Integrating Erosion Severity and Vegetation Stress Indicators for Marginal Land Mapping in Northeastern India Using a Machine Learning Approach

Marginal lands are areas with low or declining agricultural potential due to persistent soil degradation, vegetation stress, or long-term disturbances in land use. Mapping such lands is especially important in the Northeastern Region (NER) of India, where steep slopes, intense monsoonal rainfall, and widespread shifting cultivation create highly dynamic and fragile landscapes. Recent nationwide assessments have also highlighted that many districts in this region exhibit very high susceptibility to soil erosion, making the identification of marginal lands essential for restoration planning and sustainable land-use management. In this study, we propose a machine-learning framework to classify marginal lands by jointly analyzing erosion severity classes, vegetation dynamics, and bare-soil exposure. Multi-year Sentinel-2 data are used to compute pixel-wise Poor Vegetation Frequency from NDVI (Normalized Difference Vegetation Index) and Bare Soil Frequency from BSI (Bare Soil Index), providing robust indicators of vegetation stress and soil exposure. These variables are combined with potential soil loss estimates, topographic attributes, and land-use information to form a comprehensive feature set. The model is trained and evaluated using observed degradation patterns from the Desertification and Land Degradation Atlas of India, enabling an independent assessment of classification performance. The resulting marginal land maps show strong spatial agreement with known erosion-prone and degraded zones across the selected study region, with bias mass balance values generally ranging between 0.6 and 0.8. This study demonstrates the value of integrating erosion severity and vegetation dynamics within a machine-learning environment, offering a scalable approach for exploring and mapping marginal lands in complex and data-constrained regions.