Automated high-resolution detection of soil erosion and deposition features in agricultural fields using the fine-tuned Segment Anything Model

Water-induced soil erosion is a critical threat to the sustainability of agriculture worldwide, as it destroys nutrient-rich topsoil and causes significant economic costs. Conventional soil erosion monitoring methods, such as the Universal Soil Loss Equation (USLE) and its revised version (RUSLE), often face challenges with data collection for calibration or validation. Although machine learning offers promising alternatives, they usually require large data sets, which can limit their practicality. Recent advances in deep learning have introduced several innovative techniques such as transfer learning to reduce the need for extensive data. This study presents Erosion-SAM, a novel framework that fine-tunes the Segment Anything Model (SAM) to automatically identify erosion and deposition features using high-resolution remote sensing imagery. RADOLAN radar rainfall data with a spatial resolution of 1 km was used to identify erosive events and determine erosion-prone agricultural fields including grassland, vegetated cropland, and bare cropland, in southeastern Bavaria, Germany. High-resolution orthophotos (0.2 m) were taken for fields with erosive events indicating significant erosion potential. These orthophotos were then manually segmented by experts to delineate precise erosion and deposition features and subsequently used as input data for fine-tuning SAM. Three pre-processing strategies were evaluated during the fine-tuning process: resizing, cropping, and prompt-based resizing. The prompt-based resizing method performed best, especially in grassland, with an IoU of 0.75, a Dice coefficient of 0.86, a precision of 0.82 and a recall of 0.90. While the baseline SAM performed better than the cropping method in bare cropland, it overestimated erosion and deposition, which increased the recall values. The fine-tuned methods agreed well with the actual soil erosion severity ratios, with the prompt-based resizing method achieving an R² of 0.93, demonstrating superior predictive performance. Erosion-SAM showcases the potential to revolutionize soil erosion monitoring by automatically detecting erosion and deposition features across different land covers with high accuracy. Moreover, it generates high-quality, consistent data sets as valuable input for machine learning-based erosion modeling for different land covers. Its scalability and high spatial and temporal resolution also make it invaluable for large-scale erosion monitoring and risk assessment, including applications in the insurance and reinsurance industry.