IMPROVEMENT OF THE METHODS OF DETERMINING WATER AREAS ON SPECIFIC IMAGES

Abstract

This article explores the application of semantic segmentation using convolutional neural networks (CNNs) for the detection of water bodies in high-resolution multispectral images. The inclusion of near-infrared (NIR) channels enhances classification accuracy, enabling more precise differentiation between visually similar features, such as water bodies and cloud shadows. The study employs a U-Net-based architecture, trained on aerial images, to effectively identify water regions in satellite images. Additionally, it addresses challenges associated with misclassified areas caused by clouds and their shadows, recommending the Normalized Difference Water Index (NDWI) as a reliable metric for validating segmentation results. For practical applications, such as calculating the area of water surfaces, the article proposes the use of a Binarized Normalized Difference Water Index (BNDWI), which provides a straightforward and effective binary classification approach.