Earth observation derived yield forecasting and estimation in low- and lower-middle-income countries dominated by smallholder agriculture: A review

Yield forecasting and estimation using Earth Observation (EO) has become a crucial decision-support tool for stakeholders in agricultural and food security. Missing in the development of these methods are smallholders, who are important to global food security and the promotion of sustainable agricultural development. A comprehensive understanding of crop yield research in smallholder-dominated agricultural regions will not only help to formulate reasonable agricultural policies, but also for advancing scientific progress in the field of agricultural sciences. In this study, we focus on low-income and lower-middle-income countries with predominantly smallholder agricultural systems (LILMIC-SAS), and provide a comprehensive literature review on crop yield forecasting and estimation studies in these regions. Based on our search criteria, a total of 554 relevant studies were initially retrieved. After applying exclusion criteria, 268 studies (published between 2012 and 2022) remained for the follow-up analysis. We present a systematic analysis of yield modelling for 25 crops in 73 countries in terms of approaches, models, data used, input features, study scale, yield map resolution, evaluation metrics and timeliness. According to our analysis, the statistical/empirical (SE) approaches are the most widely used (49.63%), followed by machine learning (ML, 29.1%) and process-based (PB, 19.4%) approaches. Although SE approaches are most common, ML and PB approaches often achieve higher accuracy in the late and post-season stages (R 2 mostly between 0.8 and 1), and ML approaches have become much more popular since 2018. MODIS (42 studies) and Sentinel-2 (30 studies) are the most used EO data sources, and the normalized difference vegetation index (NDVI, 100 studies) is the most frequently used EO features. Temperature (136 studies) and precipitation (135 studies) are the most used weather features. We also listed 26 EO data and 21 weather data products. Finally, we present a multi-criteria decision analysis (MCDA) framework to assess the current state of crop yield modelling studies in the LILMIC-SAS regions across the criteria of quality, applicability, replicability and validity. By summarizing and presenting the status of crop yield research in the LILMIC-SAS regions, this study offers valuable insights for food security stakeholders to develop yield models for operationalization and provides directions for future research on crop yield forecasting and estimation in smallholder and low-income contexts.