Explainable Machine Learning for Wheat Biomass Integrating Sentinel-1/2, PlanetScope and In-Situ Weather Data

Francisco Zambrano , Abel Herrera, Mauricio Molina-Roco

Global food security faces increasing challenges from climate change, making accurate monitoring of essential crops like wheat (Triticum aestivum) critical. This research introduces an explainable machine learning (ML) framework to estimate and forecast wheat above-ground biomass (AGB) in central Chile across the 2020–2023 growing seasons. The study uses a two-stage approach: first, in-season AGB estimation, and second, AGB forecasting at harvest. The predictive model integrates comprehensive data from four field sites: in-situ weather (precipitation, growing degree days, and soil moisture) and satellite imagery derived from three platforms (Sentinel-1 SAR, Sentinel-2, and PlanetScope optical sensors). We evaluated 41 distinct models, derived from five ML algorithms (Random Forest, XGBoost, GLMnet, bagMLP, KNN) and an ensemble, across eight predictor variable combinations. Hyperparameter optimization was performed via cross-validation, and results were interpreted using the moDel Agnostic Language for Exploration and eXplanation (DALEX) framework to ensure model transparency. Results showed high accuracy, particularly in the estimation stage (Stage 1), where models combining Sentinel-1 and weather data achieved an R2 > 0.91 and an RMSE of approximately 3.2 t/ha. Key predictors were accumulated growing degree days, Sentinel-1 variables, and soil moisture, which exhibited complex non-linear interactions. For harvest prediction (Stage 2), soil moisture and SAR data remained the most critical factors, enabling accurate forecasting at 1–4 month leads with R2 values ranging from 0.94 to 0.86 and RMSEs of 1.11–1.74 t/ha. This cloud-resilient, highly accurate methodology offers valuable insights for precision agriculture, optimizing practices like irrigation, fertilization, and risk management. While the current model’s utility is limited by reliance on localized in-situ data, future incorporation of satellite-derived soil moisture could enhance its global applicability for yield forecasting.