Explainable Hybrid Modelling of Winter-WheatvBiomass in Bavaria Using STARFM-Fused NDVI, LUE, and Random Forest

Kurzfassung

Accurate estimation of crop biomass at fine spatial scales is essential for precision agriculture and climateresilient food systems. This study proposes a hybrid framework that couples satellite data fusion, a Light Use Efficiency (LUE) crop model, and Random Forest (RF) machine learning to map winter-wheat (WW) biomass across Bavaria, Germany, from 2001 to 2019. A temporally continuous 30 m NDVI archive was produced with the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM) by blending Landsat and MODIS observations. The fused NDVI drove an LUE model to generate annual 30 m biomass maps, which were aggregated to a 5 km2 hexagonal grid. An RF model, trained on seasonal climate statistics, elevation, slope, soil potential, the Shannon Diversity Index (SHDI), and small-woody-feature (SWF) cover, achieved R2=0.87 and RMSE=2.85dt ha-1 versus district yield records. SHapley Additive exPlanations and Partial Dependence Plots revealed three dominant interaction pairs - temperature × soil potential, SHDI × SWF, and solar radiation × elevation - that explain regional biomass gradients across Bavaria's six agroecological zones. The framework demonstrates that fusing multisensor NDVI with explainable machine learning provides a scalable, transferable solution for biomass monitoring in heterogeneous landscapes.