News
02.06.2025
Introducing Erosion-SAM: Advancing Soil Erosion Detection with AI
Our doctoral student, Hadi Shokati has recently published a significant paper titled "Erosion-SAM: Semantic Segmentation of Soil Erosion by Water" in the journal Catena.
About the Research: This study introduces Erosion-SAM, a fine-tuned version of the Segment Anything Model (SAM), tailored for the automatic detection of water-induced soil erosion features in high-resolution remote sensing imagery. The model was trained on a dataset comprising 405 manually segmented agricultural fields from erosion-prone areas, including bare cropland, vegetated cropland, and grassland.Three approaches were evaluated to pre-process the data: resizing, cropping, and prompt-based resizing. The fine-tuned models surpassed the original SAM's performance, with the prompt-based resizing method showing the highest accuracy, particularly in grassland areas (recall: 0.90, precision: 0.82, dice coefficient: 0.86, IoU: 0.75). This work represents a significant advancement in soil erosion modeling, offering a reliable alternative to traditional methods and paving the way for more accurate environmental monitoring and risk assessment.
Explore the full paper here:
https://www.sciencedirect.com/science/article/pii/S0341816225002565
Access the GitHub code here:
https://github.com/hadi1994shokati/Soil-erosion-segmentation
Visual Results: The accompanying figure illustrates the segmentation outcomes of the original SAM and Erosion-SAM using different preprocessing techniques, providing a clear comparison of their effectiveness.
Acknowledgments:
We extend our sincere gratitude to our collaborators from the University of Augsburg: Professor Peter Fiener and Kay Seufferheld, for their invaluable contributions to this research. We also express our appreciation to Professor Hendrik P.A. Lensch and Dr. Andreas Engelhardt for their significant input, particularly in the fields of computer vision.
Teilen
