Subfield maize yield prediction using Sentinel-2 land surface phenology metrics and interpretable stacked ensemble machine learning

The non-linear relationships between crop yield and phenology are not interpreted in most previous subfield yield‑prediction studies based on remote sensing and machine learning, which assume the prediction model is a black box instead. The research on land‑surface phenology metrics in these studies is particularly scarce, despite their potential to uncover these relationships in combination with model-agnostic, interpretable machine learning. The aim of the study was to explain the variance of maize yields on the subfield level by combining land‑surface phenology metrics derived from Sentinel-2 imagery with interpretable ensemble machine learning. The Normalized Difference Vegetation Index (NDVI) and Normalized Difference Moisture Index (NDMI) were used to model phenological timing and magnitude metrics through curve-fitting models, using them as covariates for maize‑yield prediction on two contrasting subsets (Subset A: Gleysols; Subset B: Phaeozems, Gleysols and Luvisols). Four individual machine‑learning methods were evaluated with a stacked ensemble approach based on repeated nested cross-validation. Model-agnostic interpretability approaches, including permutation feature importance and accumulated local effects plots, were used to interpret the relationship between the most impactful covariates and maize yield. Ensemble models based on NDMI produced the highest maize‑yield prediction accuracy on both Subset A (R2 = 0.708, RMSE = 0.543 t ha−1) and Subset B (R2 = 0.922, RMSE = 0.279 t ha−1). The findings reveal that NDMI is more resistant to the LAI saturation effect inherent to NDVI, capturing critical water-stress dynamics during the grain-filling stage. Model interpretation identified the Peak of Season (POS) magnitude and the Start of Season (SOS) timing as the most impactful drivers. These results demonstrate that interpreting ensemble machine learning through phenological windows provides a robust framework for understanding how soil-specific water dynamics and sowing timing are associated with final maize productivity.