Stratified machine learning models for wheat yield estimation using remote sensing data

Field-Level cereal yield estimation using Machine Learning (ML) models poses a significant challenge especially when applied across large areas. A large sample size is required to represent the high yield variability caused by varying topographic and climatic conditions. To enhance ML-based prediction accuracy, we propose to decompose the complexity of agricultural landscape using landforms and agro-ecological zones and use these classes as spatially explicit constraints to partition field samples. We trained three ML models using remote sensing data to estimate wheat yield. When training ML models without the mentioned spatial constraints, we achieved an R²=0.58 and RMSE=840kg/ha. Training ML separately across various landform classes increase the accuracy. For instance, wheat yield cultivated in plain areas was predicted with R2=0.72, and RMSE=809kg/ha. These results emphasized the potential of training ML separately across main landform classes for improving the accuracy of yield predictions across diverse geographical contexts.