Proxy -Labelling via Rusle and Temporal NDVI for deep learning model
This study proposes a hybrid, data-driven framework for scalable soil erosion prediction that integrates the empirical Revised Universal Soil Loss Equation (RUSLE) with five years of temporal Normalized Difference Vegetation Index (NDVI) satellite data. To overcome the scarcity of manually annotated ground-truth datasets, an automated proxy-labeling engine cross-references theoretical soil loss against historical vegetation decline to filter out seasonal agricultural cycles and generate discrete training masks. These labels train a U-Net deep learning segmentation model on multi-spectral Sentinel-2 imagery alongside topographic, rainfall, and soil parameters via Google Earth Engine. Evaluated across 50 diverse geographic regions in Maharashtra, the model achieves 95.10% overall pixel accuracy and 73.30% mIoU while reducing inference time to under 2.5 seconds. Additionally, the pipeline incorporates a rule-based mitigation engine that isolates critical erosion hotspots and recommends geo-tagged physical interventions such as reforestation or retaining walls based on local terrain metrics
Authors
- Sahil Pojee
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24422519
- Primary Topic
- Soil erosion and sediment transport
- Type
- preprint