Date of Graduation
Fall 2025
Degree
Master of Science in Computer Science
Department
Computer Science
Committee Chair
Razib Iqbal
Abstract
The increasing availability of high-resolution remote sensing data has enabled new opportunities for large-scale river studies, reducing dependence on manual field-based width measurements. While existing automated methods, ranging from thresholding to deep learning approaches such as CNNs, DeepLabV3+, and U-Net, have shown reasonable performance for wide rivers (Width>30m), they remain less effective for narrow rivers due to limited semantic representation and high computational demands. Recently, vision transformers have emerged as powerful models, offering improved semantic understanding by capturing long-range dependencies through self-attention. However, their adoption in remote sensing remains limited due to the high processing cost. In this thesis, I introduce the ‘River Water Width Extraction’ (RiWiX), a lightweight, encoder-only river water surface segmentation model integrated with a novel width extraction module. The model employs a hierarchical encoder that captures multi-scale contextual features through patch-wise self-attention. At the same time, a shallow up-sampling module efficiently restores spatial resolution without relying on a heavy transformer decoder. The accompanying width extraction algorithm further constructs a graph-based river centerline. It computes perpendicular distances along the flow direction, enabling precise, continuous width estimation even in complex river geometries. Experimental evaluation shows that RiWiX surpasses the baseline transformers both in accuracy and efficiency, achieving a dice score of 0.60 with the lowest computational cost of 2830.85 GFlops. Validation on georeferenced data further confirms its consistent performance across varying river scales, yielding a mean absolute error of 0.02% for wide rivers (width>500m), 0.06% for moderate rivers (50
Keywords
river width extraction, remote sensing, Vision Transformer, semantic segmentation, geospatial image processing
Subject Categories
Remote Sensing
Copyright
© Fahima Hasan Athina
Recommended Citation
Athina, Fahima Hasan, "RiWiX: Towards Automatic River Water Width Extraction From High Resolution Satellite Images Using Swin Transformer" (2025). Graduate Theses/Dissertations. 4116.
https://bearworks.missouristate.edu/theses/4116
