Phenology-Guided Multi-Source Data Fusion Improves Soil Organic Matter Mapping Across Chinese Farmlands

Accurate spatial mapping of soil organic matter (SOM) in croplands is essential for sustainable agriculture, but remains challenging because of crop-cover interference and regional environmental heterogeneity. Here, a phenology-guided multi-source data fusion framework was developed to predict topsoil (0–20 cm) SOM across five contrasting Chinese farmland regions. The framework integrated baseline environmental covariates, multi-year Sentinel-2 phenological parameters, and GaoFen-5 (GF-5) hyperspectral features using forward recursive feature selection (FRFS) with machine learning algorithms, and all models were evaluated over 20 random training–validation splits of each study area. Fusing phenological and hyperspectral covariates with baseline data performed best in every region, raising the mean validation R2 from 0.508 to 0.875 and lowering the mean root mean square error from 4.148 to 2.127 g kg−1. FRFS reduced predictor dimensionality by 83.8% on average, was more accurate than the all-variable strategy in all 100 paired comparisons, and lowered the split-to-split standard deviation of validation R2 from 0.129 to 0.030.