Optimization of Soil Pixel Extraction from Sentinel-2 Time-Series Based on the TDBSS-RF

Soil is a fundamental component of terrestrial ecosystems and plays a crucial role in Earth science research. As soil in the fragmented Huang-Huai-Hai Plain is frquently covered by vegetation or crop residues, it is essential to develop a effective technique for accurate extraction of soil spectral information. In this study, we integrated two-dimensional bare-soil separation (TDBSS) based on computer pattern recognition with an ensemble machine learning algorithm (Random Forest) to develop an enhanced TDBSS-RF classifier for bare soil pixel extraction. The TDBSS algorithm was optimized by integrating multiple spectral index combinations (NDI5-DFI, NDI7-DFI, NBR2-NDI5, NBR2-NDI7, NBR2-NDVI, and NDVI-STI), achieving a bare soil extraction rate of 64%. Additionally, the RF classifier was refined through feature selection and classification optimization, the extraction rate was further improved from 72% to 76%. Validation using sample points in Google Earth Engine (GEE) and comparison with a multi-band soil map showed that 67% of the samples achieved an R2 greater than 0.8. These results demonstrate that the proposed method not only generates high-spatial-resolution soil maps but also significantly optimizes computational algorithms, providing a robust and technically enhanced solution for precision agriculture and sustainable land management.