Erosion-SAM | Hadi Shokati

The Erosion-SAM project presents an automated approach for semantic segmentation of soil erosion features caused by water in agricultural landscapes. This work demonstrates how fine-tuning the Segment Anything Model (SAM) enables accurate detection of erosion and deposition features from high-resolution aerial imagery.

Overview

Soil erosion by water poses significant challenges in agricultural landscapes. Traditional detection methods rely on manual field surveys or visual interpretation of aerial imagery—processes that are time-consuming and subjective. This project leverages the Segment Anything Model (SAM), a foundation model pre-trained on over 1 billion masks, and fine-tunes it for the specialized task of identifying soil erosion features.

Study Area and Dataset

Location: Southeastern Bavaria, Germany

Data Acquisition: The project involved acquiring orthophotos following heavy rainfall events in erosion-prone areas between May and September of 2011 and 2012. To determine suitable acquisition times, RADOLAN rainfall data, with a spatial resolution of 1 × 1 km², were analyzed. When rainfall events met specific thresholds (either a total rainfall of at least 10 mm or a maximum 30-minute intensity exceeding 10 mm/h) indicating high erosion potential, aerial surveys were conducted using a small aircraft within 30 days after the identified erosive event.

Dataset: The images were then georeferenced. We manually identified and masked 405 individual images from the aerial imagery that exhibited erosion and deposition features such as ephemeral gullies, rills and sediment fans. The fields were categorized into 3 land cover categories:

Grassland: 128 images
Cropland: 277 images (vegetated: 131, bare: 146)

A) Soil erosion map of Germany and B) examples of manual segmentation of agricultural fields with erosive rainfall, showing erosion and deposition features across different land covers.

Methodology

We applied fine-tuning to adapt SAM’s pre-trained weights to soil erosion detection. Fine-tuning is a transfer learning approach that modifies a pre-trained model for a specific task, improving performance while reducing the need for large labeled datasets and computational resources.

Overview of the Segment Anything Model (SAM) architecture. The fine-tuning process adapts the model to soil erosion detection.

Data Processing:

Large orthophotos divided into 256×256 pixel patches
Empty masks filtered out
Data split: 70% training, 15% validation, 15% testing

Model Configuration:

Base: SAM (vit_b variant)
Optimizer: AdamW
Loss: Combined Dice loss and focal loss
Strategy: Fine-tuning the mask decoder

Results

The fine-tuned Erosion-SAM model successfully detected soil erosion features across different land cover types. We evaluated multiple fine-tuning approaches and compared them against the original SAM model.

Performance Metrics

The table below shows the performance of different approaches across land cover types. The numbers in blue and orange represent the increase and decrease in performance compared to the original SAM baseline.

Performance comparison of different fine-tuning approaches (Cropping, Resizing without prompt, Resizing with prompt) against the original SAM baseline across different land cover types.

The Resizing approach with prompt demonstrated the best overall performance with substantial improvements across all metrics and land cover types, particularly excelling in grassland detection (IoU: 0.75, +0.38 improvement).

Segmentation Examples

Examples of erosion feature segmentation by Erosion-SAM showing original images, ground truth, and model predictions.

Applications

The generated data sets can be applied to machine learning-based SE modeling, providing accurate and consistent training data across different land cover types, and offering a reliable alternative to traditional SE models. In addition, erosion-SAM can make a valuable contribution to the precise monitoring of SE with high temporal resolution over large areas, and its results could benefit reinsurance and insurance-related risk solutions.

Publication

Citation:

Shokati, H., Engelhardt, A., Seufferheld, K., Taghizadeh, R., Fiener, P., Lensch, H., & Scholten, T. (2025). Erosion-SAM: Semantic segmentation of soil erosion by water. CATENA, 254, 108954.

DOI: 10.1016/j.catena.2025.108954

Status: Published in CATENA (Shokati et al., 2025)

Soil erosion (SE) by water threatens global agriculture by depleting fertile topsoil and causing economic costs. Conventional SE models struggle to capture the complex, non-linear interactions between SE drivers. Recently, machine learning has gained attention for SE modeling. However, machine learning requires large data sets for effective training and validation. In this study, we present Erosion-SAM, which fine-tunes the Segment Anything Model (SAM) for automatic segmentation of water erosion features in high-resolution remote sensing imagery. The data set comprised 405 manually segmented agricultural fields from erosion-prone areas obtained from the rain gauge-adjusted radar rainfall data (RADOLAN) for bare cropland, vegetated cropland, and grassland. Three approaches were evaluated: two pre-processing techniques— resizing and cropping — and an improved version of the resizing approach with user-defined prompts during the testing phase. All fine-tuned models outperformed the original SAM, with the prompt-based resizing method showing the highest accuracy, especially for grassland (recall: 0.90, precision: 0.82, dice coefficient: 0.86, IoU: 0.75). SAM performed better than the cropping approach only on bare cropland. This discrepancy is attributed to the tendency of SAM to overestimate SE by classifying a large proportion of fields as eroded, which increases recall by covering most of the eroded pixels. All three fine-tuned approaches showed strong correlations with the actual SE severity ratios, with the prompt-enhanced resizing approach achieving the highest R2 of 0.93. In summary, Erosion-SAM shows promising potential for automatically detecting SE features from remote sensing images. The generated data sets can be applied to machine learning-based SE modeling, providing accurate and consistent training data across different land cover types, and offering a reliable alternative to traditional SE models. In addition, erosion-SAM can make a valuable contribution to the precise monitoring of SE with high temporal resolution over large areas, and its results could benefit reinsurance and insurance-related risk solutions.