Optimization of Sampling Density in Regional-Scale Soil Organic Matter Mapping and Its Impact on Prediction Accuracy

High sampling density in digital soil mapping improves accuracy but increases costs. This study optimized sampling density for soil organic matter (SOM) mapping in Longjiang County using an XGBoost model, integrating Sentinel-2 imagery (April–May) and environmental variables. We evaluated the impact of reducing a dataset of 2059 surface soil samples using a soil-type-based stratified random (SR) sampling strategy. The optimal scale was determined using a cost–performance index restricting RMSE increases to under 10%. Results showed that integrating April–May environmental variables significantly improved model accuracy. Crucially, the SR strategy allowed training samples to be reduced to just 434 while maintaining stable predictive performance (R2 = 0.499, RMSE = 5.319 g/kg). Furthermore, SOM spatial prediction patterns at this optimized, lower density aligned closely with high-density results, showing no systematic bias. In conclusion, combining environmental constraints with stratified thinning allows regional SOM mapping to maintain reliable accuracy at significantly reduced sampling densities. This offers a cost-effective approach for optimizing soil monitoring networks in black-soil regions.