1 Advanced Spatial Analytics for Policy Support Use cases from One CGIAR October 2025 Credit: ©2023CIAT/ElizabethRamirezPerez 2 Table of Contents 1. Introduction 8 1.1 Background ...................................................................................................................................................................... 8 1.2 Scope and Objectives of this synthesis report .................................................................................................................. 9 2. Spatial analytics for policy support: history and building blocks 11 2.1. Basic concepts and terminologies .................................................................................................................................. 12 2.2 The evolution of spatial analytics for policy support ...................................................................................................... 12 2.2.1 GIS for guiding public investment and service delivery. ......................................................................................... 13 2.2.2 Economic geography and spatial economics for regional development policy ....................................................... 14 2.2.3 Integrated Economic modelling with spatial explicit inputs .................................................................................... 15 2.3 Building blocks of spatial analytics for policy ............................................................................................................... 18 3. How can spatial analytics support policy? 20 3.1. Contextualizing key spatial covariates for policy making ........................................................................................ 22 3.2. Contextualizing local influence of mega trends ........................................................................................................ 22 3.3. Assessing (local) effectiveness of interventions ....................................................................................................... 23 3.4. Inform locally tailored actions and prioritization ..................................................................................................... 24 3.5. Easier visualizing & disseminating .......................................................................................................................... 25 4. Use cases of applying spatial analytics for policy support 27 Use Case 1: Fertilizer Investment Risk Explorer (FIRE) ..................................................................................................... 28 Use case 2: Spatial Production Allocation Model (SPAM)................................................................................................... 33 Use case 3: Spatial mapping of crop vulnerability for targeting investment ........................................................................ 36 Use case 4: Improving the Spatial Granularity in Impact Evaluation ................................................................................... 41 Use case 5: Simulated Land Use/Land Cover Change ......................................................................................................... 43 Use Case 6: Spatial Dimensions of Poverty and Agricultural Productivity: A Spatial Econometric Analysis ...................... 48 Use case 7: Spatial Downscaling future food demand for local food policy making ........................................................... 55 Use case 8: PLaSA as a spatial online supporting tool for policy ......................................................................................... 58 Use case 9: Spatial assessment of Landscape complexity: implications for agricultural policies ........................................ 62 Use case 10: Assessing potential impact of avian influenza on poultry in West Africa ........................................................ 69 Use case 11: Water Conflict Future: governance challenges for climate resilient peace ...................................................... 73 5. Way forward 77 5.1 Future integration of spatial analytics into economic models for policy support ........................................................... 76 5.1.1 Equilibrium models ................................................................................................................................................. 77 5.1.2 Farm/household models ........................................................................................................................................... 78 5.2 Improving capacity to harness the potential of spatial analytics for policy support. ...................................................... 80 References 82 3 Table of Figures Figure 1: Different ways spatial analytics can support policy questions. ................................................................ 9 Figure 2: Evolution of spatial analytics for policy support (source: authors) ....................................................... 12 Figure 3: Mapping each use case on impact areas and AOW1 cluster. ................................................................. 27 Figure 4: Spatial and regional patterns of fertilizer profitability and risk ............................................................. 30 Figure 5: Harvested area maps for maize in rainfed (a), irrigated (b), and all (c) farming systems, SPAM 2020. 35 Figure 6: Map showing the percentage of each pixel that is rainfed maize in 2020 (Source: SPAM2020v2r0.) .. 38 Figure 7: Projected yield change for rainfed maize for median climate model, RCP 7.0 (medium-high emissions scenario, 2005-2050) Source: Authors using DSSAT. Notes: At least 5% cropland in pixel. Maize represents at least 25% of cropland. ........................................................................................................................................... 39 Figure 8: Hotspots for rainfed maize in Africa for median climate model, RCP 7.0 (medium-high emissions scenario). Source: Authors using DSSAT. Notes: At least 5% cropland in pixel. Maize represents at least 25% of cropland................................................................................................................................................................. 39 Figure 9: Hotspots for rainfed maize in Africa for worst-case climate model, RCP 7.0 (medium-high emissions scenario) Source: Authors using DSSAT. Notes: At least 5% cropland in pixel. Maize represents at least 25% of cropland................................................................................................................................................................. 39 Figure 10: Frequency of 1-in-20-year low maize yield events in Southern Africa: comparing frequency in the 2020s to the 2060s under various emissions scenarios. Source: Thomas, TS, Robertson RD, Strzepek K, and Arndt C. (2022) Extreme Events and Production Shocks for Key Crops in Southern Africa Under Climate Change. Front. Clim. 4 .................................................................................................................................................................. 40 Figure 11: Spatial variation in impacts of watching SSU weather and farming news .......................................... 42 Figure 12: Land use changes ................................................................................................................................. 46 Figure 13: GDP per capita, PPP (2021) ................................................................................................................. 51 Figure 14: Poverty headcount ratio at $2.15, 2021 ............................................................................................... 52 Figure 15: Population density (people per sq. km), 2021 ...................................................................................... 52 Figure 16: Rice production (spam, 2020) .............................................................................................................. 52 Figure 17: Dietary diversity index 2005-2030 in Kenya (SSP2, RCP 4.5) ........................................................... 56 Figure 18: Front page, modules, and an example of PlaSA .................................................................................. 60 Figure 19: Changes in complexity of agricultural landscapes between 2005 and 2050 under the Mid scenario .................................................................................................................................................. 67 Figure 20: Flyways and outbreak areas in West Africa ......................................................................... 71 Figure 21: Estimated increase in water conflict fatality (monthly average): 2030-2040 ...................... 75 4 Contributors Section 1: Chun Song. Section 2: Chun Song, Athanasios Petsakos, Abhijeet Mishra. Section 3: Chris Mwungu, Gloria Mbabazi. Use case 1 Fertilizer response: Bisrat Gebrekidan, Jordan Chamberlin, Maxwell Mkondiwa. Use case 2 Spatial Production Allocation Model (SPAM): Shuang Zhou, Zhe Guo, Liangzhi You. Use case 3 Spatial mapping of crop vulnerability for targeting investment: Tim Thomas. Use case 4 Improving the Spatial Granularity in Impact Evaluation: Chris Mwungu, Felix Otieno, Agnes Wanjau, Anirudha Ghosh. Use case 5: Simulated Land Use/Land Cover change: Richard Robertson. Use case 6 Spatial Dimensions of Poverty and Agricultural Productivity: Valerian Pede, Bert Lenaerts. Use case 7 Spatial Downscaling future food demand for local food policy making: Chun Song, Francis Yego, Athanasios Petsakos, Elisabetta Gotor. Use case 8 PlaSA as a spatial online supporting tool for policy: Carlos Gonzalez. Use case 9 Spatial assessment of Landscape complexity: implications for agricultural policies: Nicola Cenacchi Nicola Cenacchi, Athanasios Petsakos, Richard Robertson, Chun Song, Abhijeet Mishra. Use case 10 Assessing potential impact of avian influenza on poultry in West Africa: Liangzhi You, Xinsheng Diao. Use case 11 Where to strength water governance for preventing future water conflicts? Chun Song, Athanasios Petsakos, Elisabetta Gotor. The authors thank Elisabetta Gotor, James Thurlow, Keith Wiebe for their valuable inputs, and all participants at the Foresight and Prioritization seminar, whose comments greatly improved this report. The authors thank Regina Ngethe for her excellent research assistance on this report and Ilaria Urbani for her editorial support. 5 Abbreviations AGLINK-COSIMO Agricultural Linkage – Commodity Simulation Model AI Avian Influenza AoW Area of Work CAP Common Agricultural Policy CAPRI Common Agricultural Policy Regional Impact Model CATE Conditional Average Treatment Effects CF Causal Forests CGIAR Consultive Group on International Agricultural Research CIMMYT International Maize and Wheat Improvement Center CLI Canada Land Inventory COVID-19 Coronavirus Disease 2019 DEM Digital Elevation Model DiD Difference-in-Differences DMU Decision Making Unit DnD Differences-in-Discontinuities DSSAT Decision Support System for Agrotechnology Transfer EO Earth Observation EOSDIS Earth Observing System Data and Information System EU European Union FADN Farm Accountancy Data Network FAO Food and Agriculture Organization FAOSTAT Food and Agriculture Organization Corporate Statistical Database FLW Food, Land and Water FSSIM-Dev Farming System Simulator for Developing Countries GCM General Circulation Models GEOGLAM Group on Earth Observations Global Agricultural Monitoring Initiative GIS Geographic Information System GLOBIOM Global Biosphere Management Model HLPE High Level Panel of Experts IFM-CAP Individual Farm Model for Common Agricultural Policy Analysis IFPRI International Food Policy Research Institute IGSM Integrated Global System Model IIASA International Institute for Applied Systems Analysis IMPACT International Model for Policy Analysis of Commodity Trade InVEST Integrated Valuation of Ecosystem Services and Tradeoffs IRIO Interregional Input–Output Model IRRI International Rice Research Institute ISDC Independent Science for Development Council ISFM Integrated Soil Fertility Management ITE Individual Treatment Effect LAC Latin America and the Caribbean 6 LMIC Low- and Middle-Income Country LSMS-ISA Living Standards Measurement Study – Integrated Surveys on Agriculture MAGNET Modular Applied General Equilibrium Tool MAgPIE Model of Agricultural Production and its Impact on the Environment MapSPAM Spatial Production Allocation Model MENA Middle East and North Africa MIT Massachusetts Institute of Technology NASA National Aeronautics and Space Administration NCD Non-Communicable Disease NDVI Normalized Difference Vegetation Index NEG New Economic Geography OECD Organization for Economic Cooperation and Development PIK Potsdam Institute for Climate Impact Research PISA Performance, Innovation and Strategic Analysis for Impact PlaSA Plataforma de Sistemas Alimentarios (Food Systems Platform) RCMRD Regional Centre for Mapping of Resources for Development RCSSMRS Regional Centre for Services in Surveying, Mapping and Remote Sensing RCT Randomized Control Trial RD Regression Discontinuity SDiD Spatial Difference-in-Differences SDnD Spatial Differences-in-Discontinuities SMS Short Message Service SRD Spatial Regression Discontinuity SSU Shamba Shape Up WFP World Food Programme WPS Working Paper Series WUR Wageningen University & Research WWF World Wildlife Fund 7 Executive Summary The CGIAR Science Program on Policy Innovations (“Policy Program”) is committed to driving transformation across Food, Land, and Water (FLW) systems. Identifying viable policies and investment options through Foresight and Prioritization exercises (Area of Work 1) is key to reaching this goal. However, prioritizing interventions that are relevant to local needs and conditions, while addressing global drivers and megatrends that affect FLW systems across different scales remains a challenge. This report seeks to address this challenge. It demonstrates how spatial analytics, a fast- evolving field that sits at the intersection of economics, public policy, geography, and data science, can provide actionable policy insights. It also aims to equip policymakers and partners with advanced and accessible spatial analytical tools to design and implement tailored policies, investments and programs. The report starts by providing a unified framework that brings together diverse spatial analytics approaches to support policy. It reviews the evolution of spatial analytics, spanning geographic information systems (GIS), spatial economics, and economic models with spatially explicit inputs and outputs. It also introduces a taxonomy of building blocks to illustrate how different spatial tools and methods can address various policy questions. This report draws on 11 use cases from across CGIAR centers in which spatial analytics have been applied to inform policies across Africa, Asia, and Latin America. It demonstrates how spatial analytics can identify priority intervention areas and appropriate actions at the local level while accounting for global drivers. Key challenges in scaling spatial analytics for policy application are also identified, including data gaps, methodological complexities, and computational constraints. The report concludes by outlining future directions to fully leverage spatial analytics for policy support. This report aims to advance the integration of spatial analytics across disciplines and scales, enabling the translation of local spatial patterns into regional and global policy frameworks. The curated use cases show that spatial analytics is no longer a niche technical exercise, but an operational tool that facilitates FLW systems transformation towards desirable futures. By systematically linking spatial heterogeneity to multi-scale policy needs, spatial analytics can generate actionable and scalable insights for policy development and implementation. 8 1. Introduction 1.1 Background CGIAR Science Program on Policy Innovations (Policy Program) focuses on the transformation of Food, Land, Water (FLW) Systems as a whole. To support this objective, Foresight and Prioritization (Area of Work 1) charts potential future development pathways and identifies cost-effective policy options to drive transformation in FLW systems to enhance nutrition, livelihoods, gender equity, social inclusion, climate resilience, and environmental sustainability. However, identifying policy options and modelling the impacts of these policies are confounded by a great deal of heterogeneity and interactions among agents at local level: Farming activities occur across landscapes characterized by varying soil properties, microclimates, and water availability. Farmer’s choice of varieties is affected by location- specific risks. Farmers’ decisions on crop selection are influenced by neighboring farmers and local market demands. These phenomena lead to the question: How can policymakers incorporate local variations in determining where actions are (or will be) most needed, while still accounting for drivers at aggregate or even global scales? Spatial analytics provides one way to answer this question. As part of the broader suite of quantitative disciplines in the social sciences, spatial analytics stands at the crossroads of economics, geography, and data science. Since the 1960s, spatial analytics including GIS has been supporting policy formulation and delivery, in areas ranging from rural development to forest management. The integration of subnational level information into foresight modelling improves simulation of possible future, and the robustness of projected tradeoffs and synergies between alternative policy options at local geographical scales, while still accounting for the general impacts of mega drivers. Recently, advancements in causal inferences, machine learning, and big data from Earth Observation provide opportunities to model impact pathways at a granular scale. Building on these advances, in the past decades, the global south countries have witnessed a significant increase in investments in spatial- related research and development in the FLW system (Oyewole, 2017). More operational information services have become available in Sub Sahara African countries (e.g., early warning systems for food security, pest and disease monitoring, climate information services). Spatial analytics have also been used to guide infrastructure investment in Indonesia ($267 million) (The World Bank, 2013), Haiti ($20 million) (The World Bank, 2024)) and Horn of Africa (Dappe & Lebrand, 2021). Despite the increasing capabilities that spatial analytics offer for policy support, there remains no synthesis of how these analytical tools can be leveraged for policy purposes. What are the 9 common building blocks and shared concepts for spatial analytics that bridge the disciplinary gap between economists, geographers, and policy experts? What are the examples of different use cases? This report aims to address those questions. 1.2 Scope and Objectives of this synthesis report This synthesis report provides an overview of advanced spatial analytics for policy support in the global south, including recent use cases in One CGIAR. The objective is to synthesize the state-of-the-art spatial analytics for policy, identifying key building blocks, show recent applications in different use cases, discuss persistent challenges to incorporate more spatial granularity into foresight modelling, and suggest some future directions for researchers and stakeholders. This synthesis has two parts: Part 1 provides a review of the evolution of spatial analytics in supporting policy. It reviews three strands over this evolution: GIS, spatial/regional economics, and economic models with spatial inputs. A taxonomy of the different building blocks and their properties are provided. It can be seen that spatial analytics can support the answer of two policy needs: first, contextualization of local situation for policy. Second, informing targeting and prioritization within a country. Each need can be met by different spatial analytics that allow policymakers to move beyond mapping. Figure 1 illustrates those options. Figure 1: Different ways spatial analytics can support policy questions. 11 Part 2 showcases 11 use cases that apply spatial analytics for policy support at the global scale but focus on the global south, and different ways these cases contribute to foresight and prioritization. These use cases represent a set of CGIAR projects and innovative methodological approaches. These cases demonstrate how spatial analytics can be used in different ways for supporting policy and communicating with partners and citizens. The way forward section summarizes key insights arising from this synthesis, and what are the contexts that spatial analytics can be relevant and appropriate for policy. We also reflect on the skill set, background, and experience that can make up a successful scientist’s team that uses spatial analytics for supporting policy analysis. We comment on major challenges and ways forward in improving the spatial granularity in Foresight economic modeling. Amid increasing demand for more granular evidence to support policy making, this report is also relevant for a broad audience of researchers and practitioners from a variety of policy backgrounds interested in learning about the potential to meet the growing demand for using spatial analytics for policy support. We close with a discussion of the need for a systematic change in policy analysis to fully leverage the potential of spatial analytics for improving the granularity of activities within One CGIAR. 11 2. Spatial analytics for policy support: history and building blocks Credit: ©CIAT 12 2.1. Basic concepts and terminologies We start by setting common ground for basic terminologies. In this synthesis, “spatial analytics” refers to the geo-referenced data and spatial models (including statistical and economic models) used to support various policy purposes, including integrating different spatial datasets into analytical frameworks that can capture local heterogeneity (e.g., differences in productivity, costs, or constraints across locations) and agents interactions (e.g., spillovers, diffusion of innovations, social learning). We use the term “policy” to refer broadly to government strategies and interventions designed to influence FLW systems towards desirable outcomes. We focus on policy needs that require quantitative evidence to inform different options, tradeoffs, and assess their impact. Combining these concepts, we define spatial analytics for policy support as the use of geo-referenced data and spatial tools (including statistical and economic frameworks) to provide quantitative evidence for policy design and evaluation. It integrates heterogeneous datasets (e.g., georeferenced household surveys, social-economic, biophysical and climatic data layers, and remote sensing including drone or satellite-based Earth Observation) into analyses that capture spatial interactions across Decision Making Units (DMU) (such as spillovers, diffusion, and agglomeration) and spatial heterogeneity (such as differences in productivity, costs, and constraints across /locations). 2.2 The evolution of spatial analytics for policy support We introduce the evolution of spatial analytics for supporting policy, which started over three related strands that were initially developed as separate lines but recently became more integrated for policy purposes. Figure 2 shows the evolution of these strands. Figure 2: Evolution of spatial analytics for policy support (source: authors) 13 2.2.1 GIS for guiding public investment and service delivery. In 1963, the world's first Geographic Information System (GIS) was developed in Canada by the Federal Department of Forestry and Rural Development. It was first used to store, analyze, and manipulate data collected for the Canada Land Inventory (CLI). The CLI was used by The Department to decide where to invest in rural development and how to allocate land for different purposes including infrastructure agriculture, forestry, recreation, or wildlife conservation (Maguire, 1991). Later, GIS was increasingly used in supporting different public services delivery (Hannum et al., 2025). In the late 1960s to early 1970s, the Landsat program was started for environmental monitoring, resource management, and land policy making. Since then, Governments in the global north have been relying on Landsat imagery to track whether land-use practices align with policy goals or legal requirements. For instance, in US, Landsat data has helped legislators and State and local officials to determine land use policy and by planners to project transportation demand, to identify areas where future development pressure will be greatest, in order to estimate future infrastructure requirements, and to develop more effective plans for regional development (Anderson, 1977). Brazilian government has been using Landsat data to monitor deforestation and design policies for reforestation (Shimabukuro et al., 2022). More recently, Sentinel imagery data supports the European Union’s Common Agricultural Policy (CAP) by verifying farmers’ land use declarations and corresponding subsidy eligibility (Terres et al., 1995). In the Global south, the use of GIS for policy also started in a similar period. In 1975, the Regional Centre for Mapping of Resources for Development (RCMRD) was established in Nairobi, Kenya by Kenya, Uganda, Somalia, Tanzania, and Malawi and involved in developing spatial data infrastructure among its member states, including capacity building on natural resource mapping in Africa (Regional Centre for Mapping of Resources for Development (RCMRD), 2006). In recent years, GIS has continued to offer policy support. More advanced GIS involves image classification using machine learning and platforms such as Google Earth Engine. Models like the Integrated Valuation of Ecosystem Services and Tradeoffs (InVEST) can also be integrated with GIS to assess the impact of land management practices on soil-related ecosystem services. Researchers have also developed methods for mapping crop production. These products include Spatial Production Allocation Model (SPAM) for subnational crop distribution. Policymakers can use these insights to prioritize investments, target adaptation measures, and align agricultural policy with food security and sustainability goals. 14 2.2.2 Economic geography and spatial economics for regional development policy The second strand of spatial analytics for policy begins early with location theory, which attempts to explain the geographical placement of economic activities (like firms, industries, or cities) and implications on economic outcomes. It began in 1826 when Von Thünen model (Walker, 2022) explained agricultural land use patterns based on distance to markets. The DMU in Von Thünen’s model (1826) is profit- maximizing farmer. Landowners allocate crops across space by weighing market prices against transport costs and land rents. The model shows how rational decisions by individual producers, made under assumptions of an isolated market and uniform plain, yield the famous concentric rings of land use. Although stylized, this approach established the micro-foundations of land-use theory and continues to inform agricultural policy and land zoning analyses. Later, Alfred Weber (1909) extended the theory to industrial location (Fearon, 1909). The Von Tunen framework has been widely used to study early zoning and land use policy such as land sparing (Fontes & Palmer, 2018). Building on location theory, regional science, which oriented toward the study of regions “from a multidisciplinary and spatial-analytical approach” (Kourtit et al., 2021), was formally founded by Walter Isard in 1954 (Boyce, 2004). Later, the Interregional input–output (IRIO) models capturing production and trade flows across multiple geographies, treating the sector within a region as DMU. Gravity models of interaction conceptualize the region as the decision-making entity, sending flows of people, goods, or capital to other regions in proportion to size and distance. In these frameworks, the region itself acts as the unit of analysis and decision. In 1990s, New Economic Geography (NEG) (Krugman, 1998) introduced models of agglomeration and regional development (Krugman, 1991). Firms choose locations by balancing scale economies and transport costs, while workers and households decide where to live and work based on wages, amenities, and commuting costs (Fujita et al., 1999). By embedding these micro agents within general equilibrium frameworks, modern models restored individual decision-making to the forefront while capturing emergent spatial equilibrium. This approach provides a direct link between micro level behavior and macro level spatial patterns. Urban economics emerged in the mid-20th century, focusing on the internal spatial organization of cities, and the purpose is to inform infrastructure, transport planning, housing policy, etc. Other early efforts to spatialize economics focused on understanding the distribution of economic activities across space (e.g., Maurel & Sédillot, 1999), particularly how factors such as natural resource endowments and local policies shape regional development. With the growth of optimization and equilibrium economic models, spatial dimensions were gradually incorporated into forward looking modeling frameworks for land 15 use, agricultural production, and policy evaluation. Since the 2000s, advances in Earth Observation (EO), spatial econometrics, and machine learning have transformed spatial analytics into a key tool for policy. The strand of spatial econometrics emerged in the 1970s–1980s, when researchers recognized standard econometric models ignored spatial patterns in data. The formal foundations of spatial econometrics are usually attributed to Jean Paelinck, who first coined the term in the 1970s (Paelinck & Klaassen, 1979). His early work emphasized the need to adapt econometric techniques to account for spatial dependence and heterogeneity, later further developed by Luc Anselin. Spatial econometrics provides an empirically feasible way to estimate the strength of spatial heterogeneity and spatial dependence on economic and social outcomes. 2.2.3 Integrated Economic modelling with spatial explicit inputs The term “economic model” encompasses a wide range of quantitative tools, we focus on models that satisfy the following criteria: (1) they are used for forward-looking analysis to support decision making in agriculture (ex-ante or foresight models); and (2) they can represent farmers’ and/or consumers’ decisions explicitly or implicitly, individually, for a representative sample, or on aggregate (economy-wide), allowing them to simulate the supply and/or demand of agricultural commodities at subnational to global levels. According to these criteria, we discuss the evolution of incorporation of spatial considerations in two types of economic models: farm/household optimization models, and large-scale economic equilibrium models. The relevant literature and the capacity of these models to provide evidence-based support to policy makers has evolved significantly over the past few decades, driven by advances in computational capacity, data availability, and increasing policy demands for thematically and spatially nuanced insights. Spatially explicit policy analysis with these types of economic models aims to understand how (1) locally targeted policies or other interventions (e.g., investments and new technologies) could affect the agricultural economy and various elements of the food system at finer geographies or (2) how global/aggregate drivers affect each location differently in the agrifood system outcomes. Spatial components start to enter broader economic models in the late 1990s. This is because a key limitation of the economic models covered in this report is that they are not spatially explicit by design (both the model inputs and outputs). For example, models such as Individual Farm Model for Common Agricultural Policy Analysis (IFM-CAP) and Farming System Simulator for Developing Countries (FSSIM-Dev) (developed and used by the European Commission), simulate production and consumption decisions at the farm/household level, assuming a profit or utility maximization decision criterion under policy and resource endowment constraints (Kremmydas et al., 2022; Louhichi et al., 2020). These models are built with individual or representative farm data and allow a bottom-up 16 approach to the simulation of the agricultural sector. Farm/households are typically treated as independent Decision-Making Units, although there have been some recent attempts to simulate the interaction of DMUs in a non- spatially explicit fashion (Baldi et al., 2024). Farm/household models can theoretically use spatially explicit inputs and generate spatial explicit outputs if the location of farms/households is known. However, such capacity remains unexplored for policy analysis purposes. One of the main reasons is that the use of datasets containing a sufficiently large number of farms and their geographical coordinates, like the Farm Accountancy Data Network (FADN) of the European Union (EU)1 and the Living Standards Measurement Study – Integrated Surveys on Agriculture (LSMS- ISA), is often subject to legal constraints for information protection and privacy reasons. Even though strategies for obfuscating farm household coordinates appear to have little impact on econometric estimates when such data is used for ex-post assessment purposes (Michler et al., 2022), we are not aware of any forward-looking studies using spatially explicit data for policy analysis with farm-level optimization models. On the other hand, large-scale economic models such as Common Agricultural Policy Regional Impact Analysis (CAPRI of the European Commission), International Model for Policy 1https://agridata.ec.europa.eu/extensions/FarmEconomyFocus/FA DNDatabase.html 2 Equilibrium models focus on market closure and price formation and not on individual economic agents. As such, they do not Analysis of Commodity Trade (IMPACT from International Food Policy Research Institute, IFPRI), Modular Applied General Equilibrium Tool (MAGNET by Wageningen University & Research WUR), Model of Agricultural Production and its Impact on the Environment (MAgPIE from Potsdam Institute for Climate Impact Research, PIK), Global Biosphere Management Model (GLOBIOM from International Institute for Applied Systems Analysis, IIASA), and Agricultural Linkage (AGLINK from Organization for Economic Cooperation and Development; Food and Agriculture Organization, OECD-FAO) rely on aggregate/macro-level data for some of the model inputs that is often publicly available (e.g., FAOSTAT) and thus are not subject to similar data use constraints as their farm-level counterparts. Large-scale economic models operate at various levels of aggregation and are primarily concerned with commodity prices and market interactions. They are designed to capture the systemic interactions of interventions and global megatrends, such as population growth, climate change, and policy shifts, and their influence on food, land, and water systems. As such, their main objective is to simulate market behavior and macro-level outcomes, not to offer a spatially fine-grained outlook on micro-level production decisions. In contrast to farm/household models, a DMU in large-scale economic equilibrium models is implicit2 and typically concerns the lowest explicitly consider a decision maker as do farm/household models. However, the main assumption behind the neoclassical concept of economic equilibrium is that it represents an aggregation of https://agridata.ec.europa.eu/extensions/FarmEconomyFocus/FADNDatabase.html https://agridata.ec.europa.eu/extensions/FarmEconomyFocus/FADNDatabase.html 17 spatial scale considered by the model during its solution. Thus, a DMU can be a world region, a country, and sometimes further disaggregated into agro-ecological zones or administrative districts within a country. When the spatial scale for this implicit DMU does not correspond to the scale on which prices are assumed to be formed, the aggregation of production and consumption outcomes across DMUs enables the model to ensure consistency in supply- demand balance and market clearing conditions. Some of the previously mentioned large scale models follow this approach of performing analyses at finer resolutions and aggregating outputs to the scale of market equilibrium. For instance, MAgPIE3 and CAPRI (Britz & Witzke, 2014) include modules that simulate land use decisions at a high spatial resolution (subnational) before feeding results into a broader market framework.4 This hybrid approach allows for more spatially nuanced representation of economic decision-making at a subnational level yet remains very coarse in resolution. DMUs are treated as largely spatially independent in the model structure. This means that large-scale economic models typically do not include endogenous spatial interaction, whereby, it is very plausible that changes in one economic agents’ optimal consumption and production choices within the region or country of interest. 3 https://www.pik- potsdam.de/en/institute/departments/activities/land-use- modelling/magpie 4 The spatial scale in CAPRI is NUTS2 (Nomenclature of Territorial Units for Statistics 2), which is roughly equivalent to ADMIN1, in other words, the first administrative level within a country. The DMU influence others through direct spatial processes.5 Instead, DMU interactions in many economic models arise through market and international trade (though bilateral trade is often not modelled explicitly). For example, a productivity shock in one DMU may affect world prices and, through those prices, influence land use or production decisions elsewhere. This system-wide feedback is a core strength of economic equilibrium models but differs from the localized feedback mechanisms emphasized in other spatially explicit approaches. The above discussion reveals that, although economic models are not inherently spatially explicit, they can broadly be considered as such, in the sense that their DMUs refer to a specific spatial scale. However, it is also clear that the term “spatially explicit” does not lend itself to the same interpretation as in other spatial quantitative modeling approaches such as spatial econometrics and agent-based modeling. The main reason is that the model solution depends more on the underlying assumptions and the overall economic logic in representing commodity markets and less on the spatial scale of the DMUs6. For farm/household models, spatial information is not even required as model input. Second, when DMUs are directly associated with specific spatial units (as in algorithm solves for every NUTS2 unit and simultaneously aggregates across NUTS2 to simulate the market equilibrium at national level. 5 Such direct interactions may still exist in biophysical components of an economic model, like in the IMPACT water model. 6 The spatial scale of the DMUs can, however seriously impact computation time and overall computational feasibility. We discuss this point in the next section. 18 large-scale equilibrium models), economic decision-making is represented at high levels of aggregation. On the contrary, traditional spatial analytic tools are usually applied at much finer spatial resolutions. Finally, economic models often do not simulate the interaction between different spatial units (e.g., bilateral trade), which is typically the focus of traditional spatial analytic tools. The implication for all economic models mentioned previously is that spatial heterogeneity is captured only to the extent that it is assumed to be implicit in the constraint specifications at the DMU level. For instance, crop (or land) productivity, climate shocks, or technological adoption may vary across geographical scales, and these differences are reflected in model parameters or inputs. This point reflects a fundamental difference in modeling philosophies and objectives: economic models are not statistical constructs aiming to estimate empirical relationships among spatial units. Rather, they are optimization/market clearing frameworks that seek to represent decision- making under predefined constraints. These decisions may be made by producers and/or consumers and are modeled as either direct optimization of objective functions (e.g., cost minimization, utility maximization) or as part of an equilibrium condition where supply meets demand (implicit optimization). 2.3 Building blocks of spatial analytics for policy There are several key building blocks of spatial analytics for policy. The choice of those building blocks depends on both computational and data availability and the specific needs of policymakers. Identifying the intersection of these factors ensures spatial analytics are feasible, relevant, and capable of delivering actionable, context-specific insights for effective policy decisions. Spatial Data: there are several different spatial data types. Areal (Polygon) Data, for example administrative boundaries, villages, and districts. Point Data: discrete spatial events or features. Geostatistical Data are derived from samples across a continuous surface (e.g., precipitation, temperature). Decision making unit (DMU): Decision- Making Unit means the level that economic decisions are made at. This may be the level of farm, a policy beneficiary, a community, an agro-ecological zone or administrative district within a country, or a world region. Spatial unit: The spatial unit refers to the resolution at which data are collected or analyzed (e.g., point, village, district, region), while scope pertains to the geographical extent of analysis, ranging from communities to countries or even global systems. These choices are interdependent: high-resolution data (e.g., plot-level) may not be feasible or needed for global-scale studies due to data availability or processing constraints. 19 Raw spatial data at high resolution may be aggregated to lower resolution for different purposes (Todd & Dar, 1977) pointed out that in the early example of using Landsat for land use policy, many Federal and State agencies will only be concerned with coarse tabular aggregations of change data for entire urban regions. A large State agency, for example, may be interested in yearly indications of urban growth acreage as input to the formation of a statewide land use policy. Local agencies, however, may use maps and tabular data for more detailed analysis. Multicounty, county, or municipal maps of land-use and land-cover change might be prepared with tabular aggregations at smaller units, as input to planning models and decision making. Spatial scope: refers to the spatial extent of which the analytics is applied to. Interaction: Spatial interaction refers to how different spatial units or DMU influence one another in affecting the outcomes of policy interest. There are two types of spatial interactions. One happens at the same level, for example the interaction among decision makers in nearby locations, the movement of goods and labor across locations; the improved road infrastructure in one region may increase market access in neighboring areas; conflicts may spread out and influence more communities. Diffusion of innovations(Comin et al., 2012). One is through aggregate mechanisms. For example, through global trade and market, and the diffused influence of one’s decision of using water or crop through water model (bilateral). Another example is social norms and aggregate behavior: These shape adoption of technology or health behavior across regions. Modelling the interaction of agents may not inform directly policy options but increase the accuracy of policy influence. Neighbors: The concept of “neighbors” can be defined in several ways: Geographic neighbors that are based on contiguity (e.g., "rook" or "queen" adjacency on a map grid) or distance. Economic or social neighbor based on trade flows, shared infrastructure, or social networks. The neighbors can be endogenously formed or exogenously formed. Choosing an appropriate neighbor structure is key to capture the spatial relation of interest. Spatial Patterns: Spatial analytics identifies patterns such as agglomeration, dispersion, equilibrium, or dispersion or disequilibrium. Patterns arise due to many reasons: behavioral diffusion: Local imitation or influence (e.g., farmers adopting nearby innovations). Attributional similarity: Shared characteristics due to common geography or socioeconomics. Understanding these patterns is essential for designing geographically targeted policies and evaluating their impact. Tools like Global and Local Moran’s detect autocorrelation and help in identifying clusters or hotspots requiring policy attention. 20 3. How can spatial analytics support policy? Credit: ©CIAT 22 We notice five major ways that spatial analytics has been used to support policy. 3.1.Contextualizing key spatial covariates for policy making Spatial analytics provides a tool for contextualizing data points or a local policy challenge. For example, when deciding where to target a certain intervention, overlaying various spatial layers for the targeted region can give policy makers a context about what the surrounding look like, which households might be difficult to reach, and to what extent are the target areas representative of the entire country across diverse spatial contexts. Some spatial layers such as population density and age-gender distribution from WorldPop can reveal communities with large young or elderly populations that might be disproportionately impacted by shocks related to nutrition and food security, thus informing where interventions or aid might be needed. Another example is land use policy aimed at improving soil health. Spatial analytics has played a fundamental role in supporting land use policy and soil management. They facilitate the creation of detailed soil maps, integrating data from soil surveys, laboratory analyses, sensor readings (e.g., for moisture, pH, organic matter, nutrients), and terrain models derived from Digital Elevation Models (DEMs). For example, GIS is used to assess and map soil degradation risks, such as erosion potential (often modeled using factors like slope, rainfall intensity, soil type, and land cover) and nutrient depletion or imbalances. This spatial understanding allows for the planning and implementation of targeted soil conservation measures (e.g., terracing, contour farming, cover cropping) and site-specific soil fertility management strategies, including integrated soil fertility management (ISFM) approaches. 3.2. Contextualizing local influence of mega trends Food, Land, and Water (FLW) systems are at the intersection of the mega challenges, including climate change, population growth, poverty, food and nutrition insecurity, increasing inequality, political instability, and environmental degradation (ISDC, 2023). These mega-trends will have profound policy implications and address them through policy with the necessary evidence to inform decision making requires systems thinking and a strong, coordinated, and multidisciplinary response from CGIAR and its partners. Spatial analytics plays a critical role in contextualizing the local influence of global megatrends by downscaling or disaggregating information traditionally modeled at large scales into finer spatial resolutions. Megatrends such as climate change, population growth, urbanization, and shifts in global economic structures are inherently aggregate and systemic, yet their manifestations and impacts vary significantly across local contexts. Spatial analytics such as SPAM (Spatial Production Allocation Model) (Use case 2 in this report), future land use change (Use case 5) and land complexity (Use case 9), and food demand at local level (Use case 7), provide more granular 23 results given projections or results from national statistics or global model, provide examples of how to account for these interdependencies across scales, and enabling policy makers to identify hotspots that is anticipated to experience drastic changes such that current policy may not be adequate, or where vulnerability may appear. This helps assess the uneven impacts of megatrends, and tailor policy responses accordingly. Without such localized perspectives, global models risk obscuring the heterogeneity of impacts and leading to policies that are ineffective or inequitable at local level. In this sense, spatial analytics thus provides a bridge between mega trends and local heterogeneity. 3.3. Assessing (local) effectiveness of interventions Spatial analytics has become an increasingly useful tool in impact evaluation and impact assessment, to evaluate the heterogeneous and autocorrelated causal impacts of an intervention across space. These applications emphasize that credible policy evaluation requires both recognizing localized variation and rigorously testing generalizability across diverse spatial contexts. We briefly review two strands of literature on this aspect: impact evaluation that incorporates spatial heterogeneity; and quasi- experimental design building on a spatial explicit setting. First, spatial analytics has been increasingly used for understanding spatially differential treatment effects. This is critical because policy interventions often do not have uniform impacts across space. Geographic, socioeconomic, and institutional factors can shape how treatments work in different locations, leading to systematic heterogeneity in outcomes of policy interest. Recognizing and quantifying these spatial variations matters for policy, since it allows interventions to be tailored more precisely, resources to be allocated more efficiently, and unintended spillovers to be managed effectively. This strand of methods traditionally relies on regression that specifies the form of heterogeneity parametrically, but increasingly on machine learning and non- parametric methods. For example, Causal Forests (CFs) represent an advancement in estimating Individual Treatment Effects (ITEs). First introduced by Athey & Imbens (2016), this approach has then been improved by Athey et al (2019) and Chernozhukov et al. (2018). In the past, CFs have mostly been applied to explore heterogeneity across demographic, socioeconomic, or behavioral factors. Recent work has started to use CFs in spatially explicit contexts. Kakimoto et al., (2022) applied CFs to estimate site-specific optimal nitrogen rates in precision agriculture, drawing on spatially detailed field trial data and accounting for spatially autocorrelated soil properties. Similarly, Credit & Lehnert (2024) showed that CFs can be adapted to include spatial lags of both dependent and independent variables, such as in a spatial Durbin specification, enabling the model to better capture spatial dependencies and spillover effects in treatment impacts. However, spatial methods require high- resolution data, involve more complex model 24 specifications, and may be sensitive to assumptions. Careful implementation and validation remain essential. In this report, use case 4 brings one recent example of how careful implementation is carried out in evaluating heterogeneous climate information services across space. Second, spatial econometric extensions of causal inference methods, such as spatial difference-in-differences (SDiD), spatial regression discontinuity (SRD), and spatial differences-in-discontinuities (SDnD) adapt traditional identification strategies to account for spatial dependence and spillover effects. Luc Anselin and coauthors (Kolak & Anselin, 2020) reviewed how spatial econometric technics are applied in causal inference. This is important because in spatially explicit contexts, treatment effects may diffuse across geographic boundaries, and outcomes can be correlated due to spatial proximity. Compared to traditional settings, these spatial adaptations improve external validity in geographically structured interventions, but they also demand stronger modeling choices to disentangle direct, indirect, and heterogeneous effects. Thus, spatial causal methods enhance identification where geography itself shapes exposure and outcomes. In this report, use case 11 brings one recent application of this method to evaluate where climate risks might escalate water conflicts in Africa. 3.4. Inform locally tailored actions and prioritization Spatial analytics can help prioritize and find suitable sites for future intervention and how validated interventions can be scaled to broader locations and adapt to new contexts, scaling out small scale interventions addresses the challenge of external validity, ensuring results extend beyond single trial sites. This approach helps tailor programs to local conditions, uncover contextual factors that influence outcomes, and build political credibility by providing evidence from multiple regions rather than a single site. Ultimately, scaling out ensures that policy recommendations are robust, context-sensitive, and better suited for national or regional implementation. For example, Randomized Controlled Trial (RCT) has been a “gold standard” tool to understand the causal impact of an intervention (Duflo et al., 2008) with high internal validity. Yet, if an RCT is to inform policy, it is critical to establish external validity (Rothwell, 2005) (Cartwright, 2010). This persistent gap risks producing results that, although internally sound, may have limited applicability in typical policy contexts (Peters et al. 2018, Banerjee et al., 2017). To this end, spatial analytics can examine the external validity of RCTs when participants are randomly assigned (Savoca et al., 2017). GIS method offers a visual, location- based assessment that only requires address data, making it particularly useful for community-based RCTs. Savoca et al. (2017) conclude that, while true representativeness can only be guaranteed through probabilistic 25 sampling, combining graphical spatial analysis with simple statistical models provides a practical and more nuanced way to evaluate external validity beyond standard hypothesis- testing methods. 3.5. Easier visualizing & disseminating Spatial analytics not only support identifying policy options or assessing heterogeneous impacts but also facilitates more effective visualization and dissemination of complex information to diverse stakeholders. Through maps and online mapping platforms, vast amounts of data can be communicated in a more clear, accessible, and transparent manner, and improving the decision-making process. Visualization tools allow policymakers, researchers, and practitioners to identify spatial patterns, overlaps, and disparities that are often less evident in textual or tabular data. For example, FAO Hand-in-Hand 7tool provides interactive maps for monitoring agricultural transformation. The Alliance of Bioversity and CIAT PLaSA platform supports evidence-based decisions for agrifood systems. Similarly, the World Food Programme (WFP) has developed visualization tools to assess food security and vulnerability; NASA’s Earth Observing System Data and Information System (EOSDIS) for climate and environmental monitoring; and the Global Forest Watch platform by the World Resources Institute for real-time forest change analysis. While maps carry the advantage of condensing complex variables into intuitive visuals, they require cautious interpretation to avoid oversimplification or misrepresentation of underlying data. In this report, use case 8 shows how an online platform can demonstrate a complex agrifood system to support policy making and engage various stakeholders in the food system. 7 www.fao.org/hand-in-hand/en/ 27 4. Use cases of applying spatial analytics for policy support Credit: ©2016CIAT/NeilPalmer 27 In this section, we provide 11 use cases across CGIAR centers in Policy Innovation Program AoW1, to demonstrate different spatial analytics and policy relevance. Three criteria were considered when selecting those use cases. First, the use case must reflect either spatial heterogeneity or spatial interaction among agents, Second, the use case must have clear policy relevance of either global local outlook, prioritization, and rapid response. Finally, the use case must integrate both biophysical and social economic component in the case. With those criteria, the selected use cases reflect a wide range of functions of how spatial analytics can contextualize spatial covariates, inform locally tailored action, or better visualization of complex FLW system information, in ex ante impact assessment, ex post impact evaluation, and foresight. Figure 3 shows how this use case contributes to various impact areas and Policy Innovation Area of Work 1 efforts. These cases demonstrate how spatial analytics add value by capturing heterogeneity (e.g., localized productivity–poverty relationships, climate impacts on specific crops), enabling interaction analysis (e.g., between climate, policy, and biodiversity), and improving communication through compelling, multi-scale visualizations. These methods help policymakers design targeted, cost-effective interventions, enhance resilience, and respond rapidly to emerging crises—whether conserving habitats, optimizing food supply chains, mitigating water conflicts, or managing animal disease outbreaks. Figure 3: Mapping each use case on impact areas and AOW1 cluster. 28 Use Case 1: Fertilizer Investment Risk Explorer (FIRE) Meta Data for Use Case 1 AOW1 Areas Global local outlook Prioritization and Evaluation Rapid response Impact Areas Climate adaptation & mitigation Environmental health & biodiversity Gender equality, youth & social inclusion Nutrition, health & food security Poverty reduction, livelihood and jobs Spatial Scope Global Regional Country Within country Spatial unit of the OUPTUT of the use case Pixel: 5km * 5km Community Households Other Time scope of the OUPTUT of use case analytics Historic period [2000, 2005, 2010, 2017 (South of the Sahara), 2020] Both historic + future Future 2016CIAT/GeorginaSmith 29 By Bisrat Gebrekidan, Jordan Chamberlin, Maxwell Mkondiwa | CIMMYT Introduction and policy relevance Closing Africa’s cereal yield gaps is a key factor in securing the region's future food supply, yet farm-level productivity in sub-Saharan Africa (SSA) remains well below the global average (van Ittersum et al., 2025). Poor soil fertility, characterized by acidity, depleted organic matter and unfavorable texture, is a persistent issue that limits the effectiveness of nutrient applications (Kihara et al., 2020; Sanchez, 2002; Vanlauwe et al., 2015). Rainfall variability further complicates matters: multi- decadal analyses reveal significant inter-annual fluctuations across SSA, and recent El Niño events have exacerbated production risks (Alahacoon et al., 2021; Hansen et al., 2019). Market conditions also play a critical role. Farmers often face high fertilizer costs, fragmented markets and volatile maize prices (Minot, 2014; Palmas & Chamberlin, 2020; Vorsah et al., 2025), all of which reduce incentives to adopt (Jayne & Rashid, 2013; Liverpool‐Tasie, 2017; Oyinbo et al., 2022). Policy efforts have largely relied on fertilizer subsidies to stimulate adoption (Jayne et al., 2018). While these programs have increased usage in some contexts, systematic reviews show mixed effects on yield and welfare, raising concerns about fiscal cost, targeting and displacement (Jayne et al., 2018; Mason & Smale, 2013). A central limitation of such programs is their neglect of spatial heterogeneity: profitability varies dramatically across locations, so national averages obscure where fertilizer is truly beneficial. While farmers in favorable zones may achieve high returns, those in marginal areas may incur losses, which discourages sustained adoption. This use-case addresses these issues by developing a spatially explicit, risk-aware framework to estimate fertilizer profitability across SSA. By integrating soil, climate, and market data into a machine learning model, we generate pixel-level maps of yield response, returns, and profits. Crucially, we account for uncertainty through stochastic simulations of rainfall and maize price volatility, moving beyond point estimates to probabilistic assessments. The outputs not only highlight where fertilizer is profitable, but also where risks are high and where complementary investments are needed. This provides evidence for subsidies targeting resource allocation and farmer advisory services. While demonstrated here for maize, the framework is readily adaptable to other crops by substituting relevant yield, soil, and market data. Method We propose a spatially explicit modeling framework to estimate maize yield responses, returns, and profitability under fertilizer use across sub-Saharan Africa. The framework combines machine learning with harmonized biophysical, climatic, and market data to capture the diversity of production environments. Yield responses were estimated using a non-parametric modeling approach (random forest and causal Forest) capable of 30 handling nonlinear relationships and interactions, with fertilizer application rates, soil characteristics, growing season total rainfall and variability, and market access included as key predictors. Profitability was assessed by comparing fertilized and unfertilized scenarios, translating yield differences into financial returns through spatially and temporally varying maize prices, and accounting for fertilizer costs derived from localized nitrogen price surfaces. To reflect the uncertainty farmers face, we designed a simulation procedure in which rainfall and market conditions were repeatedly sampled from historical distributions, producing multiple realizations of yield, returns, and profit outcomes. This allowed us to generate not only expected values but also measures of downside risk and the probability of achieving positive net returns. Results and key maps The framework generates spatially explicit estimates of fertilizer profitability and associated risks. It allows us to look beyond average responses and pinpoint patterns of opportunity and risk across different regions. The following maps illustrate continental-level patterns and highlight heterogeneity within countries. Figure 4: Spatial and regional patterns of fertilizer profitability and risk 31 Note: Panel A maps the model-predicted intervention areas across Sub-Saharan Africa, classifying each pixel into categories of expected returns and variability. Panel B zooms into Ethiopia to illustrate how the same framework can capture within-country heterogeneity. Panel C summarizes the model outputs at the administrative-region level, showing the share of maize area in each profitability–risk category for the top 10 producing regions. These show how modeling framework can be used not only to visualize local patterns but also to generate policy- relevant summaries that inform targeting of fertilizer subsidies, advisory services, and investment priorities. Policy recommendations The framework provides actionable insights for a wide range of stakeholders. For policymakers, it provides an evidence base to design fertilizer subsidy and extension programs that move beyond national averages. By identifying locations where fertilizer is profitable and where risks are high, governments can better target subsidies, avoid inefficient blanket recommendations, and prioritize complementary investments such as lime, credit schemes, or infrastructure in marginal areas. NGOs and development partners can utilize the outputs to strategically allocate resources, focusing interventions in areas with the highest potential for productivity, food security, and resilience gains. For private sector investors, including input suppliers and financial institutions, the results identify promising markets where demand is likely to be sustained and risks manageable, thereby supporting investment in supply chains, insurance products, and credit services. More so, this probabilistic framework can be embedded into existing decision-support tools, enriching them with risk-adjusted, spatially explicit layers offering localized, risk-adjusted guidance. This guidance can assist in mitigating financial losses and promoting greater confidence in the adoption of input. 33 Use case 2: Spatial Production Allocation Model (SPAM) Meta Data for Use Case 2 AOW1 Areas Global local outlook Prioritization and Evaluation Rapid response Impact Areas Climate adaptation & mitigation Environmental health & biodiversity Gender equality, youth & social inclusion Nutrition, health & food security Poverty reduction, livelihood and jobs Spatial Scope Global Regional Country Within country Spatial unit of the OUPTUT of the use case Pixel: 10km * 10km Community Households Other Time scope of the OUPTUT of use case analytics Historic period [2000, 2005, 2010, 2017 (South of the Sahara), 2020] Both historic + future Future Credit: ©2016CIAT/GeorginaSmith ©2017CIAT/GeorginaSmith 33 By Shuang Zhou, Zhe Guo, Liangzhi You | International Food Policy Research Institute (IFPRI) Introduction and policy relevance Feeding a growing global population while preserving forests, wetlands, and other natural ecosystems requires the development of more efficient and sustainable cropping systems (United Nations Department of Economic and Social Affairs, 2024). The expansion of farmland often comes at the expense of these critical ecosystems (Potapov et al., 2021), making it essential to improve agricultural productivity without further degrading natural resources. A crucial step in addressing this challenge is understanding the spatial distribution of crop types, their productivity, and the farming systems in which they are managed. This type of information allows policymakers, researchers, and development partners to better target interventions, enhance food security, and design effective climate adaptation strategies while conserving the ecosystem and biodiversity. Global agricultural production data are typically reported at the national or subnational administrative level (such as states or regions). However, this level of statistical data fails to capture the diversity and spatial heterogeneity of agricultural production, lacking spatial explicitness. The Spatial Production Allocation Model (SPAM) uses an entropy-based approach (You & Wood, 2006) to downscale crop production data, generating global maps of agricultural area, yield, and production at a spatial resolution of 5 arc minutes (~9km at the equator), covering more than 40 crops and crop groups. The model offers valuable insights into where different crops are grown, how much they produce, and under what production conditions—forming a strong foundation for evidence-based agricultural planning and sustainable land management. Method The SPAM model disaggregates national and subnational crop production statistics to a fine spatial resolution by integrating a wide range of geospatial datasets, including satellite-based land cover, crop suitability, irrigation maps, population density, and market accessibility. This information is compiled and integrated to generate “prior” estimates of the spatial distribution of individual crops. Priors are then submitted to an optimization model that uses cross-entropy principles and area and production accounting constraints to simultaneously allocate crops into the individual “pixels” of a GIS database (You & Wood, 2006). The result for each pixel (notionally of any size, but typically from 1 to 100 square km) is the area and production of each crop produced, split by the shares grown under irrigated and rainfed conditions (each with distinct yield levels). At the core of SPAM is a cross-entropy optimization method, which is used to allocate crop area in a way that minimizes the divergence between prior estimates and the final allocation of area shares across grid cells, crop types, and production systems. The optimization process is subject to a range of 34 constraints, such as land availability, irrigation area, and official crop statistics, ensuring that the results are both consistent with known data and yet spatially plausible (Yu et al., 2012). Results and key maps Figure 5 showcases some of the key results from SPAM2020, using maize as an illustrative example. Users can find more information for the SPAM model and the previous versions of SPAM datasets via the dedicated SPAM website (https://mapspam.info/) Policy recommendations The SPAM (Spatial Production Allocation Model) dataset series has evolved significantly since its first release in 2000, expanding from just over 20 crops to 46 in the latest SPAM2020 version. This detailed mapping of global agricultural production enables analysts and policymakers to design more targeted and effective agricultural and rural development strategies. By providing spatially disaggregated estimates of crop distribution by type and production system, SPAM supports policies that enhance food security, promote economic development, and reduce environmental impacts. For example, SPAM data can guide the allocation of input subsidies, irrigation investments, and infrastructure improvements to areas with high agricultural potential or vulnerability. It also helps identify regions suitable for adopting sustainable practices such as climate-smart agriculture or precision farming. The SPAM model and its outputs have become a critical input to many studies, models and initiatives within and beyond the CGIAR. The critical role of SPAM, as spatial data input and baseline data, is to link economic models and biophysical models (e.g., crop growth model, water model), and explore the sub-national spatial heterogeneity within a country or region. Some researchers within Policy Science Program and Climate Action Science Program use SPAM almost weekly. SPAM data are frequently downloaded and widely used by researchers and analysts from international originations, academia, governments agencies all over the world. For example, African Growth and Development Policy (AGRODEP) Modeling Consortium features SPAM on its online library and uses SPAM data as inputs for its own modeling work (http://www.agrodep.org/fr/node/1794). The Group on Earth Observations Global Agricultural Monitoring Initiative which uses SPAM to monitor food security and potential crises(www.geoclam.org). The World Bank uses SPAM inputs for its study on Africa’s Infrastructure. 35 Figure 5: Harvested area maps for maize in rainfed (a), irrigated (b), and all (c) farming systems, SPAM 2020 Compare different global crop mapping products. There are, broadly speaking, two groups of models to map cropping systems globally: one is statistical modelling which mainly relies on official statistics or survey data, the other is remote sensing-based modelling which relies on in-situ field data for both crop area and yield estimation. SPAM belongs to the first group. Others in this group include: (1) MIRCA2000: Provides monthly irrigated and rainfed crop areas around the year 2000(Siebert et al 2005). It’s widely used for climate impact assessments and water use studies. (2) GAEZ (Global Agro- Ecological Zones): Developed by FAO and IIASA, GAEZ combines climate, soil, and terrain data to assess crop suitability and potential yields (Fischer et al., 2021). (3) M3 (Harvested Area and Yield for 175 Crops): A statistical model that harmonizes crop data from various sources to produce global maps of harvested area and yield around the year 2000(Monfreda et al., 2008). Lately, CROPGRIDS updated M3 product to the year of 2020 (Tang et al., 2024). Remote sensing-based products include: (1) WorldCereal (ESA-funded. https://esa- worldcereal.org/en) – A dynamic, open-source system for seasonal crop and irrigation mapping at 10-meter resolution. It produces: Temporary crop extent maps, Seasonal maize and cereal maps, Irrigation maps, active cropland maps. WorldCereal allows users to upload their own field data to train localized crop classification models, enhancing accuracy for specific regions. (2) Crop Monitor (GEOGLAM. https://www.cropmonitor.org/) – Provides real- time monitoring of crop conditions globally using satellite data and expert assessments. It supports early warning systems for food security. While this approach is new with great promises, its application on a global scale is still quite limited. The availability of ground training and validation data is a major constraint. 36 Use case 3: Spatial mapping of crop vulnerability for targeting investment. Meta Data for Use Case 3 AOW1 Areas Global local outlook Prioritization and Evaluation Rapid response Impact Areas Climate adaptation & mitigation Environmental health & biodiversity Gender equality, youth & social inclusion Nutrition, health & food security Poverty reduction, livelihood and jobs Spatial Scope Global Regional Country Within country Spatial unit of the OUPTUT of the use case Pixel: 50 km * 50 km Community Households Other Time scope of the OUPTUT of use case analytics Historic period Both historic + future Future: 2050 Credit: ©2016CIAT/NeilPalmer 37 By Tim Thomas | IFPRI Introduction and policy relevance Crop modelling has been used for around 2 decades in the CGIAR to understand the impact of climate change on crop yields (Rosegrant et al., 2025). The work has traditionally focused on changes in average yields over time at each pixel -- typically a half-degree (roughly 54 km) square. However, while the results are often mapped only for areas where the crop is grown, the focus has not been on whether the crop is of high importance to that specific location. If it is of high importance to farmers in that location, projected losses from climate change can impact on the food security and economic well- being of households which depend on local agricultural production. We also know that farmers and policy makers care more about what will happen in “worst cases” (if the most adverse climate turns out to be the actual future climate) and in “extreme events” (in years in which climate is particularly adverse toward producing the crop – see (Murgatroyd et al., 2025; Thomas, Robertson, et al., 2022; Thomas, Schlosser, et al., 2022). In both, the well-being of households’ dependent on agriculture will be critically affected by climate in that particular year. If vulnerable areas can be identified in advance, spatially targeted policies and investments to reduce risk or adapt to climate change can be developed. Such knowledge allows them to be implemented at a lower cost than a national or global intervention. Furthermore, anticipating the problem can allow sufficient time to develop and implement solutions.  Method and results The methodology involves a multi-step process to identify agricultural “hotspots” vulnerable to climate change. First, we use the most recent version of SPAM to identify the major crop(s) in each location of the world. Figure 6 shows the SPAM data for rainfed maize in Sub- Saharan Africa. To ensure that a crop is truly “major”, we identify locations for which the crop has at least 25% of the cropland area in the grid square and that cropland represents at least 5% of the total land there. Second, we use IFPRI’s DSSAT model results for the world that feed into IMPACT (half- degree pixel or 54 kilometers at the equator) to determine the median impact of climate change across 5 climate models given an emissions scenario (Thomas, 2024; Thomas & Robertson, 2024) and identify locations that are above a specific threshold (e.g., 20%) for median losses from climate change. The median climate model results are shown in Figure 2 for rainfed maize. We note significant yield losses in West Africa, but some yield increases in the highlands of Ethiopia, Kenya, and a few other locations. Third, we interact the identified key crop locations with locations that have significant reductions in yield to create maps of “hotspots” for median climate losses across 5 climate models available for the world. The example for rainfed maize in Sub-Saharan African is shown in Figure 8. The reader can examine Figure 9 to 38 see that Figure 8 was created using the rules just described. Fourth, it is possible that the worst-case climate model at each location will be like the actual climate the location experiences in the future. To account for this type of possibility, we interact the identified key crop locations with the worst case-scenario yield change to find “hotspots” for the maximum climate losses across 5 climate models in which the losses are projected to be greater than 20%. See Figure 9 for the example of rainfed maize in SSA, noting the increase in the number of hotspots compared to Figure 8 which was based only on the median yield results from the 5 climate models. There is another way to think about hotspots, and that is by considering that climate change may alter the frequency of extreme events. Using a large ensemble of climate models, Thomas et al. (2022a, 2022b) explores the change in frequency of low-rainfall events and low-yield events under climate change. Figure 10 shows the change in frequency of 1-in-20 low-yield events for rainfed maize. In the map, a value of 20 indicates that it is a 1-in-20-year event in the 2060s based on the 1-in-20-year yield measure of a low-medium emissions scenario in the 2020s. What we see is that under the lowest emissions scenario, there is little change in the frequency of low-yield events in the 2060s. However, as the level of emissions rises, we see an increase in frequency of low- yield events, with most locations having those events every 8 years or less under the highest emissions scenario. Figure 6: Map showing the percentage of each pixel that is rainfed maize in 2020 (Source: SPAM2020v2r0.) 39 Figure 7: Projected yield change for rainfed maize for median climate model, RCP 7.0 (medium-high emissions scenario, 2005- 2050) Source: Authors using DSSAT. Notes: At least 5% cropland in pixel. Maize represents at least 25% of cropland. Figure 8: Hotspots for rainfed maize in Africa for median climate model, RCP 7.0 (medium-high emissions scenario). Source: Authors using DSSAT. Notes: At least 5% cropland in pixel. Maize represents at least 25% of cropland. Figure 9: Hotspots for rainfed maize in Africa for worst-case climate model, RCP 7.0 (medium-high emissions scenario) Source: Authors using DSSAT. Notes: At least 5% cropland in pixel. Maize represents at least 25% of cropland. 40 Figure 10: Frequency of 1-in-20-year low maize yield events in Southern Africa: comparing frequency in the 2020s to the 2060s under various emissions scenarios. Source: Thomas, TS, Robertson RD, Strzepek K, and Arndt C. (2022) Extreme Events and Production Shocks for Key Crops in Southern Africa Under Climate Change. Front. Clim. 4 Policy recommendations The hotspot methodology could be useful to policy makers, development specialists, extensions workers, and donors in identifying locations which may require special focus to help farmers in those locations to adapt to climate change. The hotspots suggest a number of possible interventions, which may include developing new cultivars resistant to or tolerant of the kind of climate shocks expected; having extension workers train farmers in growing alternative crops more suited to the future climate or in moisture-preserving or water harvesting techniques, if the climate shock is related to inadequate water; helping farmers relocate to a more hospitable climate for growing the crop that they prefer; adding irrigation, if that would reduce risk of yield shocks; developing social protection plans to assist families adversely affected by climate change and extreme events; and, if appropriate, hasten the transition of some households from agriculture to the non-agricultural sector. 41 Use case 4: Improving the Spatial Granularity in Impact Evaluation Meta Data for Use Case 4 AOW1 Areas Global local outlook Prioritization and Evaluation Rapid response Impact Areas Climate adaptation & mitigation Environmental health & biodiversity Gender equality, youth & social inclusion Nutrition, health & food security Poverty reduction, livelihood and jobs Spatial Scope Global Regional Country: Kenya Within country Spatial unit of the OUPTUT of the use case Subnational Pixel: 8km by 8km Community Households Other Time scope of the OUPTUT of use case analytics Historic period: 2021 - 2023 Both historic + future Future Credit: ©2016CIAT/GeorginaSmith 41 By Chris Mwungu, Felix Otieno, Agnes Wanjau, Anirudha Ghosh | Alliance of Bioversity & CIAT Introduction and policy relevance Shamba Shape Up (SSU) is a makeover-style reality show that aims to meet smallholder farmers’ information needs by combining entertainment with agricultural and climate advisories, helping them to improve their livelihoods in the face of climate change (CGIAR Research Program on Climate Change, 2015). This study examined spatially heterogeneous impacts of the SSU weather and farming news on smallholder farmers in Kenya on crop productivity, income, and food security. The study’s research question was to examine whether households that watched the weather and farming news segment on SSU had better outcomes in terms of crop productivity, crop income, and household food security. This aligns with the broader CGIAR goals of evaluating the role of climate information services in supporting local adaptation strategies. This study responded to a demand from Media (producers of SSU) to understand how spatial targeting of advisories can improve reach and effectiveness. By applying Causal Forests (CFs), a machine learning method for estimating heterogeneous treatment effects, the study captures not only household-level variation but also spatial patterns across counties, to inform better- targeted climate advisories for the future for Kenyan extension services, digital broadcasters like Media, NGOs, or county-level government planners. Compared to aggregate or non-spatial analyses, this approach revealed localized variation in uptake and impact, which is essential for designing tailored advisory services. Method The study used CFs together with georeferenced household data to estimate the Individual Treatment Effects (ITEs) of watching SSU weather and farming news on agricultural outputs and household welfare. A cross-sectional household survey conducted in 2023 (Kiprop et al., 2024) was used to estimate the ITEs. CFs are a machine learning method that estimates heterogeneous treatment effects by leveraging variation in observed characteristics and treatment exposure across households. This method relies on causal inference using observed data, which is different from foresight analysis that explores multiple plausible futures to help prepare for uncertainty. The spatial dimension of the method is important for answering the research question, as it helps to identify where the intervention has the greatest uptake and impact. This enables more targeted and effective decision-making to enhance the delivery of climate advisory services. Results and key maps Significant spatial heterogeneity was observed in the impact of the SSU weather and farming news on household agricultural income. Figure 11 presents results from the CF analysis, showing the spatial variation in maximum, 42 median, and minimum ITEs in agricultural income across counties in Kenya. In addition, raw ITE estimates for individual households are included to provide a more detailed view without aggregation. The map of median ITEs reveals both positive and negative treatment effects, highlighting the location-specific influence of SSU. For example, negative median effects were observed in Nyandarua and Machakos, while counties such as Migori, West Pokot, and Makueni showed clearly positive median effects. These differences may be driven by county-level variations in agroecological conditions, access to complementary extension services, and baseline exposure to agricultural advisories. The map of maximum ITEs further demonstrates the heterogeneity of impacts. Although nearly all counties showed some positive maximum effects, the size of these effects varied widely. Counties in central Kenya, such as Nyandarua, Nyeri, and Murang’a, had lower maximum ITEs, whereas Kilifi, Kwale, Migori, West Pokot, Makueni, and Isiolo had much higher values, suggesting a stronger influence of SSU on agricultural income in these counties. Figure 11: Spatial variation in impacts of watching SSU weather and farming news 42 It is important to note that exposure to SSU may be endogenous. To address potential bias, the coefficient of variation in temperature and rainfall over the past five years was used as an instrumental variable. This approach assumes that farmers experiencing greater climate variability are more likely to engage with weather-related advisory services. To get feedback on this study, study results were shared with SSU producers, who confirmed the value of county-level disaggregation for targeting advisory content. As part of advancing this work, the strength of the instrument requires further validation in follow-up studies. Policy recommendations This analysis reveals that the impact of SSU on agricultural income varies significantly across regions, with counties like Migori, Kilifi, and West Pokot showing stronger positive outcomes, while others such as Nyandarua display lower impacts. These findings highlight the importance of spatial targeting over uniform scaling. Policymakers such as Kenya’s Ministry of Agriculture and Livestock Development, county-level agricultural departments, Media, and development partners should prioritize expanding similar programs in more areas while investing in tailored outreach strategies for marginalized and underserved areas with limited observed benefits. Integrating spatial analytics into decision- making processes can enhance targeting efficiency, optimize resource allocation, and strengthen monitoring and evaluation systems. The method is most applicable in contexts where high-quality, georeferenced household- level data are available, and where treatment exposure is clearly defined. However, its reliability decreases in areas with limited data or where relevant confounding variables are unobserved. The next steps include applying this method to other advisory platforms, such as iShamba SMS services, integrating spatial ITE maps with cost-effectiveness analysis, and co- developing region-specific media strategies. Future research should also explore the use of panel data and/or randomized controlled trials (RCTs) to strengthen causal inference of advisory services. These next steps will require collaboration with various extension partners such digital advisory producers, county extension services, and development partners to co-develop region-specific scaling strategies. 43 Use case 5: Simulated Land Use/Land Cover Change Meta Data for Use Case 5 AOW1 Areas Global local outlook Prioritization and Evaluation Rapid response Impact Areas Climate adaptation & mitigation Environmental health & biodiversity Gender equality, youth & social inclusion Nutrition, health & food security Poverty reduction, livelihood and jobs Spatial Scope Global Regional Country Within country Spatial unit of the OUPTUT of the use case Pixel: ½ degree (56km north to south) Community Households Other Time scope of the OUPTUT of use case analytics Historic period Both historic + future (2005 and 2050) Future Credit: ©2009CIAT/NeilPalmer 44 By Richard Robertson | International Food Policy Research Institute (IFPRI) Introduction and policy relevance How will land use change in the future? Which kinds of ecosystems will be most disrupted by cropland expansion? These are the kinds of questions that can be addressed by this collection of spatially explicit land use simulations. The term “land use” has at least three different meanings or uses, itself. One sense is concerned with crop allocation within a single farm. Another (often non-agricultural) idea is more about externally imposed regulation like “zoning” between plots which either designate or observe categories like natural, industrial, agricultural, residential, or commercial. A more expansive version of this looks beyond human managed land to consider different kinds of natural land such as deserts versus forest. Analyzing these different versions of “land use” usually views them as resulting from different drivers such as the hyper-local scale coming from profit maximization within a single farm/firm; the regional zoning type flowing from the legal, business, political, and economic environment; and the broad global natural ecology kind of land use because of climate and geology. The datasets reported here bring parts of these different scales together into a single framework. It starts with the IMPACT model of global agriculture (Robinson, et al., 2004). IMPACT operates at roughly national levels for demand markets and some smaller regions for the production side. The supply part assumes the existence of cropland areas and their corresponding yields, both of which evolve through time subject to the forces of demographic and economic changes as well as the shocks to productivity expected as the climate changes. By taking these market-level simulations and developing internally consistent, fine grained spatial representations of the outputs, we can gain some insight at the different levels of land use thought: likely crop mixes (across broad categories) within pixels like the hyper-local, interplay between expansion/contraction of cropland and remaining natural areas like the regional and shifts within the natural land mixes driven by climate conditions. The resulting maps and various aggregations based on them can be used by modelers to assess whether their models are behaving as a changing world. Method Global maps are created showing the number of hectares of each land category that are simulated to be present in each half-degree pixel. Maps are developed for a historical case of 2005 and several future cases based on GCM simulated climates for 2050. This is done in three steps. First, a statistical model based on historical data is developed to quantify the association between the temperature and rainfall of a location and the mixture of natural land types found there. The model is then used to simulate the likely amounts of natural land to be expected for each location on the map under the various climate situations of interest. 45 Second, the regional amounts of cropland areas for several major crops along with “all others” are extracted from IMPACT simulations for the time periods and climate situations of interest. Within each region, those areas (adjusted for multicopying) are allocated, bit by bit, to individual pixels so that the pixels within the region add up to the right amount of rainfed maize area, irrigated rice area, and so on. The allocation heuristic tries to balance spreading the areas out and allowing overlapping areas against concentrating particular crops in the most favorable locations. This is guided by attractiveness indices for each crop based on simulated potential yields, distance to population centers, how flat or uneven the elevation is within the pixel, and how similar the climate is to climates historically supportive of agriculture. Finally, the natural areas in each pixel are scaled back to create space for the allocated cropland. Results and key map Pulling together the economic model (reflecting the interaction of demand and supply) along with the empirical model of natural land types (reflecting the influence of climate) allows us to extract several insights for the timeframe investigated (2005 to 2050). First, climate change is likely to, overall, make cropland less productive and hence more croplands will be needed in the future than otherwise would have been. However, the increase due to climate change is fairly small compared to the overall increase in cropland need due to basic economic and demographic factors. Second, the changes in the amounts of cropland needed for particular crops coupled with the pattern of which locations are better and worse for producing those crops do not result in major rearrangements of production regions. Third, the natural land types exhibit much greater sensitivity to climate change. Tropical forests are typically of interest. In South America, the climate will shrink the area hospitable to forests much more than cropland expands. The other tropical forests face more balanced threats from both mechanisms. Overall, climate change seems to represent a larger danger to forests than cropland expansion. The map below shows the spatial distribution of pixels exhibiting gains or losses of evergreen broadleaf forest (mainly tropical forests) with cropland changes over top. The changes in forest are broader than and more-or-less independent of cropland expansion. (Robertson et al., 2023a); Figure 12) 46 Figure 12: Land use changes Policy recommendations Forest protection will require more than fences and guards since the climate conditions conducive to forests will be found in progressively fewer places. While the broad patterns of where cropland is located may change slowly, there will still be a significant portion that experiences upheaval where the mix of crops is likely to lose a major crop or gain a new one. Producers in those cases may need to learn to grow a different set of plants as well as deal with new marketing channels. The model has a fundamental limitation in that it is a comparative static model that does not say how the world transitions between the various situations, but only what those situations are likely to look like. It also only looks at a few of the largest crops on their own; future work will expand consideration to more individual crops and reduce the size of “all others”. 48 Use Case 6: Spatial Dimensions of Poverty and Agricultural Productivity: A Spatial Econometric Analysis Met