Developmental-Stage-Aligned Explainable Machine Learning Identifies Climate-Sensitive Windows and Conditional SOM-Related Buffering Patterns in Maize Growth in Western Liaoning, China

Identifying climate-sensitive developmental windows is important for understanding maize responses to climate variability in semi-arid to semi-humid transition regions. This study integrated daily meteorological data, soil organic matter (SOM) records, regional maize field information, and maize growth processes reconstructed with the World Food Studies (WOFOST) model for 2000–2023 across 72 modelling grids and 21 counties in western Liaoning, China. The WOFOST-derived development stage (DVS), maximum aboveground dry matter (max_TAGP), final storage-organ dry weight (final_TWSO), biomass source-sink conversion efficiency (BSCE), and storage-organ filling rate (SGFR) were used as modelled process indicators rather than continuous field measurements. Four feature systems were evaluated using six machine-learning algorithms: monthly climate features (MCF; calendar-month climate, SOM, and spatial variables), DVS-aligned climate features (DCF; developmental-window climate, SOM, and spatial variables), DVS-SOM-development process features (DPF; DCF plus DVS dates and stage durations), and hybrid feature fusion (HFF; monthly, DVS-stage, SOM, developmental, and spatial variables). SOM classes were defined statistically by SOM3 tertiles, with the high-SOM tertile corresponding approximately to SOM3 > 14.29 g kg−1. Independent temporal testing showed limited-to-moderate predictive skill: MCF + CatBoost performed best for max_TAGP and final_TWSO (R2 = 0.231 and 0.362), DPF + LightGBM performed best for BSCE (R2 = 0.176), and HFF + XGBoost performed best for SGFR (R2 = 0.405). In spatial grouped cross-validation, which evaluates transfer among withheld grids rather than prediction of future years, HFF achieved the highest R2 for all four targets (0.717, 0.702, 0.726, and 0.804, respectively). SHAP analysis was interpreted as model-based association rather than causal evidence and indicated that final_TWSO was mainly associated with July precipitation and late-season humidity, BSCE with early heat exposure and developmental timing, and SGFR with late-grain-filling temperature and radiation. Higher SOM was associated with a smaller max_TAGP decline during heat years, but the pattern was not stable under compound heat–drought stress. These results suggest that stage-aligned and hybrid feature systems are useful for interpreting WOFOST-derived maize process indicators, while SOM-related buffering should be considered conditional and potentially associated with differences in soil properties and water availability.