Title: Automatic extraction method of water body boundaries in remote sensing images based on deep residual network

Authors: Qianchen Yang

Addresses: College of Computer and Information Engineering, Guangxi Vocational Normal University, Nanning, Guangxi, 530007, China

Abstract: Water body extraction, which classifies each pixel in an image as either water or background, is a fundamental task in land - cover categorisation. Accurate identification of water bodies is critical for urban hydrology applications such as water resource management and flood warning systems. While traditional index-based methods like NDWI and MNDWI have been widely used, deep convolutional neural networks (DCNNs) have recently shown promising improvements. However, training these networks requires large volumes of high-quality labelled data, which is often limited in remote sensing. PAN improves performance across various models and datasets, yielding an average increase of 0.749 in mean Intersection over Union (m Io U). Moreover, PAN surpasses previous benchmarks without altering model architecture and demonstrates the added value of incorporating multispectral data.

Keywords: patch adaptive network; PAN; StyleGAN2; Dandelion optimisation; DO; RGB/NIR; binary cross entropy; BCE.

DOI: 10.1504/IJEE.2025.151271

International Journal of Environmental Engineering, 2025 Vol.13 No.4, pp.328 - 347

Received: 21 Jun 2025
Accepted: 07 Aug 2025
Published online: 20 Jan 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article