RiWiX: Toward Automatic River Water Width Extraction From High-Resolution Satellite Imagery Using Swin Transformers

Abstract

Accurate extraction of water surface and width of river systems is essential for hydrological modeling, flood monitoring, and water resource management. Traditional field-based measurements are often labor-intensive, time-consuming, and error-prone. With advances in satellite imaging and geospatial data processing, automated methods are increasingly replacing manual approaches. Yet, remote sensing-based river width extraction remains challenging due to dynamic water levels, vegetation interference, variable discharge, spectral reflections, and heterogeneous surface appearance. Existing convolutional neural network (CNN)-based encoder-decoder models have performed well for large rivers but struggle with narrow ones due to limited high-level semantic representation. Recent transformer-based models capture long-range dependencies through self-attention but are computationally heavy and underexplored in remote sensing. This paper presents RiWiX (River Water Width Extraction), a lightweight six-level segmentation model with an integrated width extraction module. Built upon the Swin Transformer architecture, RiWiX employs an efficient encoder-only design with a simple up-sampling module to restore the output mask, thereby improving computational efficiency. A graph-based algorithm further derives continuous water widths by generating a centerline and computing perpendicular distances along the flow direction, effectively generalizing complex river geometries. Experiments on a public water dataset demonstrate that RiWiX achieves competitive accuracy with reduced complexity compared to baseline methods. Tests on geo-referenced imagery yield a mean absolute error of 0.02% for wide rivers (width >500m), 0.06% for moderate rivers (50m < width <70m), and 0.11% for narrow rivers (width <20m), underscoring the robustness and scalability of the proposed framework for river-water-width extraction. The implementation process and code has been made publicly available for researchers to use and extend this work. The full repository is accessible at: https://github.com/FahimaHasanAthina/RiWIX

Department(s)

Computer Science

Document Type

Article

DOI

10.1109/ACCESS.2026.3663480

Keywords

attention mechanism, centerline extraction, deep learning, hierarchical vision transformer, remote sensing imagery, River width extraction, spatial graph modeling, water body segmentation

Publication Date

1-1-2026

Journal Title

IEEE Access

Link to Full Text

Share

COinS