AI Model Automates Detection of Riverbank Erosion Across Bangladesh

Riverbank erosion in Bangladesh poses an urgent environmental challenge, destroying land, threatening infrastructure, and displacing communities. Yet tracking the problem across the nation's extensive river systems has remained labor-intensive and inconsistent. A new study demonstrates how machine learning can accelerate this critical monitoring work.

The Challenge#

Bangladesh faces significant erosion along its riverbanks, particularly in vulnerable areas. Analyzing satellite imagery to spot erosion patterns has historically required painstaking manual effort by researchers, making it difficult to maintain consistent assessments across large stretches of waterway.

A Specialized AI Solution#

Researchers took the Segment Anything Model (SAM)—a large-scale image segmentation tool—and adapted it specifically for riverine environments. Rather than retraining the entire system, the team kept most of the model's core components fixed and fine-tuned only its lightweight mask decoder, a parameter-efficient approach that reduces computational overhead.

The team trained the model on 500 satellite image pairs spanning 2003 to 2025, drawn from Google Earth archives of erosion hotspots including Mokterer Char, Kedarpur, and Chowhali Upazila. Each image was labeled at the pixel level to distinguish river channels, stable land, and eroded areas.

Performance and Accuracy#

The adapted model showed strong results across key metrics:

• Erosion detection: An IoU (intersection-over-union) score of 0.867 and F1-score of 0.928 on held-out test data
• Area measurement: Estimated erosion area differed from ground truth by only 0.17%, indicating reliable quantification
• Geographic transfer: The model generalized reasonably well to previously unseen riverbank regions, though accreted (deposited) land remained difficult to identify from standard RGB images alone

Implications and Next Steps#

The findings suggest that adapted foundation models can support faster, more consistent erosion monitoring across river networks. Researchers note that incorporating multi-modal remote sensing data—combining optical, radar, and other sensor types—could further improve accuracy and help tackle remaining challenges like sediment accretion detection.