Masked autoencoder and Mamba-based self-supervised segmentation of SAR imagery for riverbank erosion detection

Abstract

Riverbank erosion in Bangladesh causes significant environmental and socioeconomic challenges, including land loss that displaces communities and damages local economies. Traditional monitoring methods, such as field surveys and manual satellite imagery analysis, are labor-intensive, inaccurate, and lack continuous data. Synthetic Aperture Radar (SAR) imagery from Sentinel-1 provides precise, all-weather imaging, overcoming cloud cover limitations. We created a dataset using Sentinel Hub, retrieving daily SAR data from 2014 to 2024 for major Bangladeshi rivers, including Jamuna (Sirajganj, Tangail, Manikganj), Padma(Shariatpur, Munshiganj, Faridpur), Meghna (Chandpur, Lakshmipur, Narsingdi), Brahmaputra (Chilmari, Fulchhari, Bahadurabad), and Teesta (Lalmonirhat, Kurigram). This study addresses temporal data gaps in SAR imagery using Generative Adversarial Networks (GANs) to reconstruct missing data, ensuring continuous riverbank observations. It also employs a self-supervised learning (SSL) framework with a convolutional Masked Autoencoder (MAE) and Mamba blocks for land-water segmentation without labeled data. Our pipeline achieved an average River Intersection over Union (IoU) of 0.8550 and a Dice coefficient of 0.9038 across five rivers from 2015 to 2024, enabling robust segmentation. This approach supports scalable riverbank monitoring for disaster prevention and sustainable environmental management in Bangladesh.