Alliance Bioversity-CIAT Research Online Accepted Manuscript Proximal sensing of Urochloa grasses increases selection accuracy The Alliance of Bioversity International and the International Center for Tropical Agriculture believes that open access contributes to its mission of reducing hunger and poverty, and improving human nutrition in the tropics through research aimed at increasing the eco-efficiency of agriculture. The Alliance is committed to creating and sharing knowledge and information openly and globally. We do this through collaborative research as well as through the open sharing of our data, tools, and publications. Citation: Jiménez, Juan de la Cruz; Leiva, L.; Cardoso, J.A.; French, A.N.; Thorp, K.R. (2020) Proximal sensing of Urochloa grasses increases selection accuracy. Crop and Pasture Science 71(4) p. 401-409. ISSN: 1836- 0947 Publisher’s DOI: https://doi.org/10.1071/CP19324 Access through CIAT Research Online: https://hdl.handle.net/10568/111686 Terms: © 2020. The Alliance has provided you with this accepted manuscript in line with Alliance’s open access policy and in accordance with the Publisher’s policy on self-archiving. This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License. You may re-use or share this manuscript as long as you acknowledge the authors by citing the version of the record listed above. You may not use this manuscript for commercial purposes. For more information, please contact Alliance Bioversity-CIAT - Library Alliancebioversityciat-Library@cgiar.org Page 1 of 67 Crop & Pasture Science Summary text for the online Table of Contents Extensive areas in the Tropics are dedicated to livestock production. Grazing activities in these areas, however, are highly restricted by forage availability. By using sensors in place of conventional methods of forage evaluation, higher number of forages can be reliably evaluated, while incurring minimal additional cost. The use of digital cameras and hyperspectral sensors to evaluate forage characteristics and production were found effective and potentially useful for selecting outstanding hybrids. http://www.publish.csiro.au/nid/40.htm ly On evi ew or R F Crop & Pasture Science Page 2 of 67 1 Proximal sensing of Urochloa grasses increases selection accuracy 2 3 Juan de la Cruz Jiménez 1*, Luisa Leiva2, Juan A. Cardoso3, Andrew N. French4 and Kelly R. Thorp4 4 5 1 UWA School of Agriculture and Environment, Faculty of Science, The University of Western Australia, 6 35 Stirling Highway, Crawley, WA 6009, Australia. 7 2 Department of plant breeding, Swedish University of Agricultural Sciences, Alnarp, Sweden. 8 3 Tropical Forage Program, International Center for Tropical Agriculture (CIAT), Km 17 Recta Cali – 9 Palmira, Colombia. 10 4 USDA-ARS, U.S. Arid Land Agricultural Research Center, 21881 N Cardon Ln, Maricopa, AZ 85138, 11 United States. 12 13 14 15 * Corresponding author: Juan de la Cruz Jiménez: juan.jimenezserna@research.uwa.edu.au 16 17 18 19 20 1 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Page 3 of 67 Crop & Pasture Science 21 Abstract 22 In the American Tropics, livestock production is highly restricted by forage availability. In 23 addition, the breeding and development of new forage varieties with outstanding yield and high 24 nutritional quality is often limited by a lack of resources and poor technology. Non-destructive 25 high throughput phenotyping offers a rapid and economical means to evaluate large numbers of 26 genotypes. In this study, visual assessments, digital color images, and spectral reflectance data 27 were collected from 200 Urochloa hybrids in a field setting. Partial least squares regression 28 (PLSR) was applied to relate visual assessments, digital image analysis and spectral data with 29 shoot dry weight (DW), crude protein (CP) and chlorophyll content. Visual evaluations of biomass 30 and greenness, digital color imaging, and hyperspectral canopy data were collected in 68, 40 and 31 80 minutes, respectively. Root mean squared errors of prediction for PLSR estimations of DW, 32 CP, and chlorophyll were lower for digital image analysis followed by hyperspectral analysis and 33 visual assessments. This study showed that digital color image and spectral analysis techniques 34 have the potential to improve precision and reduce time for tropical forage grass phenotyping. 35 Keywords: High throughput phenotyping, Urochloa, tropical forage grasses, plant breeding. 36 37 Introduction 38 Livestock productivity depends on forage availability and quality. Grasses from the Urochloa (syn. 39 Brachiaria) genus have been widely planted in the tropics as forage for grazing ruminant livestock 40 and are considered the most important forages in the American Tropics (Miles et al. 2004). The 41 International Center for Tropical Agriculture (CIAT) in Colombia conducts a Urochloa breeding 42 program aimed at developing hybrids with outstanding performance on infertile, acidic soils with 2 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Crop & Pasture Science Page 4 of 67 43 superior forage productivity and nutritional quality. The hybrid development process is difficult 44 and time consuming. In a regular, three-year breeding cycle, over 7000 hybrids are produced by 45 open pollination, but fewer than 2% of these are retained for full evaluation. Approximately half 46 of the population is discarded based on their reproductive mode (sexual genotypes are discarded 47 and apomictic hybrids are kept); another major proportion is discarded based on visual evaluations; 48 and only a limited number of hybrids (approximately 100) are finally evaluated for different biotic 49 and abiotic stresses (Valheria Castiblanco, personal communication). The evaluation of genotypes 50 is restricted mainly by insufficient economic resources and technology for rapid screening. 51 Forage grasses exhibiting great biomass production and high nutritional quality are key 52 determinants of the productivity of grazing animals (Herrero et al., 2013). Therefore, evaluations 53 of shoot biomass production and quality parameters (i.e. crude protein) are among the most 54 important traits for improvement in any forag grasses breeding program. However, owing to the 55 destructive nature of these measurements and the insufficient economic resources, the evaluation 56 of these parameters is postponed to final stages of the breeding program characterized by a reduced 57 number of genotypes. Instead of analytical measurements of forage quality and destructive 58 biomass harvests, periodic visual evaluations of plant performance (i.e., plant biomass and 59 greenness) over time is traditionally used in Urochloa breeding programs to select superior plants 60 at initial stages of the breeding scheme (Miles et al. 2004; Miles 2007). These visual evaluations 61 are laborious and may not be sufficiently accurate especially in breeding populations characterized 62 by high genetic diversity and substantial genotype x environment interaction (Walter et al. 2012). 63 The use of new technologies for in-field non-destructive, high throughput phenotyping (HTP), 64 including digital image analysis and proximal hyperspectral sensing, offers the possibility to 65 precisely evaluate a larger number of genotypes than feasible in traditional ways, achieved at low 3 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 5 of 67 Crop & Pasture Science 66 cost, and implemented in a short period of time (Montes et al. 2007; White et al. 2012; Andrade- 67 Sanchez et al. 2014). Proximal hyperspectral sensing provides continuous information along the 68 visual and near-infrared electromagnetic spectrum. This information often relates to plant traits 69 and has successfully been studied in grasses to estimate quality parameters (Skidmore et al. 2010; 70 Pullanagari et al. 2012; Thulin et al. 2012; Ferner et al. 2015; Safari et al. 2016), diversity (Lopatin 71 et al. 2017) and nutrient content (Fava et al. 2009; Knox et al. 2012; Ramoelo et al. 2013; Adjorlolo 72 et al. 2015; Foster et al. 2017). Likewise, plant image analysis for phenotyping purposes is based 73 on image segmentation to separate the soil background and the plant for further quantification of 74 regions of interest (Tucker 1979; Woebbecke et al. 1995; Camargo 2004; Hunt et al. 2005). Digital 75 image analysis has also been used for quantifying vegetation indices related to plant growth, 76 greenness and nutritional status (Meyer and Camargo 2008; Hunt et al. 2013). Very few reports of 77 hyperspectral (Numata et al. 2008) or image analysis of Urochloa grasses exist in literature 78 (Jimenez et al. 2017). 79 No studies combining hyperspectral information and image analyses and comparing them to 80 conventional phenotyping methods is available. Moreover, hyperspectral data have not been used 81 to evaluate target traits in Urochloa breeding programs. In this study, in-field visual evaluations, 82 proximal hyperspectral data, and digital imaging were collected over canopies of Urochloa 83 hybrids. Partial least squares regression was used to relate hyperspectral information to field 84 measurements and machine learning (i.e. naive Bayes multiclass) was used to extract vegetation 85 indices from overhead canopy images. The objectives of this study were to: 1) develop PLSR 86 models for predicting CP, forage DW, and chlorophyll content; 2) extract plant traits from digital 87 image analysis to relate with CP, forage DW, and chlorophyll; and 3) demonstrate the superiority 88 of HTP techniques as compared to conventional visual evaluation of traits. Crude protein, forage 4 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Crop & Pasture Science Page 6 of 67 89 DW, and chlorophyll content were chosen as target traits in this study as they are key parameters 90 determining both plant and cattle productivity. The development of HTP methodologies to 91 evaluate tropical forages will increase the number of hybrids evaluated per selection cycle, thus 92 permitting more intense selection and hence, genetic gain. The identification of new hybrids with 93 outstanding performance (i.e. higher biomass, greener and high CP) will result in more productive 94 pastures with concomitant increases in milk and meat production in livestock systems in tropical 95 savannahs. 96 Materials and methods 97 Field experiment 98 Field data were obtained in August 2016 at the International Center for Tropical Agriculture 99 (CIAT) in Cali, Colombia (Lat. 3° 29’ N; Long. 76° 21’ W; altitude 965 m). Four thousand 100 Urochloa hybrids generated from crosses between the CIAT’s Urochloa breeding program 101 population SX12 and U. decumbens cv. Basilisk (CIAT 606) were initially planted in an andisol 102 soil in an augmented block design and spaced at 1.5x1.5 m. These plants were visually evaluated 103 four times (data not shown) for persistence, vigor and greenness after sequential cuttings every 104 three months for one year. After that period, 200 hybrids were randomly selected for further visual 105 and HTP analysis. These 200 hybrids, instead of the entire population, were selected for economic 106 and practical reasons. Visual evaluations of biomass and greenness, imaging and spectra collection 107 were performed after 3 months re-growth after cutting (see information below). Plant heights 108 ranged from 20 to 50 cm and shoot architecture varied with both decumbent and erectus growth. 109 Visual evaluation 110 Plant biomass was assessed using a nine-point visual scale, where level ‘9’ indicated high shoot 111 biomass with many tillers and leaves while level ‘1’ indicated stunted growth with fewer tillers 5 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 7 of 67 Crop & Pasture Science 112 and leaves. Plant greenness was visually evaluated using a five-point visual scale, where level ‘5’ 113 represented intense dark green in all the leaves of the plant and level ‘1’ indicated yellow-pale 114 color in all leaves of the plant. This visual evaluation was conducted in 68 minutes one week before 115 the HTP measurements (Table 1). 116 Imaging collection and analysis 117 Individual, digital color images for each of the 200 hybrids were taken at 1.2 m above the soil 118 surface using a commercial digital 13-Megapixel camera (Coolpix P6000, Nikon, Japan) fixed to 119 a buggy tractor. Digital images were saved in 4224 x 3168 pixel JPG format. The canopy cover 120 (CC) and six vegetation indices including the normalized green red difference index (NGRDI), 121 excess green index (ExG), excess red index (ExR), excess green minus excess red (ExGR), green 122 ratio (GR) and green leaf index (GLI) were created using the formulae as indicated in Table 2. The 123 canopy cover was extracted by dividing the total number of pixels representing the plant by the 124 total number of pixels in each image. The vegetation indices were extracted using naive Bayes 125 multiclass. Briefly, the distribution of colors in a set of digital color images (training set) was used 126 to estimate the probability density function for each of the different region of interest (i.e. plant 127 and background). Once the regions of interest were defined in the training set, the machine learning 128 process was applied to all images to accurately classify and separate regions of interest. Therefore, 129 every pixel in an image was classified into the previously defined plant and background classes. 130 Every pixel characterizing the plant (but not the background) was then decomposed into red (R), 131 green (G), and blue (B) channels. These channels were then normalized as follows: 132 𝑅 𝐺 𝐵 133 𝑟 = 𝑅 + 𝐺 + 𝐵;𝑔 = 𝑅 + 𝐺 + 𝐵;𝑏 = 𝑅 + 𝐺 + 𝐵 134 6 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Crop & Pasture Science Page 8 of 67 135 Normalization makes the variations of light intensities uniform across the spectral distribution, 136 thus, the individual color components (i.e. r,g,b) are independent from the overall brightness of 137 the image (Cheng et al. 2011). Normalized channels were further used for the quantification of the 138 vegetation indices (Table 2). Image analysis code was written in Java and run in ImageJ software 139 (National Institutes of Health, Bethesda, Maryland, USA). Images were collected early in the 140 morning to avoid beam solar radiation interferences. Digital images contained the whole plant in 141 addition to the 23-cm diameter field-of-view (as indicated below for hyperspectral measurements, 142 Supplementary Fig 1). The collection process took 40 minutes (Table 1). 143 144 Spectral collection and analysis 145 Hyperspectral field data collections were performed on clear days at full sun exposure around 11 146 am by positioning a hand-held field spectroradiometer (Fieldspec 2, Malvern Panalytical, Malvern, 147 UK) directly above the plant canopy. The instrument was used with no foreoptics, which provided 148 a 25-degree full conical angle field-of-view. To avoid soil background noise, the bare optical input 149 was positioned at 50 cm from the top of the plant canopy to yield a 23-cm diameter field of view. 150 The instrument collected information in 750 narrow wavebands from 325 to 1075 nm in 1 nm 151 intervals. One or ten spectral scans were collected per plant and 50 plants were evaluated daily in 152 about 20 minutes. Differences in the collection protocols were tested to evaluate the most effective 153 way. Different spectra collection processes (1 or 10 scans) did not yield significant differences in 154 the root mean squared error of prediction for the different traits evaluated (Supplementary Table 155 1). Radiometric collections over a 99% Spectralon panel (Labsphere, Inc., North Sutton, New 156 Hampshire) were used to describe incoming solar irradiance throughout the data collection 157 process. The radiometric collections over the calibration panel were made before starting and after 158 every five canopy scans or when slight changes in solar irradiance due to cloud cover occurred. 7 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 9 of 67 Crop & Pasture Science 159 The values of the Spectralon panel radiance were used to compute the canopy reflectance of the 160 plants in each wavelength over the time of spectra collection. Subsequently, 401 bands from 500 161 to 900 nm were used for analysis. Based on visual inspection of reflectance spectra, these bands 162 were typically less noisy, as compared to bands at the bounds of detector sensitivity. Spectral 163 collection process was run in 80 minutes (Table 1). 164 Laboratory sample collections 165 Plants were immediately harvested after spectra collection. Aboveground tissue was removed by 166 cutting the area defined by a 23-cm diameter plastic circle co-located with the spectral data 167 collection area. Tissues were packed in plastic bags and stored on ice in a cooler in the field and 168 then transported to the laboratory. The extraction of chlorophyll was performed by adding 100 mg 169 of fresh tissue to 80% (v/v) cold methanol, and the mix was homogenized using a pestle in a mortar 170 until the plant residue was clear and the solution was uniform. This solution was then filtered and 171 absorbance was determined with a spectrophotometer (Synergy HT, Biotek, Winooski, USA). 172 Total chlorophyll concentration was calculated according to Lichtenthaler and Welburn (1983). 173 Dry weight (DW) was measured on an electronic balance (PB602S, Mettler Toledo, LLC, 174 Columbus, OH, USA) after oven-drying the samples for three days at 60 °C. Nitrogen 175 concentrations in the dry tissue were determined by using an automated nitrogen-carbon analyser 176 (Sercon, Crewe, UK). Urochloa and common bean (Phaseolus vulgaris) leaves were used as 177 reference tissues for confirmation of the reliability of the analyses. The crude protein content was 178 calculated by multiplying nitrogen content with 6.25, as protein is assumed to contain 16% 179 nitrogen on average. 180 181 Statistical analysis 8 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Crop & Pasture Science Page 10 of 67 182 Visual evaluations, digital image analysis, spectral reflectance, and plant trait data were 183 incorporated into a partial least squares regression (PLSR) algorithm (Mevik and Wehrens 2007) 184 within the R Project for Statistical Computing (http://www.r-project.org). Models were developed 185 to predict each plant trait (i.e. CP, DW and chlorophyll) and to compare the precision for prediction 186 of each of the different methods of phenotyping. Partial least squares regression was used in 187 preference to conventional least squares analysis to reduce co-linearity effects. Thorp et al. (2011) 188 provided the details on the PLSR methodology used in the present study. Briefly, if Y is an n×1 189 vector of responses (i.e. CP, DW or chlorophyll content) and X is an n-observation by p-variable 190 matrix of predictors (a set of visual evaluations, digital image analysis, or spectral reflectance 191 data), PLSR aims to decompose X into a set of A orthogonal scores such that the covariance with 192 corresponding Y scores is maximized. The X-weight and Y-loading vectors that result from the 193 decomposition are used to estimate the vector of regression coefficients, βPLS, such that 194 Y = X βPLS + ε 195 where ε is an n×1 vector of error terms. 196 Leave-one-out cross validation was used to test model predictions for independent data. Results 197 were reported for PLSR models with the number of factors that minimized the root mean squared 198 error of cross validation. Pearson’s correlation coefficients were calculated for the different traits 199 extracted from digital color images taken from Urochloa hybrids. 200 Results 201 In this study, visual evaluations of biomass and greenness, digital color imaging and hyperspectral 202 data were collected on 200 Urochloa hybrids in 68, 40 or 80 minutes, respectively (Table 1). High 203 variability for the different characteristics of DW, CP and chlorophyll content evaluated on 200 204 Urochloa hybrids was found (Table 3). 9 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 11 of 67 Crop & Pasture Science 205 Visual assessments 206 Partial least squares regressions for measured traits of DW, CP and chlorophyll based on visual 207 evaluations of biomass and greenness performed with a root mean square error of prediction 208 (RMSEP) of 8.47 g plant-1, 1.76% and 0.60 mg g FW respectively (Fig 1). 209 Spectral data and digital image phenotyping 210 The PLSR models developed from the digital image analysis estimated DW, CP and chlorophyll 211 with a RMSEP of 7.81 g plant-1, 1.53% and 0.57 mg g FW, respectively (Fig 2). Differences on 212 the correlation coefficients among traits extracted from image analysis indicated that including 213 different indices into the model added independent information to build stronger PLSR models 214 (Supplementary Fig 2). The contribution of each trait extracted from digital image analysis to the 215 overall prediction of each destructively-measured trait is shown in Table 4. The GLI had the 216 stronger positive influence on the PLSR model for predicting DW. The ExGR had the stronger 217 positive influence on the PLSR model for predicting both CP and chlorophyll content. 218 The fitted PLSR models developed from 401 wavebands of canopy spectral reflectance estimated 219 DW, CP and chlorophyll with a RMSEP of 7.90 g plant-1, 1.63% and 0.55 mg g FW, respectively 220 (Fig 3). The contribution of each spectral waveband to the overall prediction of each destructively- 221 measured trait is shown in the Fig 4. In the PLSR model for DW, local extrema in regression 222 coefficients were found at 701 and 674 nm, corresponding to red light near the inflection band and 223 red light, respectively (Fig 4a). Strong positive contribution to DW estimation were with NIR (700- 224 750), and a strong negative contribution with red light (674-640). In the PLSR models for CP and 225 chlorophyll, regression coefficient plots exhibited strong positive contribution for traits estimation 226 in the visible green light (Fig 4b and c). The PLSR models for CP contrasted wavebands in the 227 visible spectrum with positive contribution from wavebands around 503 nm and negative 10 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Crop & Pasture Science Page 12 of 67 228 contributions from wavebands at 678 nm. Similarly, regression coefficients for total chlorophyll 229 indicated strong positive contribution in the visible spectrum around 504 nm and negative 230 contribution throughout the visible wavebands, especially at 625 and 643 nm (Fig 4c). This is 231 sensible considering visible light absorption is increased with additional leaf chlorophyll. 232 Discussion 233 The results from this study demonstrate that the current visual assessment methodology at initial 234 steps of the breeding cycle in the CIAT Urochloa breeding program can be improved using non- 235 destructive HTP techniques. Color imaging, hyperspectral analysis, and PLSR models are more 236 precise and faster than visual evaluations, thus increasing the number of plants evaluated in the 237 tropical forage breeding program. 238 Visual evaluations of plant growth and greenness (characteristics associated with N content, and 239 therefore CP and chlorophyll concentration in leaves) have traditionally been used to discard 240 Urochloa hybrids at initial stages of plant phenotyping. The visual evaluation of an entire breeding 241 population (i.e., 7,000 hybrids) is a slow, costly and tedious process, and is often biased by 242 subjectivity and human fatigue, especially when phenotypic variation of such traits is high (Table 243 3). In this study, the estimation of DW, CP, and chlorophyll content was more precisely and 244 consistently estimated by HTP techniques. Dry weight and CP predictions were more accurate 245 using digital image analysis, followed by spectral analysis and visual evaluations. Chlorophyll 246 content was better estimated by the analysis of 401 spectral wavebands, followed by color image 247 analysis and finally visual evaluations (Fig 1, 2 and 3). The time required to run non-destructive 248 HTP evaluations was considerably shorter by 28 minutes per 200 plants for color image analysis 249 than visual evaluations, but longer by 12 minutes per 200 plants in hyperspectral than in visual 250 evaluations (Table 1). 11 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 13 of 67 Crop & Pasture Science 251 The moderate trends in the relationship between Urochloa canopy imaging and reflectance and 252 measured DW, CP and chlorophyll may indicate that the method is not appropriate for very precise 253 estimations of these traits. However, for breeding purposes where a large percentage of hybrids 254 are discarded without detailed evaluation due to scarce resources, a difference in DW of 7.90 g 255 plant-1 or a difference of 1.63% in the CP content of plants may be acceptable during initial stages 256 of plant breeding. This moderate trend between Urochloa canopy analysis and measured traits in 257 this study can be explained by dissimilarities in the canopy architecture of the Urochloa genotypes 258 (Numata et al. 2008), as well as different growth patterns during recovery from cutting. The further 259 evaluation of breeding populations with contrasting canopy architecture will improve the accuracy 260 of the PLSR model to predict the targeted traits. Nonetheless, by combining both digital image and 261 hyperspectral analysis techniques, higher precision accuracy for DW, CP and Chlorophyll content 262 can be achieved. 263 The vegetation indices (see Table 2) extracted from color images of 200 Urochloa hybrids were 264 originally developed to separate green plants from the background by extracting green and red 265 colors from digital images. These indices have been related to different plant characteristics 266 including biomass, chlorophyll content and nutritional status (Tucker 1979; Woebbecke et al. 267 1995; Camargo 2004; Hunt et al. 2005; Meyer and Camargo 2008; Hunt et al. 2013; Lee and Lee 268 2013; Wang et al. 2013). In this study, digital image analysis performed better than hyperspectral 269 scanning analysis to estimate DW and CP (Fig 2 and 3). Nonetheless, the use of spectral analysis 270 over grasses becomes more important when this technique is used to detect either nutritional or 271 anti-nutritional compounds (i.e. metabolisable energy, digestibility, fiber) that are better estimated 272 with the near-infrared regions of the electromagnetic spectra (Curran 1989; Pullanagari et al. 2012; 273 Ferner et al. 2015). In this sense, the use of digital color image analysis and hyperspectral analysis 12 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Crop & Pasture Science Page 14 of 67 274 is complementary because by using both techniques a diverse set of plant traits can accurately be 275 predicted and by adding extra factors to the prediction model, higher prediction accuracy can be 276 achieved (cf. Numata et al. 2008). Future efforts will use data mining to fine-tune the spectral 277 bands included in the PLSR model (Thorp et al. 2017), which can reduce model error and improve 278 model fit statistics. Although testing multiple methods of analysis was not the intention of this 279 study, future research could also test other techniques (e.g., artificial neural networks) for relating 280 HTP measurements to plant traits. 281 The regression coefficients for the PLSR for DW and chlorophyll content obtained in this study 282 highlight that the key wavelengths for the prediction of these traits occur in the green, red, red- 283 edge and NIR regions of the electromagnetic spectrum (Fig 4). Previous hyperspectral studies have 284 highlighted those regions as being highly representative for dry mass and chlorophyll content in 285 plants (Lichtenthaler et al. 1996; Thenkabail et al. 2000; Mutanga and Skidmore 2004; Fava et al. 286 2009; Thorp et al. 2011; Adjorlolo et al. 2015; Dou et al, 2018). Although some similarities were 287 found between wavebands among the different traits, the general regression coefficients differed 288 among the traits, thus demonstrating that the reflectance data in a given waveband contributed 289 differently toward the estimation of a given trait. Given the logistical burden to collect and analyze 290 hyperspectral scans, the identification of informative key bands associated with each evaluated 291 trait can improve the HTP process (Thorp et al. 2017). Results from this study will help guide 292 selection of optimal bands in the construction of multispectral sensors tailored to predict specific 293 traits of interest in tropical forage breeding programs. 294 The PLSR models for predicting DW, CP and chlorophyll content can be now used to evaluate the 295 next generation of hybrids from the same Urochloa gene pool (i.e. U. ruziziensis – U. brizantha – 296 U. decumbens). The accuracy of this prediction models relies on collection protocols similar to the 13 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 15 of 67 Crop & Pasture Science 297 explained in the Materials and Methods section and evaluations on plants with comparable growth 298 characteristics as the hybrids evaluated here (i.e. about three months after regrowth). The 299 prediction accuracy will likely be reduced on larger plants with higher biomass (Hill 2004) and a 300 greater proportion of senescent leaves (Asner 1998). The development of more precise PLSR 301 models to predict variables of interest in a breeding program requires an ongoing effort. The 302 collection of ground data every year while making improvements to standardize collection 303 protocols and incorporate wider range of genotypes will result in more accurate and robust models. 304 Larger data sets will increase estimation precision. 305 306 Conclusions 307 In this study, 200 Urochloa hybrids were monitored in 40 and 80 minutes by digital imaging and 308 spectral analysis, respectively (Table 1). At this pace, more than 1000 Urochloa hybrids could be 309 evaluated in a period of less than 7 hours. This means that forage biomass and quality in a high 310 number of genotypes would be reliably evaluated with minimal increased acquisition costs relative 311 to destructive harvest. This demonstrates the superiority of HTP techniques as compared to 312 conventional visual evaluation of traits. The PLSR models for predicting CP, forage DW, and 313 chlorophyll content developed in this study supports the evaluation of higher numbers of genotypes 314 at initial stages of the breeding program. The greater numbers of plants evaluated reliably every 315 year in the Urochloa breeding program, the greater the genetic gain will be. Therefore, the use of 316 image analysis and hyperspectral monitoring over Urochloa hybrids canopies will benefit the on- 317 going breeding program. The application of this HTP method could be of great help in rural remote 318 areas lacking facilities to perform destructive harvest and plant chemical analysis. Research is 319 underway to improve the utility of proximal sensing by considering a greater range of canopy 14 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Crop & Pasture Science Page 16 of 67 320 architectural configurations and evaluating the potential to assess nutritional quality, including 321 characteristics such as metabolisable energy, fiber, digestibility, lignin and cellulose fractions in 322 Urochloa grasses. 323 324 Conflict of interest 325 The authors have no conflicts of interest to declare. 326 Acknowledgements 327 We thank Dr. John W. Miles for helpful suggestions and comments to early version of this 328 manuscript. JCJ is grateful to the USDA Foreign Agricultural Service, Norman Borlaug 329 Fellowship Program for a training fellowship. JCJ thanks the CIAT’s Young Scientist Award 2016 330 Program for travel assistance. This work was partially undertaken as part of the CGIAR Research 331 Program on Livestock. We thank all donors that globally support our work through their 332 contributions to the CGIAR system. 333 References 334 Adjorlolo C, Mutanga O, Cho MA (2015) Predicting C3 and C4 grass nutrient variability using in 335 situ canopy reflectance and partial least squares regression. International Journal of Remote 336 Sensing 36(6), 1743–1761. 337 338 Andrade-Sanchez P, Gore MA, Heun JT, Thorp KR, Carmo-Silva AE, French AN, Salvucci ME, 339 White JW (2014) Development and evaluation of a field-based high throughput phenotyping 340 platform. Functional Plant Biology 41, 68–79. 341 15 http://www.publish.csiro.au/nid/40.htm ly On vie w e or R F Page 17 of 67 Crop & Pasture Science 342 Camargo JN (2004) A combined statistical—soft computing approach for classification and 343 mapping weed species in minimum tillage systems. University of Nebraska, Lincoln, NE 344 345 Currant PJ (1989) Remote sensing of foliar chemistry. Remote Sensing of the Environment 30, 346 271–278. 347 348 Dou Z, Cui L, Li J, Zhu Y, Gao C, Pan X, Lei Y, Zhang M, Zhao X, Li W (2018) Hyperspectral 349 estimation of the chlorophyll content in short-term and long-term restorations of mangrove in 350 Quanzhou Bay Estuary, China. Sustainability 10, 1127; doi:10.3390/su10041127 351 352 Fava F, Colombo R, Bocchi S, Meroni M, Sitzia M, Fois N, Zucca C (2009) Identification of 353 hyperspectral vegetation indices for Mediterranean pasture characterization. International Journal 354 of Applied Earth Observation and Geoinformation 11, 233–243. 355 356 Ferner J, Linstadter A, Sudekum KH, Schmidtlein S (2015) Spectral indicators of forage quality 357 in West Africa’s tropical savannas. International Journal of Applied Earth Observation and 358 Geoinformation 41, 99–106. 359 360 Foster AJ, Kakani VG, Mosali J (2017) Estimation of bioenergy crop yield and N status by 361 hyperspectral canopy reflectance and partial least square regression. Precision Agriculture 18, 362 192–209. 363 364 Herrero M, Havlík P, Valin H, Notenbaert A, Rufino MC, Thornton PK, Blümmel M, Weiss F, 365 Grace D, Obersteiner M (2013) Biomass use, production, feed efficiencies, and greenhouse gas 16 http://www.publish.csiro.au/nid/40.htm nly iew O v or Re F Crop & Pasture Science Page 18 of 67 366 emissions from global livestock systems. Proceedings of the National Academy of Science of the 367 United States of America 110(52), 20888–20893. 368 369 Hunt Jr ER, Cavigelli M, Daughtry CST, McMurtrey J, Walthall CL (2005) Evaluation of digital 370 photography from model aircraft for remote sensing of crop biomass and nitrogen status. Precision 371 Agriculture 6, 359–378. 372 373 Hunt Jr ER, Doraiswamy PC, McMurtrey J, Daughtry CST, Perry EM, Akhmedov B (2013) A 374 visible band index for remote sensing leaf chlorophyll content at the canopy scale. International 375 Journal of Applied Earth Observation and Geoinformation 21, 103–112. 376 377 Jiménez JC, Cardoso JA, Leiva LF, Gil J, Forero MG, Worthington ML, Miles JW, Rao IM (2017) 378 Non-destructive phenotyping to identify brachiaria hybrids tolerant to waterlogging stress under 379 field conditions. Frontiers in Plant Science 8, 167. doi: 10.3389/fpls.2017.00167 380 381 Knox NM, Skidmore AK, Prins HHT, Heitkönig IMA, Slotow R, van der Waal C, de Boer WF 382 (2012) Remote sensing of forage nutrients: Combining ecological and spectral absorptiond feature 383 data. ISPRS Journal of Photogrammetry and Remote Sensing 72, 27–35. 384 385 Meyer GE, Hindman TW, Lakshmi K (1998) Machine vision detection parameters for plant 386 species identification. In ‘Precision Agriculture and Biological Quality’. (Eds GE Meyer, JA De 387 Shazer) pp. 327–335. (Proceedings of SPIE. vol. 3543, Bellingham, WA) 388 17 http://www.publish.csiro.au/nid/40.htm ly w O n vie r R e Fo Page 19 of 67 Crop & Pasture Science 389 Meyer GE, Camargo J (2008) Verification of color vegetation indices for automated crop imaging 390 applications. Computers and Electronics in Agriculture 63, 282–293. 391 392 Mevik BH, Wehrens R (2007) The pls Package: Principal component and partial least squares 393 regression in R. Journal of Statistical Software 18(2), 1-23. 394 395 Miles JW, Do Valle CB, Rao IM, Euclides VPB (2004) Brachiariagrasses. American Society of 396 Agronomy, Crop Science Society of America, Soil Science Society of America, 677 S. Segoe Rd., 397 Madison, WI 53711, USA. Warm Season (C4) Grasses, Agronomy Monograph no. 45. 398 399 Miles JW (2007) Apomixis for cultivar development in tropical forage grasses. Crop Science 400 47(S3), S238–S249. 401 402 Montes JM, Melchinger AE, Reif JC (2007) Novel throughput phenotyping platforms in plant 403 genetic studies. Trends in Plant Science 12, 433–436. 404 405 Mutanga O, Skidmore AK (2004) Narrow band vegetation indices overcome the saturation 406 problem in biomass estimation. International Journal of Remote Sensing 25(19), 3999–4014. 407 408 Numata I, Roberts DA, Chadwick OA, Schimel JP, Galvão LS, Soares JV (2008) Evaluation of 409 hyperspectral data for pasture estimate in the Brazilian Amazon using field and imaging 410 spectrometers. Remote Sensing of Environment 112, 1569–1583. 411 18 http://www.publish.csiro.au/nid/40.htm ly On vie w e or R F Crop & Pasture Science Page 20 of 67 412 Lee KJ, Lee BW (2013) Estimation of rice growth and nitrogen nutrition status using color digital 413 camera image analysis. European Journal of Agronomy 48, 57–65. 414 415 Lichtenthaler HK, Wellburn AA (1983) Determination of total carotenoids and chlorophylls a and 416 b of leaf extracts in different solvents. Biochemical society transactions 603, 591–592. 417 418 Lichtenthaler HK, Gitelson A, Lang M (1996) Non-destructive determination of chlorophyll 419 content of leaves of a green and an aurea mutant of tobacco by reflectance measurements. Journal 420 of Plant Physiology 148, 483–493. 421 422 Lopatin J, Fassnacht FE, Kattenborn T, Schmidtlein S (2017) Mapping plant species in mixed 423 grassland communities using close range imaging spectroscopy. Remote Sensing of Environment 424 201, 12–23. 425 426 Louhaichi M, Borman MM, Johnson DE (2001) Spatially located platform and aerial photography 427 for documentation of grazing impacts on wheat. Geocarto International 16(1), 65-70. 428 429 Pullanagari RR, Yule IJ, Tuohy MP, Hedley MJ, Dynes RA, King WM (2012) In-field 430 hyperspectral proximal sensing for estimating quality parameters of mixed pasture. Precision in 431 Agriculture 13, 351–369. 432 433 Ramoelo A, Skidmore AK, Schlerf M, Heitkönig IWA, Mathieu R, Cho MS (2013) Savanna grass 434 nitrogen to phosphorous ratio estimation using field spectroscopy and the potential for estimation 19 http://www.publish.csiro.au/nid/40.htm nly iew O v or Re F Page 21 of 67 Crop & Pasture Science 435 with imaging spectroscopy. International Journal of Applied Earth Observation and 436 Geoinformation 23, 334–343. 437 438 Safari H, Fricke T, Wachendorf M (2016) Determination of fibre and protein content in 439 heterogeneous pastures using field spectroscopy and ultrasonic sward height measurements. 440 Computers and Electronics in Agriculture 123, 256–263. 441 442 Skidmore AK, Ferwerda JG, Mutanga O, Van Wieren SE, Peel M, Grant RC, Prins HHT, Balcik 443 FB, Venus V (2010) Forage quality of savannas - Simultaneously mapping foliar protein and 444 polyphenols for trees and grass using hyperspectral imagery. Remote Sensing of Environment 114, 445 64–72. 446 447 Thenkabail PS, Smith RB, Pauw ED (2000) Hyperspectral vegetation indices and their 448 relationships with agricultural crop characteristics. Remote Sensing of Environment 71, 158–182. 449 450 Thorp KR, Dierig DA, French AN, Hunsaker DJ (2011) Analysis of hyperspectral reflectance data 451 for monitoring growth and development of lesquerella. Industrial Crops and Products 33, 524– 452 531. 453 454 Thorp KR, Wang G, Bronson KF, Badaruddin M, Mon J (2017) Hyperspectral data mining to 455 identify relevant canopy spectral features for estimating durum wheat growth, nitrogen status, and 456 grain yield. Computers and Electronics in Agriculture 136, 1-12. 457 20 http://www.publish.csiro.au/nid/40.htm ly On vie w e or R F Crop & Pasture Science Page 22 of 67 458 Thulin S, Hill MJ, Held A, Jones S, Woodgate P (2012) Hyperspectral determination of feed 459 quality constituents in temperate pastures: Effect of processing methods on predictive relationships 460 from partial least squares regression. International Journal of Applied Earth Observation and 461 Geoinformation 19, 322–334. 462 463 Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. 464 Remote sensing of environment 8, 127-150. 465 466 Walter A, Studer B, Kölliker R (2012) Advanced phenotyping offers opportunities for improved 467 breeding of forage and turf species. Annals of Botany 110, 1271–1279. 468 469 Wang Y, Wang D, Zhang G, Wang J (2013) Estimating nitrogen status of rice using the image 470 segmentation of G-R thresholding method. Field Crops Research 149, 33–39. 471 472 Woebbecke DM, Meyer GE, Von Bargen K, Mortensen DA (1995) Color indices for weed 473 identification under various soil, residue, and lighting conditions. Transactions of the ASAE 38(1), 474 259-269. 475 476 477 478 479 480 481 21 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 23 of 67 Crop & Pasture Science 482 Table 1. Phenotyping techniques used in the present study, the time of evaluation, its application, 483 advantages and disadvantages. 484 Phenotyping Time of technique evaluation* Applications Advantages Disadvantages Easy operation, low Evaluation of low Visual Visual observations of cost, evaluations can be number of 68 min different plant performed under genotypes, evaluation characteristics diverse conditions and evaluation environments subjected to human bias and fatigue Quantification of Easy operation, low Changes in ambient canopy cover and cost, greater number of light conditions limit calculation of Image analysis 40 min vegetation indices in plants evaluated, the visible determination of vegetation indices, electromagnetic several vegetation and data analysis is spectrum water indices moderately complex Canopy reflectance Moderately easy information in the operation, greater Low solar radiation visible and near infra- number of plants or cloudy days limit Hyperspectral red regions of the evaluated, analysis, sensor and 80 min electromagnetic determination of white reference analysis spectrum. Information nutritional and calibration is can be used to predict biochemical frequently needed, biochemical composition of data analysis is composition of plants leaf/canopy complex 485 * The time of evaluation refers to 200 Urochloa plants evaluated under the conditions of the 486 present study. 487 488 489 490 491 492 493 494 22 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Crop & Pasture Science Page 24 of 67 495 Table 2. Canopy cover and vegetation indices calculated from digital images of 200 Urochloa 496 hybrids. Vegetation indices were extracted using a naive Bayes multiclass machine learning 497 approach. Indices were then incorporated into a PLSR model to predict crude protein, dry weight 498 biomass and chlorophyll content. 499 Plant traits Name Formula* Reference CC** Canopy cover Nc/Nt - NGRDI Normalized green red difference index (g-r)/(g+r) Hunt et al., 2005 ExG Excess green index 2g-r-b Woebbecke et al., 1995 ExR Excess red index 1.3r-g Meyer et al., 1998 ExGR Excess green minus excess red ExG-ExR Camargo 2004 GR Green ratio g/(r+g+b) Tucker 1979 GLI Green leaf index (2g-r-b)/(2g+r+b) Louhaichi et al. 2001 *r, g and b denote the normalized pixel values of each channel on the RGB colour mode. ** No normalization was performed for the canopy cover quantification. Nc= total number of pixels representing the canopy, Nt= total number of pixels in the picture. 500 501 502 503 23 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 25 of 67 Crop & Pasture Science 504 Table 3. Plant traits measured in 200 Urochloa hybrids. 505 Trait Min Max Mean CV (%) Dry Weight (g plant-1) 6.74 64.1 30.22 34.81 Crude Protein (%) 6.76 21.58 11.23 19.68 Chlorophyll (mg g FW) 0.87 6.41 2.88 24.31 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 24 http://www.publish.csiro.au/nid/40.htm On ly vie w e or R F Crop & Pasture Science Page 26 of 67 521 Table 4. Regression coefficients of the fitted partial least square regression models of seven traits 522 extracted from digital image analysis. Positive and negative coefficients indicate positive and 523 negative influence on the prediction model, respectively. Traits* Dry weigth (g.plant-1) Crude protein (% DW) Chlorophyll (mg.g-1) CC 3.760545 -0.23114345 -0.01673706 NGRDI 9.948634 0.08600517 -0.03532585 ExG -14.3163 0.07760486 0.0730642 ExR -32.126212 -0.39106455 -0.07758547 ExGR 3.724492 0.26592971 0.10326555 GR -34.770799 -0.31158671 -0.03624834 GLI 80.87152 -0.31108316 -0.0357783 524 * CC= canopy cover, NGRDI= normalized green red difference index, ExG= excess green index, 525 ExR= excess red index, ExGR= excess green minus excess red, GR= green ratio and GLI= green 526 leaf index. 527 528 529 530 531 532 533 534 535 536 537 538 539 25 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 27 of 67 Crop & Pasture Science 540 Fig 1. Modeled versus measured dry weight, crude protein and chlorophyll content when fitting 541 partial least square regression models to relate each biophysical characteristic to visual evaluations 542 of biomass and greenness of 200 Urochloa hybrids. 543 544 60 18 5 (a) (b) (c) 50 16 4 14 40 12 3 30 10 2 20 8 RMSEP=8.47 g.plant-1 RMSEP=1.76% RMSEP=0.60 mg.g-1 10 6 1 10 20 30 40 50 60 6 8 10 12 14 16 18 1 2 3 4 5 Measured dry weight (g.plant-1) Measured CP (% DW) Measured chlorophyll (mg.g-1) 545 546 547 548 549 550 551 552 553 554 555 556 26 http://www.publish.csiro.au/nid/40.htm Modeled dry weight (g.plant-1) Modeled CP (% DW) Modeled chlorophyll (mg.g-1) nly O iew Re v r Fo Crop & Pasture Science Page 28 of 67 557 Fig 2. Modeled versus measured dry weight, crude protein and chlorophyll content when fitting 558 partial least square regression models to relate each biophysical characteristic to digital image 559 analysis of 200 Urochloa hybrids. 560 561 562 60 18 5 (a) (b) (c) 50 16 4 14 40 12 3 30 10 2 20 8 RMSEP=7.81 g.plant-1 RMSEP=1.53% RMSEP=0.57 mg.g-1 10 6 1 10 20 30 40 50 60 6 8 10 12 14 16 18 1 2 3 4 5 563 Measured dry weight (g.plant -1) Measured N (% DW) Measured chlorophyll (mg.g-1) 564 565 566 567 568 569 570 571 27 http://www.publish.csiro.au/nid/40.htm Modeled dry weight (g.plant-1) Modeled N (% DW) Modeled chlorophyll (mg.g-1) On ly w vie r R e Fo Page 29 of 67 Crop & Pasture Science 572 Fig 3. Modeled versus measured dry weight, crude protein and chlorophyll content when fitting 573 partial least square regression models to relate each biophysical characteristic to canopy spectral 574 reflectance of 200 Urochloa hybrids. 575 576 577 60 18 5 (a) (b) (c) 50 16 4 14 40 12 3 30 10 2 20 8 RMSEP=7.90 g.plant-1 RMSEP=1.63% RMSEP=0.55 mg.g-1 10 6 1 10 20 30 40 50 60 6 8 10 12 14 16 18 1 2 3 4 5 578 Measured dry weight (g.plant-1) Measured CP (% DW) Measured chlorophyll (mg.g-1) 579 580 581 582 583 584 585 586 28 http://www.publish.csiro.au/nid/40.htm Modeled dry weight (g.plant-1) Modeled CP (% DW) Modeled chlorophyll (mg.g-1) On ly w vie r R e Fo Crop & Pasture Science Page 30 of 67 587 Fig 4. Regression coefficients of the fitted partial least squares regression models for dry weight, 588 crude protein and chlorophyll content. The regression coefficients represents the contribution of 589 each spectral waveband to the overall prediction of each destructively-measured trait. 590 6 1.0 0.3 (a) (b) (c) 0.8 4 0.6 0.2 2 0.4 0.1 0.2 0 0.0 0.0 -0.2 -2 -0.4 -0.1 -4 -0.6 -0.8 -0.2 -6 -1.0 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000 591 nm nm nm 592 593 594 595 596 597 598 599 600 601 602 603 29 http://www.publish.csiro.au/nid/40.htm Regression dry weight Regression crude protein Regression chlorophyll nly O iew Re v r Fo Page 31 of 67 Crop & Pasture Science 604 Supplementary information 605 Supplementary Table 1. Different protocols of spectral data collection and their respective root 606 mean squared error of prediction (RMSEP) for crude protein, dry weight and chlorophyll content. 607 Collection Trait Factors¥ RMSEP Dry weight (g plant-1) 4 9.23 608 Day 1* Crude protein (%) 11 1.29 Chlorophyll (mg g FW) 4 0.49 Dry weight (g plant-1) 4 8.20 609 Day 2* Crude protein (%) 10 1.26 Chlorophyll (mg g FW) 7 0.50 610 Dry weight (g plant-1) 4 7.63 Day 3** Crude protein (%) 11 2.07 Chlorophyll (mg g FW) 5 0.54 611 Dry weight (g plant-1) 6 8.14 Day 4** Crude protein n (%) 2 1.21 612 Chlorophyll (mg g FW) 3 0.58 Dry weight (g plant-1) 6 7.90 All days Crude protein (%) 5 1.63 613 Chlorophyll (mg g FW) 5 0.55 614 Fifty plants were evaluated daily * One scan collected per plant. ** Ten scans collected per plant. 615 ¥ Number of factors for which the root mean squared error of prediction was minimized in the 616 model prediction. 617 618 619 620 621 622 623 624 625 30 http://www.publish.csiro.au/nid/40.htm ly On vie w e or R F Crop & Pasture Science Page 32 of 67 626 Supplementary Fig 1. Schematic representation of the observation geometry of hyperspectral 627 analysis (a) and digital image analysis (b) techniques evaluated in 200 Urochloa hybrids. White 628 circle positioned at the center of the plant canopy in figure (a) represents the 23-cm field of view 629 of the spectroradiometer at a distance of 50mm from the plant canopy. For the digital image 630 analysis (figure b), the whole plant, and not the 23-cm section, was used for segmentation and 631 further analysis. Scale bar= 10 cm. 632 633 634 635 636 637 638 639 640 641 31 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 33 of 67 Crop & Pasture Science 642 Supplementary Fig 2. Binary relationships and Pearson’s correlation coefficients between seven 643 plant traits extracted from digital images of 200 Urochloa hybrids. CC= canopy cover, NGRDI= 644 normalized green red difference index, ExG= excess green index, ExR= excess red index, ExGR= 645 excess green minus excess red, GR= green ration and GLI= green leaf index. Pearson’s correlation 646 coefficients are indicated with their statistical significance as follows: *P≤0.1, **P≤0.01, 647 ***P≤0.001. 648 649 650 32 http://www.publish.csiro.au/nid/40.htm On ly vie w e or R F Crop & Pasture Science Page 34 of 67 1 Proximal sensing of Urochloa grasses increases selection accuracy 2 3 Juan de la Cruz Jiménez 1*, Luisa Leiva2, Juan A. Cardoso3, Andrew N. French4 and Kelly R. Thorp4 4 5 1 UWA School of Agriculture and Environment, Faculty of Science, The University of Western Australia, 6 35 Stirling Highway, Crawley, WA 6009, Australia. 7 2 Department of plant breeding, Swedish University of Agricultural Sciences, Alnarp, Sweden. 8 3 Tropical Forage Program, International Center for Tropical Agriculture (CIAT), Km 17 Recta Cali – 9 Palmira, Colombia. 10 4 USDA-ARS, U.S. Arid Land Agricultural Research Center, 21881 N Cardon Ln, Maricopa, AZ 85138, 11 United States. 12 13 14 15 * Corresponding author: Juan de la Cruz Jiménez: juan.jimenezserna@research.uwa.edu.au 16 17 18 19 20 1 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Page 35 of 67 Crop & Pasture Science 21 Abstract 22 In the American Tropics, livestock production is highly restricted by forage availability. In 23 addition, the breeding and development of new forage varieties with outstanding yield and high 24 nutritional quality is often limited by a lack of resources and poor technology. Non-destructive 25 high throughput phenotyping offers a rapid and economical means to evaluate large numbers of 26 genotypes. In this study, visual assessments, digital color images, and spectral reflectance data 27 were collected from 200 Urochloa hybrids in a field setting. Partial least squares regression 28 (PLSR) was applied to relate visual assessments, vegetation indicesdigital image analysis and 29 spectral data with shoot dry weight, nitrogen (N) content (DW), crude protein (CP) and chlorophyll 30 content. Visual evaluations of biomass and greenness, digital color imaging, and hyperspectral 31 canopy data were collected in 68, 40 and 80 minutes, respectively. Root mean squared errors of 32 prediction for PLSR estimations of dry weight, NDW, CP, and chlorophyll were lower for 33 vegetation indices digital image analysis followed by hyperspectral analysis and visual 34 assessments. This study showed that digital color image and spectral analysis techniques have the 35 potential to improve precision and reduce time for tropical forage grass phenotyping. 36 Keywords: High throughput phenotyping, Urochloa, tropical forage grasses, plant breeding. 37 38 Introduction 39 Livestock productivity depends on forage availability and quality. Grasses from the Urochloa (syn. 40 Brachiaria) genus have been widely planted in the tropics as forage for grazing ruminant livestock 41 and are considered the most important forages in the American Tropics (Miles et al. 2004). The 42 International Center for Tropical Agriculture (CIAT) in Colombia conducts a Urochloa breeding 2 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Crop & Pasture Science Page 36 of 67 43 program aimed at developing hybrids with outstanding performance on infertile, acidic soils with 44 superior forage productivity and nutritional quality. The hybrid development process is difficult 45 and time consuming. In a regular, three-year breeding cycle (three years),, over 7000 hybrids are 46 produced by open pollination, but fewer than 2% of these are retained for full evaluation. 47 Approximately half of the population is discarded based on their reproductive mode (sexual 48 orgenotypes are discarded and apomictic hybrids are kept); another major proportion is discarded 49 based on visual evaluations; and only a limited number of hybrids (approximately 100) are finally 50 evaluated for different biotic and abiotic stresses (Valheria Castiblanco, personal communication). 51 The evaluation of genotypes is restricted mainly by insufficient economic resources and lacking 52 technology for rapid screening. 53 PeriodicForage grasses exhibiting great biomass production and high nutritional quality are key 54 determinants of the productivity of grazing animals (Herrero et al., 2013). Therefore, evaluations 55 of shoot biomass production and quality parameters (i.e. crude protein) are among the most 56 important traits for improvement in any forage grasses breeding program. However, owing to the 57 destructive nature of these measurements and the insufficient economic resources, the evaluation 58 of these parameters is postponed to final stages of the breeding program characterized by a reduced 59 number of genotypes. Instead of analytical measurements of forage quality and destructive 60 biomass harvests, periodic visual evaluations of plant performance (i.e., plant biomass and 61 greenness) over time has beenis traditionally used in the CIAT’s Urochloa breeding 62 programprograms to select superior plants at initial stages of the breeding scheme (Miles et al. 63 2004; Miles 2007). These visual evaluations are laborious and may not be sufficiently accurate 64 especially in breeding populations characterized by high genetic diversity and substantial genotype 65 x environment interaction (Walter et al. 2012). In this sense, the 3 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 37 of 67 Crop & Pasture Science 66 The use of new technologies for in-field non-destructive, high throughput phenotyping (HTP), 67 including digital image analysis and proximal hyperspectral sensing, representsoffers the 68 possibility to precisely evaluate a larger number of genotypes than feasible in traditional ways, 69 achieved at low cost, and implemented in a short period of time (Montes et al. 2007; White et al. 70 2012; Andrade-Sanchez et al. 2014). 71 Proximal hyperspectral sensing provides continuous information along the visual and near-infrared 72 electromagnetic spectrum. This information often relates to plant traits and has successfully been 73 studied in grasses to estimate quality parameters (Skidmore et al. 2010; Pullanagari et al. 2012; 74 Thulin et al. 2012; Ferner et al. 2015; Safari et al. 2016), diversity (Lopatin et al. 2017) and nutrient 75 content (Fava et al. 2009; Knox et al. 2012; Ramoelo et al. 2013; Adjorlolo et al. 2015; Foster et 76 al. 2017). Likewise, plant image analysis for phenotyping purposes is based on image 77 segmentation to separate the soil background (i.e., soil) and the plant for further quantification of 78 regions of interest (Tucker 1979; Woebbecke et al. 1995; Camargo 2004; Hunt et al. 2005). Digital 79 image analysis has also been used for quantifying vegetation indices related to plant growth, 80 greenness and nutritional status (Meyer and Camargo 2008; Hunt et al. 2013). Very few reports of 81 hyperspectral (Numata et al. 2008) or image analysis of Urochloa grasses exist in literature 82 (Jimenez et al. 2017). 83 No studies combining hyperspectral information and image analyses and comparing them to 84 conventional phenotyping methods is available. Moreover, hyperspectral data have not been used 85 to evaluate target traits in Urochloa breeding programs. In this study, in-field visual evaluations, 86 proximal hyperspectral data, and digital imaging were collected over canopies of Urochloa 87 hybrids. Partial least squares regression (PLSR) was used to relate hyperspectral information to 88 field measurements and machine learning (i.e. naive Bayes multiclass) was used to extract 4 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Crop & Pasture Science Page 38 of 67 89 vegetation indices from overhead canopy images. The objectives of this study were to: 1) develop 90 PLSR models for predicting NCP, forage dry weightDW, and chlorophyll content; 2) compute 91 vegetation indicesextract plant traits from digital image analysis to relate with NCP, forage dry 92 weightDW, and chlorophyll; and 3) demonstrate the superiority of HTP techniques as compared 93 to conventional visual evaluation of traits. Hyperspectral data or image analysis have not been 94 used to evaluate forage biomass, N or chlorophyll in Urochloa breeding programs. Moreover, no 95 studies combining hyperspectral information and image analysis and comparing it to conventional 96 phenotyping methods is available. Nitrogen, forage dry weightCrude protein, forage DW, and 97 chlorophyll content were chosen as target traits in this study as they are key parameters 98 determining both plant and cattle productivity. The development of HTP methodologies to 99 evaluate tropical forages will increase the number of hybrids evaluated per selection cycle, thus 100 permitting more intense selection and hence, genetic gain. The identification of new hybrids with 101 outstanding performance (i.e. higher biomass, greener and high N contentCP) will result in more 102 productive pastures with concomitant increases in milk and meat production in livestock systems 103 in tropical savannahs. 104 Materials and methods 105 Field experiment 106 Field data were obtained in August 2016 at the International Center for Tropical Agriculture 107 (CIAT) in Cali, Colombia (Lat. 3° 29’ N; Long. 76° 21’ W; altitude 965 m). Four thousand 108 Urochloa hybrids generated from crosses between the CIAT’s Urochloa breeding program 109 population SX12 and U. decumbens cv. Basilisk (CIAT 606) were initially planted in an andisol 110 soil in an augmented block design and spaced at 1.5x1.5 m. These plants were visually evaluated 111 four times (data not shown) for persistence, vigor and greenness after sequential cuttings every 5 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 39 of 67 Crop & Pasture Science 112 three months for one year. After that period, 200 hybrids were randomly selected for further visual 113 and HTP analysis. These 200 hybrids (and not, instead of the entire population), were selected for 114 economic and practical reasons. Visual evaluations of biomass and greenness, imaging and spectra 115 collection were performed after 3 months regrowthre-growth after cutting (see information below). 116 Plant heights ranged from 20 to 50 cm and shoot architecture varied from very prostrate towith 117 both decumbent and erectus growth. 118 Visual evaluation 119 Plant biomass was assessed using a nine-point visual scale, where level ‘9’ indicated high shoot 120 biomass with many tillers and leaves while level ‘1’ indicated stunted growth with fewer tillers 121 and leaves. Plant greenness was visually evaluated using a five-point visual scale, where level ‘5’ 122 represented intense dark green in all the leaves of the plant and level ‘1’ indicated yellow-pale 123 color in all leaves of the plant. This visual evaluation was conducted in 68 minutes one week before 124 the HTP measurements (Table 1). 125 Imaging collection and analysis 126 Individual, digital color images for each of the 200 hybrids were taken at 1.2 m above the soil 127 surface using a commercial digital 13-Megapixel camera (Coolpix P6000, Nikon, Japan) fixed to 128 a buggy tractor. Digital images were saved in 4224 x 3168 pixel JPG format and vegetation indices 129 were analyzed. The canopy cover (CC) and six vegetation indices including the normalized green 130 red difference index (NGRDI), excess green index (ExG), excess red index (ExR), excess green 131 minus excess red (ExGR), green ratio (GR) and green leaf index (GLI) were created using the 132 formulae as indicated in Table 2. The canopy cover was extracted by dividing the total number of 133 pixels representing the plant by the total number of pixels in each image. The vegetation indices 134 were extracted using naive Bayes multiclass. Briefly, the distribution of colors in a set of digital 135 color images (training set) was used to estimate the probability density function for each of the 6 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Crop & Pasture Science Page 40 of 67 136 different region of interest (i.e. plant and background). Once the regions of interest were defined, 137 in the training set, the machine learning process was applied to all images to accurately classify 138 and separate regions of interest; therefore. Therefore, every new pixel in an image was classified 139 into the previously defined plant and background classes. Every pixel characterizing the plant (but 140 not the background) was then decomposed into red, (R), green, (G), and blue (RGBB) channels 141 for . These channels were then normalized as follows: 142 𝑅 𝐺 𝐵 143 𝑟 = 𝑅 + 𝐺 + 𝐵;𝑔 = 𝑅 + 𝐺 + 𝐵;𝑏 = 𝑅 + 𝐺 + 𝐵 144 145 Normalization makes the variations of light intensities uniform across the spectral distribution, 146 thus, the individual color components (i.e. r,g,b) are independent from the overall brightness of 147 the image (Cheng et al. 2011). Normalized channels were further used for the quantification of the 148 vegetation indices. Seven vegetation indices including canopy cover, normalized red green 149 difference index (Tucker 1979), excess green index (Woebbecke et al. 1995), excess red index 150 (Meyer et al. 1998), excess green minus excess red (Camargo 2004), green ratio, and green leaf 151 index (Louhaichi et al. 2001) were calculated. (Table 2). Image analysis code was written in Java 152 and run in ImageJ software (National Institutes of Health, Bethesda, Maryland, USA). Images 153 were collectingcollected early in the morning to avoid beam solar radiation interferences. Digital 154 images contained the whole plant in addition to the 23-cm diameter field-of-view (as indicated 155 below for hyperspectral measurements, Supplementary Fig 1). The collection process took 40 156 minutes (Table 1). 157 158 Spectral collection and analysis 7 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 41 of 67 Crop & Pasture Science 159 Hyperspectral field data collections were performed on clear days at full sun exposure around 11 160 am by positioning a hand-held field spectroradiometer (Fieldspec 2, Malvern Panalytical, Malvern, 161 UK) directly above the plant canopy. The instrument was used with no foreoptics, which provided 162 a 25-degree full conical angle field-of-view. To avoid soil background noise, the bare optical input 163 was positioned at 50 cm from the top of the plant canopy to yield a 23-cm diameter field of view. 164 The instrument collected information in 750 narrow wavebands from 325 to 1075 nm in 1 nm 165 intervals. One or ten spectral scans were collected per plant and 50 plants were evaluated daily in 166 about 20 minutes. Differences in the collection protocols were deliberately done for comparison 167 purposes.tested to evaluate the most effective way. Different spectra collection 168 paradigmsprocesses (1 or 10 scans) did not yield significant differences in the root mean squared 169 error of prediction for the different traits evaluated (Supplementary FigTable 1). Radiometric 170 collections over a 99% Spectralon panel (Labsphere, Inc., North Sutton, New Hampshire) were 171 used to describe incoming solar irradiance throughout the data collection process. The radiometric 172 collections over the calibration panel were made before starting and after every five canopy scans 173 or when slight changes in solar irradiance due to cloud cover occurred. The values of the 174 Spectralon panel radiance were used to compute the canopy reflectance of the plants in each 175 wavelength over the time of spectra collection. Subsequently, 401 bands from 500 to 900 nm were 176 used for analysis. Based on visual inspection of reflectance spectra, these bands were typically less 177 noisy, as compared to bands at the bounds of detector sensitivity. Spectral collection process was 178 run in 80 minutes (Table 1). 179 Laboratory sample collections 180 Plants were immediately harvested after spectra collection. Aboveground tissue was removed by 181 cutting the area defined by a 23-cm diameter plastic circle co-located with the spectral data 8 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Crop & Pasture Science Page 42 of 67 182 collection area. Tissues were packed in plastic bags and stored on ice in a cooler in the field and 183 then transported to the laboratory. The extraction of chlorophyll was performed by adding 100 mg 184 of fresh tissue to 80% (v/v) cold methanol, and the mix was homogenized using a pestle in a mortar 185 until the plant residue was clear and the solution was uniform. This solution was then filtered and 186 absorbance was determined with a spectrophotometer (Synergy HT, Biotek, Winooski, USA). 187 Total chlorophyll concentration was calculated according to Lichtenthaler and Welburn (1983). 188 Dry weight (DW) was measured on an electronic balance (PB602S, Mettler Toledo, LLC, 189 Columbus, OH, USA) after oven-drying the samples for three days at 60 °C. Nitrogen 190 concentrations in the dry tissue were determined by using an automated nitrogen-carbon analyser 191 (Sercon, Crewe, UK). Urochloa and common bean (Phaseolus vulgaris) leaves were used as 192 reference tissues for confirmation of the reliability of the analyses. The crude protein content was 193 calculated by multiplying nitrogen content with 6.25, as protein is assumed to contain 16% 194 nitrogen on average. 195 196 Statistical analysis 197 Visual evaluations, vegetation indicesdigital image analysis, spectral reflectance, and plant trait 198 data were incorporated into a partial least squares regression (PLSR) algorithm (Mevik and 199 Wehrens 2007) within the R Project for Statistical Computing (http://www.r-project.org)). Models 200 were developed to estimatepredict each plant trait (i.e. CP, DW and chlorophyll) and to compare 201 the precision for prediction of each of the different methods of phenotyping. Partial least squares 202 regression was used in preference to conventional least squares analysis to reduce co-linearity 203 effects. Thorp et al. (2011) provided the details on the PLSR methodology used in the present 204 study. Briefly, if Y is an n×1 vector of responses (i.e. N, dry weightCP, DW or chlorophyll content) 9 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 43 of 67 Crop & Pasture Science 205 and X is an n-observation by p-variable matrix of predictors (a set of visual evaluations, vegetation 206 indicesdigital image analysis, or spectral reflectance data), PLSR aims to decompose X into a set 207 of A orthogonal scores such that the covariance with corresponding Y scores is maximized. The 208 X-weight and Y-loading vectors that result from the decomposition are used to estimate the vector 209 of regression coefficients, βPLS, such that 210 Y = X βPLS + ε 211 where ε is an n×1 vector of error terms. 212 Leave-one-out cross validation was used to test model predictions for independent data. Results 213 were reported for PLSR models with the number of factors that minimized the root mean squared 214 error of cross validation. Pearson’s correlation coefficients were calculated for the different traits 215 extracted from digital color images taken from Urochloa hybrids. 216 Results 217 In this study, visual evaluations of biomass and greenness, digital color imaging and hyperspectral 218 data were collected on 200 Urochloa hybrids in 68, 40 or 80 minutes, respectively (Table 1). High 219 variability for the different characteristics of dry weight, nitrogenDW, CP and chlorophyll content 220 evaluated on 200 Urochloa hybrids was found (Table 23). 221 Visual assessments 222 Partial least squares regressions for measured traits of DW, NCP and chlorophyll andbased on 223 visual evaluations of biomass and greenness performed with a root mean square error of prediction 224 (RMSEP) of 8.47 g plant-1, 1.76% and 0.60 mg g FW respectively (Fig 1). 225 Spectral data and digital image phenotyping 10 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Crop & Pasture Science Page 44 of 67 226 The PLSR models developed from seven vegetation indices the digital image analysis estimated 227 DW, NCP and chlorophyll with a RMSEP of 7.7981 g plant-1, 1.53% and 0.57 mg g FW, 228 respectively (Fig 2). Differences on the correlation coefficients among traits extracted from image 229 analysis indicated that including different indices into the model added independent information 230 to build stronger PLSR models (Supplementary Fig 2). The contribution of each trait extracted 231 from digital image analysis to the overall prediction of each destructively-measured trait is shown 232 in Table 4. The GLI had the stronger positive influence on the PLSR model for predicting DW. 233 The ExGR had the stronger positive influence on the PLSR model for predicting both CP and 234 chlorophyll content. 235 The fitted PLSR models developed from 401 wavebands of canopy spectral reflectance estimated 236 DW, NCP and chlorophyll with a RMSEP of 7.90 g plant-1, 1.63% and 0.55 mg g FW, respectively 237 (Fig 3). 238 The contribution of each spectral waveband to the overall prediction of each destructively- 239 measured trait is shown in the Fig. 4. In the PLSR model for DW, three bands characterized the 240 dry weight of Urochloa. Locallocal extrema in regression coefficients were found at 543, 668701 241 and 744674 nm, corresponding to visible green light, red light near the inflection band and NIR 242 radiationred light, respectively (Fig. 4a). Strong positive contribution to dry weightDW estimation 243 were with green light (543) and NIR (744700-750), and a strong negative contribution with red 244 light (668674-640). In the PLSR models for NCP and chlorophyll, regression coefficient plots 245 exhibited a noisy pattern with less defined extrema.strong positive contribution for traits estimation 246 in the visible green light (Fig 4b and c). The PLSR models for NCP contrasted wavebands in the 247 visible spectrum with positive contribution from wavebands around 513503 nm and negative 248 contributions from wavebands at 676678 nm. Wavebands at 600 nm and in the NIR contributed 11 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 45 of 67 Crop & Pasture Science 249 less to the model for N (Fig. 4b). RegressionSimilarly, regression coefficients for total chlorophyll 250 indicated strong positive contribution from NIR wavelengths and strong in the visible spectrum 251 around 504 nm and negative contribution throughout the visible wavebands, especially from 525 252 toat 625 nm, and at the red edge at 705643 nm (Fig. 4c). This is sensible considering visible light 253 absorption is increased with additional leaf chlorophyll. In the PLSR model for chlorophyll, local 254 extrema in regression coefficients were found at 567, 674, 705 and 763 nm, which correspond to 255 green light at the edge of yellow, red light, red light near the red inflection band and NIR radiation. 256 Discussion 257 The results from this study demonstrate that the current visual assessment methodology at initial 258 steps of the breeding cycle in the CIAT Urochloa breeding program can be improved by the use 259 ofusing non-destructive high throughput phenotypingHTP techniques. The use of colorColor 260 imaging, hyperspectral analysis, and PLSR models isare more precise and faster than visual 261 evaluations, thus increasing the number of plants evaluated in the tropical forage breeding 262 program. 263 Visual evaluations of plant growth and greenness (characteristics associated with N content, and 264 therefore CP and chlorophyll concentration in leaves) have traditionally been used to discard 265 Urochloa hybrids at initial stages of plant phenotyping. The visual evaluation of an entire breeding 266 population (i.e., 40007,000 hybrids) is a slow, costly and tedious process, and is often biased by 267 subjectivity and human fatigue, especially when phenotypic variation of such traits is high (Table 268 23). In this study, the estimation of DW, NCP, and chlorophyll content was more precisely and 269 consistently estimated by HTP techniques. Dry weight and NCP predictions were more accurate 270 using vegetation indicesdigital image analysis, followed by spectral analysis and visual 271 evaluations. Chlorophyll content was better estimated by the analysis of 401 spectral wavebands, 12 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Crop & Pasture Science Page 46 of 67 272 followed by color image analysis and finally visual evaluations (Fig 1, 2 and 3). Likewise, theThe 273 time required to run non-destructive HTP evaluations was considerably shorter by 28 minutes per 274 200 plants for color image analysis than visual evaluations, but longer by 12 minutes per 200 plants 275 in hyperspectral than in visual evaluations (Table 1). 276 The moderate trends in the relationship between Urochloa canopy imaging and reflectance and 277 measured DW, NCP and chlorophyll may indicate that the method is not appropriate for very 278 precise estimations of these traits. However, for breeding purposes where a large percentage of 279 hybrids are discarded without detailed evaluation due to scarce resources, a difference in DW of 280 7.90 g plant-1 or a difference of 1.63% in the NCP content of plants may be acceptable during 281 initial stages of plant breeding. This moderate trend between Urochloa canopy analysis and 282 measured traits in this study can be explained by dissimilarities in the Urochloa genotypes canopy 283 architecture of the Urochloa genotypes (Numata et al. 2008), as well as different growth patterns 284 during recovery from cutting. The further evaluation of breeding populations with contrasting 285 canopy architecture will improve the accuracy of the PLSR model to predict the targeted traits. 286 Nonetheless, by combining both digital image and hyperspectral analysis techniques, higher 287 precision accuracy for DW, CP and Chlorophyll content can be achieved. 288 The vegetation indices (see materials and methodsTable 2) extracted from color images of 200 289 Urochloa hybrids were originally developed to extractseparate green plants from the background 290 by extracting green and red colors from the image data to estimatedigital images. These indices 291 have been related to different plant characteristics including biomass, chlorophyll content and the 292 nutritional status of plants (Tucker 1979; Woebbecke et al. 1995; Camargo 2004; Hunt et al. 2005; 293 Meyer and Camargo 2008; Hunt et al. 2013; Lee and Lee 2013; Wang et al. 2013). In this study, 294 vegetation indices digital image analysis performed better than hyperspectral scanning analysis to 13 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 47 of 67 Crop & Pasture Science 295 estimate DW and NCP (Fig 2 and 3). Nonetheless, the use of spectral analysis over grasses 296 becomes more important when this technique is used to detect either nutritional or anti-nutritional 297 compounds (i.e. proteinmetabolisable energy, digestibility, fiber) that are better estimated with the 298 near-infrared regions of the electromagnetic spectra (Curran 1989; Pullanagari et al. 2012; Ferner 299 et al. 2015). In this sense, the use of digital color image analysis and hyperspectral analysis is 300 complementary because by using both techniques a diverse set of plant traits can accurately be 301 predicted. and by adding extra factors to the prediction model, higher prediction accuracy can be 302 achieved (cf. Numata et al. 2008). Future efforts will use data mining to fine-tune the spectral 303 bands included in the PLSR model (Thorp et al. 2017), which can reduce model error and improve 304 model fit statistics. Although testing multiple methods of analysis was not the intention of this 305 study, future research could also test other techniques (e.g., artificial neural networks) for relating 306 HTP measurements to plant traits. 307 The regression coefficients for the PLSR for DW and chlorophyll content obtained in this study 308 highlight that the key wavelengths for the prediction of these traits were locatedoccur in the green, 309 red, red -edge and NIR regions of the electromagnetic spectrum (Fig 4). Previous hyperspectral 310 studies have highlighted those regions as being highly representative for dry mass and chlorophyll 311 content in plants (Lichtenthaler et al. 1996; Thenkabail et al. 2000; Mutanga and Skidmore 2004; 312 Fava et al. 2009; Thorp et al. 2011; Adjorlolo et al. 2015; Dou et al, 2018). Although some 313 similarities were found between wavebands among the different traits, the general regression 314 coefficients differed among the traits, thus demonstrating that the reflectance data in a given 315 waveband contributed differently toward the estimation of a given trait. Given the logistical burden 316 to collect and analyze hyperspectral scans, the identification of informative key bands associated 317 with each evaluated trait can improve the HTP process (Thorp et al. 2017). Results from this study 14 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Crop & Pasture Science Page 48 of 67 318 will help guide selection of optimal bands in the construction of multispectral sensors tailored to 319 predict specific traits of interest in tropical forage breeding programs. 320 The PLSR models for predicting DW, CP and chlorophyll content can be now used to evaluate the 321 next generation of hybrids from the same Urochloa gene pool (i.e. U. ruziziensis – U. brizantha – 322 U. decumbens). The accuracy of this prediction models relies on collection protocols similar to the 323 explained in the Materials and Methods section and evaluations on plants with comparable growth 324 characteristics as the hybrids evaluated here (i.e. about three months after regrowth). The 325 prediction accuracy will likely be reduced on larger plants with higher biomass (Hill 2004) and a 326 greater proportion of senescent leaves (Asner 1998). The development of more precise PLSR 327 models to predict variables of interest in a breeding program requires an ongoing effort. The 328 collection of ground data every year while making improvements to standardize collection 329 protocols and incorporate wider range of genotypes will result in more accurate and robust models. 330 Larger data sets will increase estimation precision. 331 332 Conclusions 333 In this study, 200 Urochloa hybrids were successfully monitored in 40 and 80 minutes by digital 334 imaging and spectral analysis, respectively (Table 1). At this pace, more than 1000 Urochloa 335 hybrids cancould be evaluated in a period of less than 7 hours. This means morethat forage biomass 336 and quality in a high number of genotypes couldwould be reliably evaluated with minimal 337 increased acquisition costs (comparedrelative to destructive harvest).. This demonstrates the 338 superiority of HTP techniques as compared to conventional visual evaluation of traits. The PLSR 339 models for predicting CP, forage DW, and chlorophyll content developed in this study supports 15 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 49 of 67 Crop & Pasture Science 340 the evaluation of higher numbers of genotypes at initial stages of the breeding program. The greater 341 numbernumbers of plants evaluated reliably every year in the Urochloa breeding program, the 342 greater the genetic gain will be. Therefore, the use of image analysis and hyperspectral monitoring 343 over Urochloa hybrids canopies will benefit the on-going breeding program. Likewise, theThe 344 application of this methodologyHTP method could be of great help in rural remote areas without 345 appropriatelacking facilities to perform destructive harvest and plant chemical analysis. Additional 346 studies on Urochloa plants with contrasting architectures need to be performed to optimize PLSR 347 models. Moreover, more careful field measurements over plants with similar regrowth capacity 348 are requiredResearch is underway to improve the prediction models. Furthermore,utility of 349 proximal sensing by considering a greater range of canopy architectural configurations and 350 evaluating the potential to assess nutritional quality traits, including proteincharacteristics such as 351 metabolisable energy, fiber, digestibility and non-digestible fractions of the forage (, lignin and 352 cellulose) must be evaluated through proximal hyperspectral sensing to improve phenotyping 353 fractions in Urochloa grasses. 354 355 Conflict of interest 356 The authors have no conflicts of interest to declare. 357 Acknowledgements 358 We thank Dr. John W. Miles for helpful suggestions and comments to early version of this 359 manuscript. JCJ is grateful to the USDA Foreign Agricultural Service, Norman Borlaug 360 Fellowship Program for a training fellowship. JCJ thanks the CIAT’s Young Scientist Award 2016 361 Program for travel assistance. This work was partially undertaken as part of the CGIAR Research 16 http://www.publish.csiro.au/nid/40.htm On ly w vie r R e Fo Crop & Pasture Science Page 50 of 67 362 Program on Livestock. We thank all donors that globally support our work through their 363 contributions to the CGIAR system. 364 References 365 Adjorlolo C, Mutanga O, Cho MA (2015) Predicting C3 and C4 grass nutrient variability using in 366 situ canopy reflectance and partial least squares regression. International Journal of Remote 367 Sensing 36(6), 1743–1761. 368 369 Andrade-Sanchez P, Gore MA, Heun JT, Thorp KR, Carmo-Silva AE, French AN, Salvucci ME, 370 White JW (2014) Development and evaluation of a field-based high throughput phenotyping 371 platform. Functional Plant Biology 41, 68–79. 372 373 Camargo JN (2004) A combined statistical—soft computing approach for classification and 374 mapping weed species in minimum tillage systems. University of Nebraska, Lincoln, NE 375 376 Currant PJ (1989) Remote sensing of foliar chemistry. Remote Sensing of the Environment 30, 377 271–278. 378 379 Dou Z, Cui L, Li J, Zhu Y, Gao C, Pan X, Lei Y, Zhang M, Zhao X, Li W (2018) Hyperspectral 380 estimation of the chlorophyll content in short-term and long-term restorations of mangrove in 381 Quanzhou Bay Estuary, China. Sustainability 10, 1127; doi:10.3390/su10041127 382 17 http://www.publish.csiro.au/nid/40.htm On ly vie w e or R F Page 51 of 67 Crop & Pasture Science 383 Fava F, Colombo R, Bocchi S, Meroni M, Sitzia M, Fois N, Zucca C (2009) Identification of 384 hyperspectral vegetation indices for Mediterranean pasture characterization. International Journal 385 of Applied Earth Observation and Geoinformation 11, 233–243. 386 387 Ferner J, Linstadter A, Sudekum KH, Schmidtlein S (2015) Spectral indicators of forage quality 388 in West Africa’s tropical savannas. International Journal of Applied Earth Observation and 389 Geoinformation 41, 99–106. 390 391 Foster AJ, Kakani VG, Mosali J (2017) Estimation of bioenergy crop yield and N status by 392 hyperspectral canopy reflectance and partial least square regression. Precision Agriculture 18, 393 192–209. 394 395 Herrero M, Havlík P, Valin H, Notenbaert A, Rufino MC, Thornton PK, Blümmel M, Weiss F, 396 Grace D, Obersteiner M (2013) Biomass use, production, feed efficiencies, and greenhouse gas 397 emissions from global livestock systems. Proceedings of the National Academy of Science of the 398 United States of America 110(52), 20888–20893. 399 400 Hunt Jr ER, Cavigelli M, Daughtry CST, McMurtrey J, Walthall CL (2005) Evaluation of digital 401 photography from model aircraft for remote sensing of crop biomass and nitrogen status. Precision 402 Agriculture 6, 359–378. 403 404 Hunt Jr ER, Doraiswamy PC, McMurtrey J, Daughtry CST, Perry EM, Akhmedov B (2013) A 405 visible band index for remote sensing leaf chlorophyll content at the canopy scale. International 406 Journal of Applied Earth Observation and Geoinformation 21, 103–112. 18 http://www.publish.csiro.au/nid/40.htm ly w O n vie or Re F Crop & Pasture Science Page 52 of 67 407 408 Jiménez JC, Cardoso JA, Leiva LF, Gil J, Forero MG, Worthington ML, Miles JW, Rao IM (2017) 409 Non-destructive phenotyping to identify brachiaria hybrids tolerant to waterlogging stress under 410 field conditions. Frontiers in Plant Science 8, 167. doi: 10.3389/fpls.2017.00167 411 412 Knox NM, Skidmore AK, Prins HHT, Heitkönig IMA, Slotow R, van der Waal C, de Boer WF 413 (2012) Remote sensing of forage nutrients: Combining ecological and spectral absorptiond feature 414 data. ISPRS Journal of Photogrammetry and Remote Sensing 72, 27–35. 415 416 Meyer GE, Hindman TW, Lakshmi K (1998) Machine vision detection parameters for plant 417 species identification. In ‘Precision Agriculture and Biological Quality’. (Eds GE Meyer, JA De 418 Shazer) pp. 327–335. (Proceedings of SPIE. vol. 3543, Bellingham, WA) 419 420 Meyer GE, Camargo J (2008) Verification of color vegetation indices for automated crop imaging 421 applications. Computers and Electronics in Agriculture 63, 282–293. 422 423 Mevik BH, Wehrens R (2007) The pls Package: Principal component and partial least squares 424 regression in R. Journal of Statistical Software 18(2), 1-23. 425 426 Miles JW, Do Valle CB, Rao IM, Euclides VPB (2004) Brachiariagrasses. American Society of 427 Agronomy, Crop Science Society of America, Soil Science Society of America, 677 S. Segoe Rd., 428 Madison, WI 53711, USA. Warm Season (C4) Grasses, Agronomy Monograph no. 45. 429 19 http://www.publish.csiro.au/nid/40.htm nly iew O v or Re F Page 53 of 67 Crop & Pasture Science 430 Miles JW (2007) Apomixis for cultivar development in tropical forage grasses. Crop Science 431 47(S3), S238–S249. 432 433 Montes JM, Melchinger AE, Reif JC (2007) Novel throughput phenotyping platforms in plant 434 genetic studies. Trends in Plant Science 12, 433–436. 435 436 Mutanga O, Skidmore AK (2004) Narrow band vegetation indices overcome the saturation 437 problem in biomass estimation. International Journal of Remote Sensing 25(19), 3999–4014. 438 439 Numata I, Roberts DA, Chadwick OA, Schimel JP, Galvão LS, Soares JV (2008) Evaluation of 440 hyperspectral data for pasture estimate in the Brazilian Amazon using field and imaging 441 spectrometers. Remote Sensing of Environment 112, 1569–1583. 442 443 Lee KJ, Lee BW (2013) Estimation of rice growth and nitrogen nutrition status using color digital 444 camera image analysis. European Journal of Agronomy 48, 57–65. 445 446 Lichtenthaler HK, Wellburn AA (1983) Determination of total carotenoids and chlorophylls a and 447 b of leaf extracts in different solvents. Biochemical society transactions 603, 591–592. 448 449 Lichtenthaler HK, Gitelson A, Lang M (1996) Non-destructive determination of chlorophyll 450 content of leaves of a green and an aurea mutant of tobacco by reflectance measurements. Journal 451 of Plant Physiology 148, 483–493. 452 20 http://www.publish.csiro.au/nid/40.htm ly On vie w e or R F Crop & Pasture Science Page 54 of 67 453 Lopatin J, Fassnacht FE, Kattenborn T, Schmidtlein S (2017) Mapping plant species in mixed 454 grassland communities using close range imaging spectroscopy. Remote Sensing of Environment 455 201, 12–23. 456 457 Louhaichi M, Borman MM, Johnson DE (2001) Spatially located platform and aerial photography 458 for documentation of grazing impacts on wheat. Geocarto International 16(1), 65-70. 459 460 Pullanagari RR, Yule IJ, Tuohy MP, Hedley MJ, Dynes RA, King WM (2012) In-field 461 hyperspectral proximal sensing for estimating quality parameters of mixed pasture. Precision in 462 Agriculture 13, 351–369. 463 464 Ramoelo A, Skidmore AK, Schlerf M, Heitkönig IWA, Mathieu R, Cho MS (2013) Savanna grass 465 nitrogen to phosphorous ratio estimation using field spectroscopy and the potential for estimation 466 with imaging spectroscopy. International Journal of Applied Earth Observation and 467 Geoinformation 23, 334–343. 468 469 Safari H, Fricke T, Wachendorf M (2016) Determination of fibre and protein content in 470 heterogeneous pastures using field spectroscopy and ultrasonic sward height measurements. 471 Computers and Electronics in Agriculture 123, 256–263. 472 473 Skidmore AK, Ferwerda JG, Mutanga O, Van Wieren SE, Peel M, Grant RC, Prins HHT, Balcik 474 FB, Venus V (2010) Forage quality of savannas - Simultaneously mapping foliar protein and 475 polyphenols for trees and grass using hyperspectral imagery. Remote Sensing of Environment 114, 476 64–72. 21 http://www.publish.csiro.au/nid/40.htm ly w O n vie r R e Fo Page 55 of 67 Crop & Pasture Science 477 478 Thenkabail PS, Smith RB, Pauw ED (2000) Hyperspectral vegetation indices and their 479 relationships with agricultural crop characteristics. Remote Sensing of Environment 71, 158–182. 480 481 Thorp KR, Dierig DA, French AN, Hunsaker DJ (2011) Analysis of hyperspectral reflectance data 482 for monitoring growth and development of lesquerella. Industrial Crops and Products 33, 524– 483 531. 484 485 Thorp KR, Wang G, Bronson KF, Badaruddin M, Mon J (2017) Hyperspectral data mining to 486 identify relevant canopy spectral features for estimating durum wheat growth, nitrogen status, and 487 grain yield. Computers and Electronics in Agriculture 136, 1-12. 488 489 Thulin S, Hill MJ, Held A, Jones S, Woodgate P (2012) Hyperspectral determination of feed 490 quality constituents in temperate pastures: Effect of processing methods on predictive relationships 491 from partial least squares regression. International Journal of Applied Earth Observation and 492 Geoinformation 19, 322–334. 493 494 Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. 495 Remote sensing of environment 8, 127-150. 496 497 Walter A, Studer B, Kölliker R (2012) Advanced phenotyping offers opportunities for improved 498 breeding of forage and turf species. Annals of Botany 110, 1271–1279. 499 22 http://www.publish.csiro.au/nid/40.htm ly On vie w e or R F Crop & Pasture Science Page 56 of 67 500 Wang Y, Wang D, Zhang G, Wang J (2013) Estimating nitrogen status of rice using the image 501 segmentation of G-R thresholding method. Field Crops Research 149, 33–39. 502 503 Woebbecke DM, Meyer GE, Von Bargen K, Mortensen DA (1995) Color indices for weed 504 identification under various soil, residue, and lighting conditions. Transactions of the ASAE 38(1), 505 259-269. 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 23 http://www.publish.csiro.au/nid/40.htm nly O iew Re v r Fo Page 57 of 67 Crop & Pasture Science 524 525 526 Table 1. Phenotyping techniques used in the present study, the time of evaluation, its application, 527 advantages and disadvantages. 528 Phenotyping Time of technique evaluation* Applications Advantages Disadvantages Easy operation, low Evaluation of low Visual Visual observations of cost, evaluations can be number of evaluation 68 min different plant performed under genotypes, characteristics diverse conditions and evaluation environments subjected to human bias and fatigue Changes in ambient Quantification of canopy Easy operation, low cost, greater number of light conditions limit cover and vegetation calculation of Image analysis 40 min indices in the visible plants evaluated, determination of vegetation indices, electromagnetic several vegetation and data analysis is spectrum water indices moderately complex Canopy reflectance information in the visible Moderately easy and near infra-red operation, greater Low solar radiation or cloudy days limit regions of the number of plants evaluated, analysis, sensor and Hyperspectral 80 min electromagnetic analysis spectrum. Information determination of white reference can be used to predict nutritional and calibration is biochemical frequently needed, biochemical data analysis is compositionscomposition composition of leaf/canopy complexof plants 529 * The time of evaluation refers to 200 Urochloa plants evaluated under the conditions of the 530 present study. 531 532 533 534 535 536 24 http://www.publish.csiro.au/nid/40.htm On ly vie w e or R F Crop & Pasture Science Page 58 of 67 537 538 539 Table 2. Canopy cover and vegetation indices calculated from digital images of 200 Urochloa 540 hybrids. Vegetation indices were extracted using a naive Bayes multiclass machine learning 541 approach. Indices were then incorporated into a PLSR model to predict crude protein, dry weight 542 biomass and chlorophyll content. 543 Plant traits Name Formula* Reference CC** Canopy cover Nc/Nt - NGRDI Normalized green red difference index (g-r)/(g+r) Hunt et al., 2005 ExG Excess green index 2g-r-b Woebbecke et al., 1995 ExR Excess red index 1.3r-g Meyer et al., 1998 ExGR Excess green minus excess red ExG-ExR Camargo 2004 GR Green ratio g/(r+g+b) Tucker 1979 GLI Green leaf index (2g-r-b)/(2g+r+b) Louhaichi et al. 2001 *r, g and b denote the normalized pixel values of each channel on the RGB colour mode. ** No normalization was performed for the canopy cover quantification. Nc= total number of pixels representing the canopy, Nt= total number of pixels in the picture. 544 545 25 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 59 of 67 Crop & Pasture Science 546 547 548 Table 3. Plant traits measured in 200 Urochloa hybrids. 549 Trait Min Max Mean CV (%) Dry Weight (g plant-1) 6.74 64.1 30.22 34.81 NitrogenCrude Protein (%) 6.76 21.58 11.23 19.68 Chlorophyll (mg g FW) 0.87 6.41 2.88 24.31 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 26 http://www.publish.csiro.au/nid/40.htm nly iew O v or Re F Crop & Pasture Science Page 60 of 67 565 Table 4. Regression coefficients of the fitted partial least square regression models of seven traits 566 extracted from digital image analysis. Positive and negative coefficients indicate positive and 567 negative influence on the prediction model, respectively. Traits* Dry weigth (g.plant-1) Crude protein (% DW) Chlorophyll (mg.g-1) CC 3.760545 -0.23114345 -0.01673706 NGRDI 9.948634 0.08600517 -0.03532585 ExG -14.3163 0.07760486 0.0730642 ExR -32.126212 -0.39106455 -0.07758547 ExGR 3.724492 0.26592971 0.10326555 GR -34.770799 -0.31158671 -0.03624834 GLI 80.87152 -0.31108316 -0.0357783 568 * CC= canopy cover, NGRDI= normalized green red difference index, ExG= excess green index, 569 ExR= excess red index, ExGR= excess green minus excess red, GR= green ratio and GLI= green 570 leaf index. 571 572 573 574 575 576 577 578 579 580 581 582 583 27 http://www.publish.csiro.au/nid/40.htm nly w O vie r R e Fo Page 61 of 67 Crop & Pasture Science 584 Fig 1. Modeled versus measured dry weight, nitrogencrude protein and chlorophyll content when 585 fitting partial least square regression models to relate each biophysical characteristic to visual 586 evaluations of biomass and greenness of 200 Urochloa hybrids. 587 588 60 18 5 (a) (b) (c) 50 16 4 14 40 12 3 30 10 2 20 8 RMSEP=8.47 g.plant-1 RMSEP=1.76% RMSEP=0.60 mg.g-1 10 6 1 10 20 30 40 50 60 6 8 10 12 14 16 18 1 2 3 4 5 Measured dry weight (g.plant-1) Measured CP (% DW) Measured chlorophyll (mg.g-1) 589 590 591 592 593 594 595 596 597 598 599 600 28 http://www.publish.csiro.au/nid/40.htm Modeled dry weight (g.plant-1) Modeled CP (% DW) Modeled chlorophyll (mg.g-1) ly w O n ev ie or R F Crop & Pasture Science Page 62 of 67 601 Fig 2. Modeled versus measured dry weight, nitrogencrude protein and chlorophyll content when 602 fitting partial least square regression models to relate each biophysical characteristic to vegetation 603 indicesdigital image analysis of 200 Urochloa hybrids. 604 605 606 60 18 5 (a) (b) (c) 50 16 4 14 40 12 3 30 10 2 20 8 RMSEP=7.81 g.plant-1 RMSEP=1.53% RMSEP=0.57 mg.g-1 10 6 1 10 20 30 40 50 60 6 8 10 12 14 16 18 1 2 3 4 5 607 Measured dry weight (g.plant -1) Measured N (% DW) Measured chlorophyll (mg.g-1) 608 609 610 611 612 613 614 615 29 http://www.publish.csiro.au/nid/40.htm Modeled dry weight (g.plant-1) Modeled N (% DW) Modeled chlorophyll (mg.g-1) ly w O n ev ie or R F Page 63 of 67 Crop & Pasture Science 616 Fig 3. Modeled versus measured dry weight, nitrogencrude protein and chlorophyll content when 617 fitting partial least square regression models to relate each biophysical characteristic to canopy 618 spectral reflectance of 200 Urochloa hybrids. 619 620 621 60 18 5 (a) (b) (c) 50 16 4 14 40 12 3 30 10 2 20 8 RMSEP=7.90 g.plant-1 RMSEP=1.63% RMSEP=0.55 mg.g-1 10 6 1 10 20 30 40 50 60 6 8 10 12 14 16 18 1 2 3 4 5 622 Measured dry weight (g.plant-1) Measured CP (% DW) Measured chlorophyll (mg.g-1) 623 624 625 626 627 628 629 630 30 http://www.publish.csiro.au/nid/40.htm Modeled dry weight (g.plant-1) Modeled CP (% DW) Modeled chlorophyll (mg.g-1) nly w O vie r R e Fo Crop & Pasture Science Page 64 of 67 631 Fig 4. Regression coefficients of the fitted partial least squares regression models for dry weight, 632 nitrogen and chlorophyll contentcrude protein and chlorophyll content. The regression coefficients 633 represents the contribution of each spectral waveband to the overall prediction of each 634 destructively-measured trait. 635 6 1.0 0.3 (a) (b) (c) 0.8 4 0.6 0.2 2 0.4 0.1 0.2 0 0.0 0.0 -0.2 -2 -0.4 -0.1 -4 -0.6 -0.8 -0.2 -6 -1.0 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000 400 500 600 700 800 900 1000 636 nm nm nm 637 638 639 640 641 642 643 644 645 646 647 31 http://www.publish.csiro.au/nid/40.htm Regression dry weight Regression crude protein Regression chlorophyll nly O iew Re v r Fo Page 65 of 67 Crop & Pasture Science 648 649 Supplementary information 650 Supplementary Table 1. Different protocols of spectral data collection and their respective root 651 mean squared error of prediction (RMSEP) for nitrogencrude protein, dry weight and chlorophyll 652 content. 653 Collection Trait Factors¥ RMSEP Dry weight (g plant-1) 4 9.23 654 Day 1* NitrogenCrude protein (%) 11 1.29 Chlorophyll (mg g FW) 4 0.49 Dry weight (g plant-1) 4 8.20 655 Day 2* NitrogenCrude protein (%) 10 1.26 Chlorophyll (mg g FW) 7 0.50 656 Dry weight (g plant-1) 4 7.63 Day 3** NitrogenCrude protein (%) 11 2.07 Chlorophyll (mg g FW) 5 0.54 657 Dry weight (g plant-1) 6 8.14 Day 4** NitrogenCrude protein n 658 (%) 2 1.21 Chlorophyll (mg g FW) 3 0.58 Dry weight (g plant-1) 6 7.90 659 All days NitrogenCrude protein (%) 5 1.63 Chlorophyll (mg g FW) 5 0.55 660 Fifty plants were evaluated daily * One scan collected per plant. ** Ten scans collected per plant. 661 ¥ Number of factors for which the root mean squared error of cross validationprediction was 662 minimized in the model prediction. 663 664 665 666 667 668 669 670 32 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Crop & Pasture Science Page 66 of 67 671 672 Supplementary Fig 1. Schematic representation of the observation geometry of hyperspectral 673 analysis (a) and digital image analysis (b) techniques evaluated in 200 Urochloa hybrids. White 674 circle positioned at the center of the plant canopy in figure (a) represents the 23-cm field of view 675 of the spectroradiometer at a distance of 50mm from the plant canopy. For the digital image 676 analysis (figure b), the whole plant, and not the 23-cm section, was used for segmentation and 677 further analysis. Scale bar= 10 cm. 678 679 680 681 682 683 684 685 686 33 http://www.publish.csiro.au/nid/40.htm ly w O n ev ie or R F Page 67 of 67 Crop & Pasture Science 687 688 Supplementary Fig 2. Binary relationships and Pearson’s correlation coefficients between seven 689 plant traits extracted from digital images of 200 Urochloa hybrids. CC= canopy cover, NGRDI= 690 normalized green red difference index, ExG= excess green index, ExR= excess red index, ExGR= 691 excess green minus excess red, GR= green ration and GLI= green leaf index. Pearson’s correlation 692 coefficients are indicated with their statistical significance as follows: *P≤0.1, **P≤0.01, 693 ***P≤0.001. 694 695 696 34 http://www.publish.csiro.au/nid/40.htm On ly vie w e or R F