High-resolution maize yield mapping across Africa using earth observation and machine learning, deep learning, and foundation model

Africa's food security is increasingly threatened by climate change and population growth, creating an urgent need for high-resolution crop yield maps to support precision agriculture and climate adaptation; however, much of the continent remains poorly monitored due to limited ground observations. This study introduces a novel 250-m spatial resolution maize yield prediction framework for 42 African countries. Our methodology utilizes Net Primary Production (NPP) data to spatially disaggregate national FAO yield statistics, generating fine-scale training data for supervised learning. A comprehensive feature set of 296 variables was constructed by integrating multi-source Earth observation, climate, and soil data. We evaluated multiple benchmark models for tabular data, including XGBoost, LightGBM, a Hybrid Deep Neural Network (HDNN), and, for the first time in this context, a tabular foundation model, the Tabular Prior data Fitted Network (TabPFN). Using an expanding-window temporal cross-validation strategy, XGBoost achieved the highest temporal R 2 (0.78), while TabPFN demonstrated superior spatial generalization and the lowest mean absolute percentage error (MAPE ≈ 25 %). Causal inference and ablation analyses highlighted the predictive importance of vegetation indices (e.g., NDVI, NDWI), drought metrics, and soil properties. Model outputs closely matched national yields (R 2 > 0.75; MAPE ≈ 26–28 %). This study provides a scalable framework for yield monitoring in data-scarce regions and represents the first successful application of tabular foundation models in continental-scale agricultural prediction.