<b>Dataset: Enhancing Cross-Regional SOC Prediction with Feature Selection and Transfer Learning: A Case Study in Temperate and Subtropical Soils in China</b>

Soil organic carbon (SOC) content in the 0–20 cm topsoil layer was used as the target variable for model development and validation. To ensure spatial consistency and alignment with global standards, SOC values were obtained from the SoilGrids250m product developed by ISRIC(Hengl et al., 2017b). This dataset provides harmonized, machine-learning-based estimates of SOC concentrations, derived from an extensive archive of globally distributed soil profiles.For each study region, 2000 spatially randomized sampling points were generated within the defined study boundaries. SOC values at these locations were extracted from the SoilGrids raster using bilinear interpolation to achieve spatial smoothing and continuity. These extracted estimates served as proxy ground-truth values for model training, validation, and cross-regional transfer analysis.This approach ensured the use of standardized, high-resolution SOC data, enabling robust comparative modeling across regions with contrasting edaphoclimatic conditions.