   (model   to   simulate   or  predict  atmospheric  circulation)  was  combined   with  PROMET,  where  the  45km  pixels  of  the  MM5   model  were  successfully  combined  with  the  1  km  PROMET  model  to  deliver  an  estimate  for  the  average   annual  evapotranspiration  (Zabel  et  al.  2012).       The   above-­‐presented   examples   show   that   high   resolution   satellite   images   now   allow   observing   the   current  crop  status  at  various  scales.  The  heterogeneities  of  the   land  surface  can  thus  be  captured.  By   assimilation   of   satellite   data,   improved   modelling   of   the   water   and   carbon   cycle   can   be   achieved.   PROMET  is  capable  of  predicting  crop  yields  at  field  scale,  meso-­‐scale  and  even  macro-­‐scale  using  the   same  physical  principles  and  procedures.         33     3.12 MARS  operational  crop  monitoring  and  yield  forecasting  activities  in   Europe   Gregory  Duveiller,  AGRI4CAST  Action,  MARS  Unit,  JRC,  Ispra,  Italy   The   European   Commission   requires   in-­‐season   crop   yield   forecasts   at   a   European   level   as   part   of   the   decision  making  process  on  market  intervention  and  for  policy  support.  For  the  past  twenty  years,  the   Monitoring  Agricultural  Resources  (MARS)  Unit  of  the  European  Commission  Joint  Research  Centre  (JRC)   has   operationally   produced   such   forecasts   for   European  member   states   and   for   countries   in   the   EU   periphery   in   a   tight   monthly   schedule.   This   is   done   using   the   MARS   Crop   Yield   Forecasting   System   (MCYFS),  a  modelling  infrastructure  driven  by  agro-­‐meteorological  data  and  assisted  by  remotely  sensed   observations.  The  MCYFS  is  a  decision  support  system  driven  by  expert  knowledge  and  relying  on  four   main  data  infrastructures:  a  meteorological  data  infrastructure,  a  remote  sensing  data  infrastructure,  a   crop  simulation   infrastructure  and  a   statistical   infrastructure.  The  system  uses  meteorological  data   to   run   crop   growth   models   that   provide   information   on   crop   status,   such   as   biomass   production,   soil   moisture   or   biomass   of   the   storage   organs.   Remote   sensing   provides   an   independent   assessment   of   crop  status  through  the  use  of  global  and  pan-­‐European  low-­‐resolution  imagery  in  near  real-­‐time  (NRT).   Finally,  the  statistical  infrastructure  includes  methods  used  to  analyse,  along  the  season,  historical  yield   records  against  the  information  about  crop  status  generated  by  crop  models  to  produce  a  forecast  that   is  presented  in  a  monthly  bulletin  to  decision-­‐makers   in  Brussels.  Of  course,  the  team  of  analysts  that   needs   to   decide  what   is   the   most   adequate   information   to   base   the   forecast   upon   is   the   keystone   of   this   approach,   and   is   supported   by   a   skilled   IT   team.   The   system   is   articulated   by   a   spatial   framework   defining   the   spatial   reference   upon   which   all   the   data   is   generated   (reference   grids,   administrative   units,   static   spatial   layers   used   by   crop   models   and   remote   sensing,   etc.).   Some  research  questions  that  are  currently  being  investigated  or  that  are  foreseen  in  the  coming  future   include  the   following:   (1)  using  remote  sensing     to  provide   improved  crop  calendars  which  could  help   recalibrate  models  better   in   the   crop  modelling   infrastructure;   (2)  develop  a  method   to   identify  pure   enough   crop   specific   time   series   from   MODIS   that   can   be   used   from   crop   growth   monitoring;   (3)   exploiting   global   solar   radiation  LANSAF  products   (derived   from  MSG)  as   input   to   the   crop  models   to   produce  simulations  of  better  quality.     34     3.13 Experiences  with  data  assimilation  for  regional  crop  yield  forecasting   Allard  de  Wit,  Alterra,  Wageningen,  The  Netherlands   The   result   of   a   study   investigating  whether   data   assimilation   techniques   could   improve   regional   crop   yield  forecasting  for  Europe  was  presented.  A  study  area  was  selected  in  the  Walloon  area  of  Belgium   and   in   Northern   France.   In   this   area   wheat   is   a   dominant   crop.   The  WOFOST   crop   yield  model   was   applied  on  these  areas  on  a  10x10  km  grid  scĂůĞ͘dŚĞŵĞƚŚŽĚŽůŽŐLJŝŶǀŽůǀĞĚƚŚĞŝĚĞŶƚŝĨŝĐĂƚŝŽŶŽĨ͞ƉƵƌĞ͟ wheat   pixels   (see   graphics)   using   LAI   temporal   profiles.   MODIS   GAI   ingestion   was   used   for   selected   wheat  pixels  for  the  years  2000  to  2009.   The   research   involved  heavy   quality   control   on  MODIS  GAI   after  which   the  GAI   data  were   applied   in   WOFOST   using   parameter   optimization.   Finally   results   were   validated   with   the   EUROSTAT   regional   statistics.   Limitations   to   the   use   of   the   Kalman   Ensembles   data   assimilation   technique   were   found.   As   EnKF   originates   from   meteorology   and   oceanography   (Evensen  1994)  it  works  best  with   integration   of   rates   of   change   according   to   atmospheric   physics/hydrodynamics.   However,   crop   models   have   two   processes   running:   growth   and   phenology.   Phenology   can   be   seen   as   a   parallel   controlling   process   that   complicated   the   application  of  the  EnKF  filter.   The  main  conclusions  of  the  research  are  (de  Wit  et  al.,  2012):     x The   Ensemble   Kalman   filter  must   be   applied  with   care.   It   proved   to   be   suitable   for   soil  moisture   assimilation,  where  there  is  no  phenology  effect.   x The  data  assimilation  recalibration  strategy  seems  more  suitable  in  general  for  assimilating  canopy   variables  although  crop-­‐specific  estimates  are  needed  (no  mixed  pixels).   x MODIS   GAI   estimates   have   shown   to   be   very   noisy   in  W-­‐Europe,   as   a   result   of   the   high   level   of   landscape  fragmentation.    Post-­‐processing  and  quality  control  are  very  important.   x MODIS  GAI  estimates  have  demonstrated  to  be  useful  in  updating  crop  model  parameters.  One  of   the   findings   was   that   the   inter-­‐annual   variability   in   the   distributions   of   the   optimized   model   parameters  was  larger  than  expected.       35     3.14 Crop  Monitoring  and  Early  Warning  Service  in  Africa   Bakary  Djaby,  University  of  Liege,  Arlon,  Belgium   University   of   Liege   (ULg)   plays   an   important   role   in   the   GMFS   (Global  Monitoring   for   Food   Security)   project,   funded  by   the   European   Space  Agency   (see  www.gmfs.info).   The  main   partners   in  Africa   are   located  in  Sudan,  Malawi  and  Niger  (AGHRYMET).   The  ULg  aims  at  improving  early  warning  services  with  quantitative  estimates  of  crop  yield  and  pasture   biomass  in  two  regions:   x West  Africa:  Development  of   crop   yields   forecast  models  using   remote   sensing  data   in  Niger   and   Senegal;   x East   Africa   and  West   Africa:   improvement   of   Livestock   Early  Warning   systems   products   in  West   Africa  (Niger  and  Senegal)  and  East  Africa  (Ethiopia)   In   GMFS   Phase   2,   the   focus   of   the   project   has   been   on   intensive   training   of   users   in   Africa,   and   the   integration  of  model  results  into  the  countries  food  security  bulletins.  The  model  used  is  the  FAO  Crop   specific  soil  water  balance  model  (CSSWB),  which  has  been  implemented  in  the  software  AgrometShell.   Input  data  differ  from  application  to  application,  but  in  general  these  datasets  are  used:   x Rainfall  estimates  from  stations  and  ECMWF  era  interim  reanalysis  data   x Remote  sensing  NOAA-­‐AVHRR  GAC,  SPOT  and  MERIS  imagery  (NDVI,  Fapar  and  DMP)   x Land  use  data:  LULC/  Globcover  and  FAO  crop  calendars.   x National  Statistics  for  production  data   Remote   Sensing   data   are   used   to   assess   the   seasonality   (sowing,   harvest).   For   this,   time   series   are   analyzed  with   the  adaptive  Savitzky-­‐Golay   filtering  method.  From  this   fitted  model   the  beginning  and   end  of  the  growing  season  can  be  extracted.  In  a  next  step,  these  data  are  input  into  the  water  balance   calculations.   For  validation  extensive   field   surveys  have  been  conducted   in  Niger  and  Senegal   for  5  years.   In  Niger   18343  fields  with  millet,  8548  fields  with  sorghum  and  1791  fields  with  maize  In  Senegal  3122  fields  with   millet  and  2743  fields  with  peanuts.  The  models  are  validated  by  leaving  part  of  the  input  data  out  of  the   calculations  and  check  calculated  values  against   input  data   later.     If   little  data  are  available   leave-­‐one-­‐ out  cross  validation  techniques  are  used  as  well  as  resampling  (bootstrap).   Some   difficulties   were   experienced   with   the   accuracy   of   climate   information,   e.g.   ECMWF   versus   country  station  data  and  uncertainties  in  land  use  and  country  statistical  figures.     Developments  planned  in  the  near  future:   x Integration  of  SAR  soil  moisture  data;   x Two  forecast  periods  in  Niger  and  Senegal  for  this  season  (August  and  September);   x Comparative  study  of  USGS  WRSI  input  versus  AgrometShell  Water  balance  and  impact.   36     3.15 Data  Assimilation  based  on  the  Integration  of  Satellite  Data  and  Field   Sensor  Data  for  Drought  Monitoring   Kiyoshi  Honda͕/Ŷƚ͛ůŝŐŝƚĂůĂƌƚŚƉƉůŝĞĚ^ĐŝĞŶĐĞZĞƐĞĂƌĐŚ͕ĞŶƚĞƌ  (IDEAS),  Chubu  Institute  for   Advanced  Studies,  Chubu  University,  Japan   Chubu  University  develops  methods  for  crop  model  calibration  based  on  the  Integration  of  satellite  data   and   field   sensor   data.   In   an   effort   to   standardize   and  have   systems   communicate   easily,   cloud-­‐based   web  services  have  been  developed  to  dissimilate  field  sensor  data.   The  Field  sensor  network  cloudSense  is  based  on  small  and  low-­‐cost  sensors   that  provide  data  through  mobile  Internet  communication.  Potentially  these   sensors   can   gather   information   in   real-­‐time   from   anywhere   in   the   world.   Possible   applications   are:   disaster   preparedness,   agriculture,   logistics,   security,  etc.     As  the  sensor  network  is  essentially  open  source,  anyone  can  add  a  sensor  to   the  network.  A  simple  protocol  based  on  an  input  form  needs  to  be  filled  in   order  to  add  the  sensor  to  the  network.   For   analysis   and   visualisation,   applications   are   developed   for   mobile   phones   and   various   computer   operating  systems.  One  of  the  applications  aims  at  fostering  confidence  in  food  safety  among  consumers   It  essentially  displays  crop   information   to  end  users  of   the  crop,  while   the  crop   is   still  on   the   field.   In   another  application  greenhouse  gasses  (CH4  and  N2O)  are  measured  and  visualised  in  Thailand.  Sensors   are  fitted  onto  fixed  poles  as  well  as  low-­‐ĐŽƐƚŚĞůŝĐŽƉƚĞƌƐĂŶĚŽƚŚĞƌhs͛Ɛ͘   Remote  sensing  data  can  be  used  in  crop  models  through  data  assimilation.  However,  remote  sensing   generally   provides   just   a   few   parameters   such   as   LAI,   Eta   etc.   Important   parameters   such   as   soil   hydraulic  parameters,  sowing  date  etc.  are  difficult  to  base  on  satellites.  Field  sensor  data  fill  this  gap.  As   an  example,  in  Thailand,  rice  is  frequently  damaged  by  dry  spells.  The  damage  is  assessed  in  real-­‐time  by   running   the   SWAP   model   assimilated   with   remote   sensing   and   field   sensor   data.   This   research   has   shown  that  low  October  rainfall  has  the  highest  adverse  impact  on  rice  production.   Field   Sensor   data   have   been   successfully   used   to   correct   MODIS   LAI   data,   as   MODIS   LAI   generally   underestimates   the   LAI   on   the   ground.   The   satellite   LAI   was   calibrated   with   ground   measurements   before  it  was  used  in  the  assimilation  process.   Measured   soil   moisture   information   is   very   valuable   as   assimilated   input   into   crop   models.   As   this   cannot  be  done  with  satellite  measurements,  a  ground  sensor  network  is  proven  to  be  very  helpful.   Generally  low-­‐resolution  satellite  data  are  available  at  a  high  frequency,  while  high-­‐resolution  data  are   available   at   low   frequency.   With   an   algorithm   developed   at   Chubu   University,   both   sources   can   be   combined  into  a  more  valuable  source  of  data.  As  an  example,  high  resolutions  LANDSAT  /  ASTER  data   have  successfully  been  combined  with  low  resolution  AVHRR  /  MODIS  data  (Ines  and  Honda,  2005).     37     3.16 Data  assimilation  for  the  carbon  cycle  in  Sudan  savannah  smallholder   communities   Pierre  Traoré,  ICRISAT,  Bamako,  Mali   Stable  soil  organic  carbon  (SOC)  plays  an  important  role  in  soils  while  it  retains  water  and  improves  the   structure   of   the   soil.   Increasing   SOC   contents   in   the   soil   could   also   potentially   help   reduce   the   CO2   content   of   the   air.   These   are   long-­‐term   processes   prove   difficult   to   quantify.   ICRISAT   took   up   the   challenge  and  used  the  DSSAT  model  (DSSAT-­‐CENTURY)  together  with  field  measurements  and  remote   sensing  to  quantify  the  carbon  cycle.     This  was  applied  in  Sudanian  agricultural  systems  in  Southern  Mali,  Burkina  Faso  and  Ghana  (see  map).   These  areas  have  heterogeneous  management  techniques  and  quite  extensive  mixed  cropping  practices,   often   with   low-­‐yielding   traditional   varieties.   The   most   important   crops   were   maize,   yam,   millet,   sorghum  and  peanut.    Even  within  a  crop  like  sorghum,  8  to  10  different  varieties  have  been  identified   that  react  differently  to  management  practices.   Information   on   the   very   detailed   cropping   patterns   was   obtained   though   high-­‐ resolution   imagery   in   combination   with   field   work   (based   on   QuickBird   NDVI   anomalies).   In   time,   SOC   measurements   and   model   outcomes   have   been   studied   at   both   point   level   and   aggregated   to   areas,   where   the   aim  was   to  minimize   uncertainty.   At   point   level   simplified   DSSAT  simulations  of  SOC  have  been  assimilated  with  field  measurements  using  the  Ensemble  Kalman   Filter.     At  point-­‐level  (Jones  &  al.,  2004,  2007;  Koo,  2007),  using  the  EnKF  reduced  measurement  uncertainty  by   around  60%.  Furthermore,  over  space  the  EnKF  reduced  uncertainty  by  50%,  although  results  proved  to   be  very  sensitive   to   initial  estimates  of  parameters.   In  other  words,   there   is  uncertainty  on  departure   from  steady  state  as  well  as  uncertainty  on  planting  dates.   Besides   this   research,   ICRISAT   is   instrumental   in   the   worldwide   AgMIP   project.   This   is   a   distributed   climate-­‐scenario   simulation   exercise   for   historical   model   intercomparison   and   future   climate   change   conditions  that  goes  further  than  just  crop  modelling.  Many  crop  and  agricultural  economics  modelling   groups   around   the   world   are   contributing.   The   goals   of   AgMIP   are   to   improve   substantially   the   characterization   of   risk   of   hunger   and   world   food   security   due   to   climate   change   and   to   enhance   adaptation  capacity  in  both  developing  and  developed  countries.   38     3.17 Soil-­‐water-­‐crop  modelling  for  decision  support  in  Sub-­‐Saharan  west   Africa:  experiences  from  Niger  and  Benin   Pierre  B.  Irénikatché    AKPONIKPÈ,  Faculty  of  Agronomy,  University  of  Parakou,  Benin   The  Sahel  region   in  West  Africa  suffers  from  low  grain  yields  (millet  yield  often   lower  than  500kg/ha),   caused   by   limited   and   uncertain   rainfall   (300-­‐600mm   per   year)   compounded   by   low   soil   fertility.   Although   numerous   improvements   have   been   proposed   over   the   years,   the   impact   of   agricultural   research  is  still   low.  Small  scale  farmers  rarely  adopt  new  management  methods  and  inputs.  The  main   reason  seems  to  be  that  farmers  seek  to  reduce  risk  while  scientists  try  to  increase  yields.     The   University   of   Parakou   in   Benin   has   investigated   this   phenomenon.   It   has   studied   climate   risk   management   in   S-­‐W   Niger   where   a   high   temporal   rainfall   variability   is   normal   (annual   coefficient   of   variance  of    17  to  36  %,  even  78%  at  a  daily  basis).  There  is  also  a  high  spatial  rainfall  variability.  Farmers   seem  to  adapt  to  the  spatial  variability  by  dispersing  their  fields  within  the  village  territory.     The   University   set   out   to   investigate   the   hypothesis   whether   farmers   disperse   their   fields   to   reduce   agro-­‐climatic  risk.   ͞ŚŽƵƐĞŚŽůĚĨŝĞůĚĚŝƐƉĞƌƐŝŽŶŝŶĚĞdž͟ŚĂƐďĞĞŶĚĞǀĞůŽƉĞĚƚŽƚĞƐƚƚŚĞŚLJƉŽƚŚĞƐŝƐ͘dŚŝƐŝŶĚĞdžŝƐƐĞŶƐŝƚŝǀĞƚŽ the  distance  between  fields,  but  independent  of  the  number  of  farms  in  the  village  as  well  as  the  total   farm  ĂƌĞĂŽĨƚŚĞĨĂƌŵĞƌ͘&ƵƌƚŚĞƌŵŽƌĞĂ͞LJŝĞůĚŝŶƐƚĂďŝůŝƚLJŝŶĚĞdž͟;ƚŽŵĞĂƐƵƌĞŝŶƚĞƌ-­‐annual  variation  of  the   ŚŽƵƐĞŚŽůĚͿĂŶĚĂ͞LJŝĞůĚĚŝƐƉĂƌŝƚLJŝŶĚĞdž͟;ƚŽŵĞĂƐƵƌĞƚŚĞŝŶƚĞƌ-­‐annual  variation  of  yields  relative  to  the   village  area)  were  constructed.  Soil  fertility  gradients  were  taken  into  account.  Closer  to  the  village  soil   fertility  is  usually  higher.     The  main  conclusions  were  as  follows  (Akponikpe  et  al.,  2011):   ƒ There  is  no  relation  between  cumulated  annual  rainfall  and  yield  (see  graph);   ƒ Large  spatial  rainfall  variability  generates  an  even  larger  spatial  variability  in  yields;     ƒ Field  dispersion,  as  practiced  by  farmers  in  western  Niger,  allows  to  mitigate  inter-­‐annual  yield   variability  at  the  household  level,  albeit  to  a  limited  extent.   A   second   study   was   carried   out   in   Northern   Benin   investigating  the  optimal  amount  of  nitrogen  that  can   be   applied   to   farmers   fields.,   the   current   recommendation  being  30  kg  per  ha.   The   University   found   that   grain   yields   were   considerably   lower   than   those   assumed   with   the   recommendation   above.   In   part   this   is   explained   by   farmers   using   un-­‐improved     varieties   of   millet.   The   study  concludes  that  that  around  15  kg  of  nitrogen  per  ha  is  the  best  optimum.       39     3.18 Wheat  yield  modelling  in  a  stochastic  framework  within  and  post  season   yield  estimation  in  Tunisia   Eduardo  Marinho  and    Michele  Meroni,  FOODSEC  Action,  MARS  Unit,  JRC,  Ispra,  Italy   As  it  is  not  possible  to  directly  measure  and  model  grain  yields  production,  it  is  assumed  that  grain  yields   are   highly   correlated   to   biomass   yields.   Three   proxies   for   wheat   biomass   production   and   different   statistical  modelling  solutions  have  been  investigated  for  Tunisia.  The  aim  was  to  select  the  proxy  and   statistical   model   providing   the   best   predictive   capacity   in   yield   estimation   avoid   over/under-­‐ parameterization.   The   study   area   encompassed   10   governorates   representing   88%   of   national   production   of   wheat   in   Tunisia.  The  remote  sensing  data  used  were  13  years  of  SPOT-­‐VGT  fAPAR  &  NDVI  as  well  as  area  fraction   masks   for   cereals   from   aerial   photographs.   National   yield   statistics   were   available   on   the   level   of   governorates.   The  biomass  proxies  tested  were  (1)  NDVI  and  (2)  fAPAR  at  a  given  dekad,  and  (3)  the  Integral  of  fAPAR   during  the  period  of  plant  activity,  юĨWZ.  The  start  and  end  of  the  season  have  been  extracted  pixel  by   pixel   from   the   fAPAR   time   series,   analyzing   the   shape   of   the   curve   and   setting   a   priori   percentage   thresholds.   The   relation   between   these   proxies  and  the  final  grain  yield  was  assumed   to   be   linear   and   it   was   modelled   under   different   statistical   assumptions   (see   figure).   All   the   models   have   been   assessed   through   Jackknife   technique,   leaving   one   year  out   at   time.  It  proved  to  be  important  to  couple  the   phenology   of   the   crop   to   the   timing   of   the   remote  sensing  imagery  used.  In  this  study,  if   no   phenological   information   is   extracted   from   the   imagery   itself,   the   end   of   April   imagery  proved  to  deliver  the  best  results.  The  most  important  findings  are:   x High   yield   variability   in   Tunisia   can   be   estimated   by   remote   sensing   techniques,   without   the   involvement  of  a  crop  model;   x Improved   statistical  models   (i.e.,   fixed   and   random  effect)   have   a   significantly   positive   impact   on   yield  accuracy  estimation;   x In  Tunisia,  юĨWZoutperforms  other  biomass  proxies  for  yield  estimation;   x /ŶƚŚĞĂďƐĞŶĐĞŽĨŐƌŽƵŶĚĚĂƚĂ͕ƚŚĞюĨWZŝƐƚŚĞďĞƐƚŽption  for  measuring  crop  yields  because  it  is   linearly  related  to  pooled  yield  data  (no  distinction  among  governorates);   x Finally,   the   role   played   by   data   scarcity   in   determining   the   most   suitable   approach   for   yield   estimation  was  addressed.  The  trade-­‐off  between  the  ability  of  modeling  regional  specificities  and   over-­‐parameterization  has  been  emphasized  in  the  case  of  a  reduced  sample  size.  Results  indicate   that   the   selection   of   the   model   specification   should   take   into   account   the   number   of   available   observations,  and  not  only   the  expected  spatial  heterogeneity  on   the  yield-­‐biophysical  parameter   relationship.     40         41     4 References     Akponikpe,  P.B.I.,  J.  Minet,  B.  Gerard,  P.  Defourny,  and  C.L.  Bielders,  2011.  Spatial  fields  dispersion  as  a   farmer  strategy  to  reduce  agro-­‐climatic  risk  at  the  household  level  in  pearl  millet-­‐based  systems  in  the   Sahel:  a  modelling  perspective.  Agricultural  and  Forest  Meteorology,  151:  215-­‐227.   de   Wit,   A.,   Duveiller,   G.,   Defourny,P.,   2012,   Estimating   regional   winter   wheat   yield   with   WOFOST   through  the  assimilation  of  green  area  index  retrieved  from  MODIS  observations.  Agricultural  and  Forest   Meteorology,  in  press.   Drewry,   D.T.,   P.   Kumar,   S.   Long,   C.   Bernacchi,   X.-­‐Z.   Liang,   and   M.   Sivapalan,   2010.     Ecohydrological   responses  of  dense   canopies   to  environmental   variability:   1.   Interplay  between  vertical   structure   and   photosynthetic  pathway.  Journal  of  Geophysical  Research,  115,  G04023,  doi:10.1029/2010JG001341.   Duveiller,  G.,  F.  Baret,  and  P.  Defourny,  2012.  Crop  specific  green  area  index  retrieval  from  MODIS  data   at  regional  scale  by  controlling  pixel-­‐target  adequacy.  Remote  Sensing  of  Environment,  in  press.   Fang,  H.,  S.  Liang,  G.  Hoogenboom,  2011.   Integration  of  MODIS  products  and  a  crop  simulation  model   for  crop  yield  estimation.  International  Journal  of  Remote  Sensing,  32(4):  1039-­‐1065.   Hammer,  G.L.,   J.W.  Hansen,   J.G.  Phillips,   J.W.  Mjelde,  H.  Hill,  A.   Love,  A.  Potgieter,  2001.  Advances   in   application  of  climate  prediction  in  agriculture.  Agricultural  Systems  70:  515ʹ553.     Hank,  T.,  Bach,  H.,  Spannraft,  K.,   Friese,  M.,  Frank,  T.,  Mauser,  W.  2012.   Improving   the  process-­‐based   simulation   of   growth   heterogeneities   in   agricultural   stands   through   assimilation   of   earth   observation   data,  IEEE,  IGARSS  2012  Proceedings.     Hansen,   J.W.,  A.  Challinor,  A.   Ines,  T.  Wheeler,  and  V.  Moron,  2006.  Translating  climate   forecasts   into   agricultural  terms:  advances  and  challenges.  Climate  Research.  33:  27ʹ41.     Hansen,   J.W.,   Challinor,   A.,   Ines,   A.,   Wheeler,   T.,   Moron,   V.   2006.   Translating   climate   forecasts   into   agricultural  terms:  advances  and  challenges.  Climate  Research.  33:  27ʹ41.     Hansen,   J.W.   2005.   Integrating   seasonal   climate   prediction   and   agricultural   models   for   insights   into   agricultural  practice.  Philosophical  Transactions  of  the  Royal  Society.  B.  360:  2037-­‐2047     Hansen  J.W.,  Indeje,  M.  2004.  Linking  dynamic  seasonal  climate  forecasts  with  crop  simulation  for  maize   yield  prediction  in  semi-­‐arid  Kenya.  Agricultural  and  Forest  Meteorology.  125:  143ʹ157.     Immerzeel,   W.W.,   and   P.   Droogers,   2008.   Calibration   of   a   distributed   hydrological   model   based   on   satellite  evapotranspiration.  Journal  of  Hydrology.  349:  411-­‐424.   Ines,  A.,  Das,  N.N.,  Hansen,  J.W.  and  E.G.Njoku,  2012.  Assimilation  of  remotely  sensed  soil  moisture  and   vegetation  with  a  crop  simulation  model.  Remote  Sensing  of  Environment.  Under  review.   42     Ines,   A.,   and   K.   Honda,   2005.  On   quantifying   agricultural   and  water  management   practices   from   low   spatial   resolution   RS   data   using   genetic   algorithms:   A   numerical   study   for  mixed   pixel   environment.   Advances  in  Water  Resources.  28:  856-­‐870.   Mauser,  W.,  Bach  H.   (2009):  PROMET  ʹ  a  Physically  Based  Hydrological  Model   to  Study  the   Impact  of   Climate  Change  on  the  Water  Flows  of  Medium  Sized,  Mountain  Watersheds,  J.  Hydrol.,  376(2009)362-­‐ 377.   Pauwels,  V.  R.  N.,  and  G.  J.  M.  De  Lannoy  .2009.  Ensemble-­‐based  assimilation  of  discharge  into  rainfall-­‐ runoff     models:   A   comparison   of   approaches   to   mapping   observational   information   to   state   space.   Water  Resources  Research.  45,  W08428,  doi:10.1029/2008WR007590.   Rojas,  O.,  2007.  Operational  maize  yield  model  development  and  validation  based  on   remote  sensing   and  agroǦmeteorological  data  in  Kenya.  International  Journal  of  Remote  Sensing,  28:  3775-­‐3793.   Sehgal,  V.K.,  Jain,  S.,  Aggarwal,  P.K.,  and  S.  Jha,  2011.  Deriving  Phenology  Metrics  and  Their  Trends  Using   Times  Series  of  NOAA-­‐AVHRR  NDVI  Data.  Journal  of  Indian  Society  of  Remote  Sensing  39:  373-­‐381.     Vermeulen,  S.,  Zougmore,  R.,  Wollenberg,  E.,  Thornton,  P.,  Nelson,  G.,  Kristjanson,  P.,  Kinyangi,  J.,  Jarvis,   A.,  Hansen,   J.,  Challinor,  A.,  Campbell,  B.  and  P.  Aggarwal.  2012.  Climate  change,  agriculture  and  food   security:   a   global   partnership   to   link   research   and   action   for   low-­‐income   agricultural   producers   and   consumers.  Environmental  Sustainability.  4:128ʹ133     Vermeulen,  S.J.,  Aggarwal,  P.K.,  Ainslie,  A.,  Angelone,  C.,  Campbell,  B.M.,  Challinor,  A.J.,  Hansen,   J.W.,   Ingram,   J.S.I.,   Jarvis,   A.,   Kristjanson,   P.,   Lau,   C.   ,   Nelson,   G.C.,   Thornton,   P.K.,   E.   Wollenberg.   2012.   Options   for   support   to   agriculture   and   food   security   under   climate   change.   Environmental   Science  &   Policy  15:136-­‐144.     Varella,  H.,  M.  Guérif,  S.  Buis,  and  N.  Beaudoin,  2010.  Soil  properties  estimation  by  inversion  of  a  crop   model  and  observations  on  crops   improves  the  prediction  of  agro-­‐environmental.  European  Journal  of   Agronomy.  33:  139-­‐147.   Vazifedoust,  M.,  J.C.  van  Dam,  W.G.M.  Bastiaanssen  and  R.A.  Feddes,  2009.  Assimilation  of  satellite  data   into  agrohydrological  models  to  improve  crop  yield  forecasts.  International  Journal  of  Remote  Sensing,   30:  2523-­‐2545.   Verhoef,   W.,   Bach,   H.  2012.  Simulation   of   Sentinel-­‐3   images   by   four   stream   surface   atmosphere   radiative  transfer  modeling  in  the  optical  and  thermal  domains.    Remote  Sensing  of  Environment,  120  :   197-­‐207.   Zabel,  F.,  Mauser,  W.,  Marke,  T.,  Pfeiffer,  A.,  Zangl,  G.,  and  C.  Wastl.  2012.  Inter-­‐comparison  of  two  land-­‐ surface  models   applied   at   different   scales   and   their   feedbacks  while   coupled  with   a   regional   climate   model.  Hydrology  and  Earth  System  Science.  16:  1017ʹ1031.       43     5 Acronyms  and  Abbreviations   AGRHYMET   Centre  for  Agriculture,  Hydrology  and  Meteorology   AgMIP     Agricultural  Model  Intercomparison  and  Improvement  Project   AGROMETS HELL   FAO  Water  Balance  Model  implementation   AMSR   Advanced  Microwave  Scanning  Radiometer   AUV   Unmanned  Aerial  Vehicle   AVHRR     Advanced  Very  High  Resolution  Radiometer   C3   Carbon  fixation  method  in  photosynthesis  for  most  crops  in  temperate  regions  (e.g.,  wheat)   C4   Carbon  fixation  method  in  photosynthesis  for  some  crops  in  tropical  regions  (e.g.,  maize)   CCAFS   Climate  Change,  Agriculture  and  Food  Security  research  program  of  the  CGIAR   CCE   Crop  Cutting  Experiments     CGIAR    Research  Program  on  Climate  Change,  Agriculture  and  Food  Security     CSM   Cropping  System  Model   DSSAT   Decision  Support  System  for  Agro-­‐technology  Transfer   ECMWF     European  Centre  for  Medium-­‐Range  Weather  Forecasts   EnKF     Ensemble  Kalman  Filter   EOS   Earth  Observing  System,  a  coordinated  series  of  polar-­‐orbiting  and  low  inclination  satellites   ESSP   Earth  System  Science  Partnership   ETA   Actual  Crop  Evapotranspiration   EVI   Enhanced  vegetation  Index   FACE   Free  Air  Carbon  Enrichment     FAO     Food  and  Agriculture  Organisation  of  the  United  Nations   FAPAR   Fraction  of  Absorbed  Photosynthetically  Active  Radiation   FASAL   Forecasting  Agricultural  Output  Using  Space,  Agromet  and  Land  Observations  (India)   GAI   Green  Area  Index   GWSI     Global  Water  Satisfaction  Index   IARI   Indian  Agricultural  Research  Institute   ICRISAT     International  Crops  Research  Institute  for  Semi-­‐Arid  Tropics   INRA   French  National  Institute  for  Agricultural  Research   IRI   International  Research  Institute  for  Climate  and  Society   JRC   Joint  Research  Centre  of  the  European  Commission   LAI   Leaf  Area  Index     LINGRA   A  grass  growth  model  developed  by  ALTERRA,  Wageningen.  Based  on  LINTUL   LINTUL   Light  INTerception  and  UtiLization  simulator.  A  simple  general  crop  growth  model   MARS   dŚĞ͞Monitoring  Agriculture  with  Remote  ^ĞŶƐŝŶŐ͟  project  of  the  JRC  -­‐  AGRI4CAST   MERIS     MEdium  Resolution  Imaging  Spectrometer   MLCan   Vertically  resolved  canopy-­‐atmosphere  exchange  model   MM5   Mesoscale  crop  growth  model  of  Pennsylvania  State  University   MODFLOW   Groundwater  model   44     MODIS   MODerate-­‐resolution  Imaging  Spectroradiometer   MSG   METEOSAT  Second  Generation   N   Nitrogen   NADAMS   National  Agricultural  Drought  Assessment  &  Monitoring  System  (India)   NASA   National  Aeronautics  and  Space  Administration  (USA)   NDVI   Normalized  Difference  Vegetation  Index   NGO   Non-­‐governmental  organization   NIR   Near  Infrared   NOAA   National  Oceanic  and  Atmospheric  Administration   OLS   Ordinary  Least  Squares     ORYZA1   Eco-­‐physiological  model  for  irrigated  rice  production.   PROMET   Crop  Growth  Model  of  VISTA  (German  company)   PROSAIL   Radiative  transfer  model   RS   Remote  Sensing   SAR   Synthetic  Aperture  Radar   SOC   Stable  soil  organic  carbon     SPOT   Système  Pour  l'Observation  de  la  Terre  (French  satellites)   STICS   Generic  model  for  the  simulation  of  crops  and  their  water  and  nitrogen  balances.   SUCROS   Simple  and  Universal  CROp  growth  Simulator   SWAP   Soil  Water  Atmosphere  Plant  model   TM   Thematic  Mapper   TRMM   Tropical  Rainfall  Measuring  Mission   USGS   United  States  Geological  Survey   VGT   VEGETATION  sensor  on  board  the  SPOT  satellite   WARM   Rice  crop  model  used  at  JRC   WFP   World  Food  Programme   WOFOST   WOrld  FOod  Studies.  Simulation  model  for  the  quantitative  analysis  of  the  growth  and   production  of  annual  field  crops   WRSI   Water  Requirement  Satisfaction  Index   WTGROWS   Crop  simulation  model  for  regional  wheat  yield  mapping         45     6 Workshop  Program     tĞĚŶĞƐĚĂLJ͕ϭϯ:ƵŶĞϮϬϭϮ   KƉĞŶŝŶŐƐĞƐƐŝŽŶ   Ϭϵ͗ϬϬ-­‐Ϭϵ͗ϭϬ   tĞůĐŽŵĞĂĚĚƌĞƐƐ͕:Z   Ϭϵ͗ϭϬ-­‐Ϭϵ͗Ϯϱ   ƌŽƉ&ŽƌĞĐĂƐƚŝŶŐǁŝƚŚŝŶƚŚĞ&^WƌŽŐƌĂŵ͕:ĂŵĞƐ,ĂŶƐĞŶ͕&^͕/Z/-­‐ŽůƵŵďŝĂhŶŝǀĞƌƐŝƚLJ   Ϭϵ͗Ϯϱ-­‐Ϭϵ͗ϰϬ   dŚĞĐŚĂůůĞŶŐĞƐŽĨĂŶŽƉĞƌĂƚŝŽŶĂůĐƌŽƉLJŝĞůĚĨŽƌĞĐĂƐƚŝŶŐƐLJƐƚĞŵŝŶ^Ƶď-­‐^ĂŚĂƌĂŶĨƌŝĐĂ͘/ƐƚŚĞƌĞĂ ƌĞĂůŝƐƚŝĐĂŶĚĞĨĨĞĐƚŝǀĞƐŽůƵƚŝŽŶ͍&ƌĂŶĐŽŝƐ<ĂLJŝƚĂŬŝƌĞ͕:Z   Ϭϵ͗ϰϬ-­‐ϭϬ͗ϭϬ   /ŶƚĞŐƌĂƚŝŽŶŽĨĂŐƌŽ-­‐ŚLJĚƌŽůŽŐŝĐĂůŵŽĚĞůůŝŶŐ͕ƌĞŵŽƚĞƐĞŶƐŝŶŐĂŶĚŐĞŽŐƌĂƉŚŝĐĂůŝŶĨŽƌŵĂƚŝŽŶ͕:ŽƐ ǀĂŶĂŵ͕tĂŐĞŶŝŶŐĞŶhŶŝǀĞƌƐŝƚLJ   ^ĞƐƐŝŽŶϭ   ϭϬ͗ϰϱ-­‐ϭϭ͗ϭϱ   ƐƐŝŵŝůĂƚŝŽŶŽĨƌĞŵŽƚĞƐĞŶƐŝŶŐŽďƐĞƌǀĂƚŝŽŶƐŝŶƚŽĂĐƌŽƉŵŽĚĞůŝŵƉƌŽǀĞƐƉƌĞĚŝĐƚŝǀĞ ƉĞƌĨŽƌŵĂŶĐĞĨŽƌƐƉĂƚŝĂůĂƉƉůŝĐĂƚŝŽŶ͕DĂƌƚŝŶĞ'ƵĞƌŝĨ͕ /EZ   ϭϭ͗ϭϱ-­‐ϭϭ͗ϰϱ   ZĞŐŝŽŶĂůƌŽƉ^ŝŵƵůĂƚŝŽŶDŽĚĞůůŝŶŐĨŽƌzŝĞůĚWƌĞĚŝĐƚŝŽŶƵƐŝŶŐZĞŵŽƚĞ^ĞŶƐŝŶŐĂŶĚ'/^͗/ŶĚŝĂŶ džƉĞƌŝĞŶĐĞƐ͕sŝŶĂLJ^ĞŚŐĂů͕/Z/-­‐/ŶĚŝĂ   ϭϭ͗ϰϱ-­‐ϭϮ͗ϭϱ   hƐŝŶŐDK/^>/ƚŽĞƐƚŝŵĂƚĞŵĂŝnjĞLJŝĞůĚ͕'Ğƌƌŝƚ,ŽŽŐĞŶŵ͕tĂƐŚŝŶŐƚŽŶ^ƚĂƚĞhŶŝǀĞƌƐŝƚLJ   ϭϮ͗ϭϱ-­‐ϭϮ͗ϯϬŝƐĐƵƐƐŝŽŶ   ^ĞƐƐŝŽŶϮ   ϭϰ͗ϬϬ-­‐ϭϰ͗ϯϬ   džƉůŽƌŝŶŐƚŚĞĐůŝŵĂƚŝĐƌĞƐƉŽŶƐĞŽĨƚŚĞĐĞŶƚƌĂůh^ĂŐƌŽ-­‐ĞĐŽƐLJƐƚĞŵ͕ĂƌƌĞŶƌĞǁƌLJ͕E^-­‐:W>   ϭϱ͗ϯϬ-­‐ϭϲ͗ϬϬ   KŶƚŚĞĂƐƐŝŵŝůĂƚŝŽŶŽĨƌĞŵŽƚĞůLJƐĞŶƐĞĚƐŽŝůŵŽŝƐƚƵƌĞĂŶĚǀĞŐĞƚĂƚŝŽŶǁŝƚŚĐƌŽƉƐŝŵƵůĂƚŝŽŶ ŵŽĚĞůƐ͕ŵŽƌ/ŶĞƐ͕/Z/͕ŽůƵŵďŝĂhŶŝǀĞƌƐŝƚLJ   ^ĞƐƐŝŽŶϯ   ϭϲ͗ϯϬ-­‐ϭϳ͗ϬϬ   ^ŝŵƵůƚĂŶĞŽƵƐƐƚŝŵĂƚŝŽŶŽĨDŽĚĞů^ƚĂƚĞsĂƌŝĂďůĞƐĂŶĚKďƐĞƌǀĂƚŝŽŶĂŶĚ&ŽƌĞĐĂƐƚŝĂƐĞƐƵƐŝŶŐ ĂdǁŽ-­‐^ƚĂŐĞ,LJďƌŝĚ<ĂůŵĂŶ&ŝůƚĞƌ͕sĂůĞŶƚŝũŶWĂƵǁĞůƐ͕'ŚĞŶƚhŶŝǀĞƌƐŝƚLJ   ϭϳ͗ϬϬ-­‐ϭϳ͗ϯϬ   ^ĂƚĞůůŝƚĞŝŵĂŐĞƐŝŵƵůĂƚŝŽŶƐĨŽƌĚĂƚĂĂƐƐŝŵŝůĂƚŝŽŶĂƚŵƵůƚŝƉůĞƐĐĂůĞƐ͕,ĞŝŬĞĂĐŚ;s/^dͿĂŶĚ tŽůĨƌĂŵDĂƵƐĞƌ;hŶŝǀĞƌƐŝƚLJŽĨDƵŶŝĐŚͿ   ϭϳ͗ϯϬ-­‐ϭϴ͗ϬϬ   ƐƚŝŵĂƚŝŶŐĐƌŽƉďŝŽƉŚLJƐŝĐĂůƉƌŽƉĞƌƚŝĞƐĨƌŽŵƌĞŵŽƚĞƐĞŶƐŝŶŐĚĂƚĂďLJŝŶǀĞƌƚŝŶŐůŝŶŬĞĚƌĂĚŝĂƚŝǀĞ ƚƌĂŶƐĨĞƌĂŶĚĞĐŽƉŚLJƐŝŽůŽŐŝĐĂůŵŽĚĞůƐ͕<ĞůůLJZ͘dŚŽƌƉ͕h^   46     ϭϴ͗ϬϬ-­‐ϭϴ͗ϭϱ   ŝƐĐƵƐƐŝŽŶ   dŚƵƌƐĚĂLJ͕ϭϰ:ƵŶĞϮϬϭϮ   ^ĞƐƐŝŽŶϰ   ϵ͗ϬϬ-­‐Ϭϵ͗ϯϬ   DZ^ŽƉĞƌĂƚŝŽŶĂůĐƌŽƉŵŽŶŝƚŽƌŝŶŐĂŶĚLJŝĞůĚĨŽƌĞĐĂƐƚŝŶŐĂĐƚŝǀŝƚŝĞƐŝŶƵƌŽƉĞĂŶĚƉŽƐƐŝďůĞ ŝŵƉƌŽǀĞŵĞŶƚƐďĂƐĞĚŽŶƌĞŵŽƚĞƐĞŶƐŝŶŐĚĂƚĂ͕'ƌĞŐŽƌLJƵǀĞŝůůĞƌ͕ :Z   Ϭϵ͗ϯϬ-­‐ϭϬ͗ϬϬ   džƉĞƌŝĞŶĐĞƐǁŝƚŚƐĂƚĞůůŝƚĞĚĂƚĂĂƐƐŝŵŝůĂƚŝŽŶĨŽƌƌĞŐŝŽŶĂůĐƌŽƉLJŝĞůĚĨŽƌĞĐĂƐƚŝŶŐ͕ůůĂƌĚĚĞtŝƚ͕ ůƚĞƌƌĂ   ϭϬ͗ϬϬ-­‐ϭϬ͗ϯϬ   KƉĞƌĂƚŝŽŶĂůĐƌŽƉLJŝĞůĚĨŽƌĞĐĂƐƚƵƐŝŶŐƌĞŵŽƚĞƐĞŶƐŝŶŐĂŶĚĂŐƌŽŵĞƚĞŽƌŽůŽŐŝĐĂůŝŶtĞƐƚĨƌŝĐĂ͕ ĞƌŶĂƌĚdLJĐŚŽŶĂŶĚĂŬĂƌLJũĂďLJ͕hŶŝǀĞƌƐŝƚLJŽĨ>ŝĞŐĞ   ^ĞƐƐŝŽŶϱ   ϭϬ͗ϰϱ-­‐ϭϭ͗ϭϱ   ĂƚĂƐƐŝŵŝůĂƚŝŽŶďĂƐĞĚŽŶƚŚĞ/ŶƚĞŐƌĂƚŝŽŶŽĨ^ĂƚĞůůŝƚĞĂƚĂĂŶĚ&ŝĞůĚ^ĞŶƐŽƌĂƚĂĨŽƌƌŽƵŐŚƚ DŽŶŝƚŽƌŝŶŐ͕<ŝLJŽƐŚŝ,ŽŶĚĂ͕ŚƵďƵhŶŝǀĞƌƐŝƚLJ   ϭϭ͗ϭϱ-­‐ϭϭ͗ϰϱ   ĂƚĂĂƐƐŝŵŝůĂƚŝŽŶĨŽƌƚŚĞĐĂƌďŽŶĐLJĐůĞŝŶ^ƵĚĂŶŝĂŶƐŵĂůůŚŽůĚĞƌĐŽŵŵƵŶŝƚŝĞƐ͕^ŝďŝƌLJdƌĂŽƌĞ͕ /Z/^d   ϭϭ͗ϰϱ-­‐ϭϮ͗ϭϱ   ^Žŝů-­‐ǁĂƚĞƌ-­‐ĐƌŽƉŵŽĚĞůŝŶŐĨŽƌĚĞĐŝƐŝŽŶƐƵƉƉŽƌƚŝŶƐƵď-­‐ƐĂŚĂƌĂŶtĞƐƚĨƌŝĐĂ͗ĞdžƉĞƌŝĞŶĐĞƐĨƌŽŵ EŝŐĞƌĂŶĚĞŶŝŶ͕WŝĞƌƌĞ/ƌĠŶŝŬĂƚĐŚĠ