A Learning Framework With Multispectral Band-Differentiated Encoding for Remote Sensing Water Body Detection

• Published in: IEEE journal of selected topics in applied earth observations and remote sensing Vol. 17; pp. 6278 - 6289
• Main Authors: Wei, Debin, Xie, Hongji, Li, Pinru, Xu, Yongqiang
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Coding; Deep learning; Image enhancement; Image segmentation; Indexes; Multispectral remote sensing; negative Laplacian filtering (NLF); Neural networks; Remote sensing; Satellites; semantic segmentation; Spatial resolution; Water; Water bodies
• ISSN: 1939-1404; 2151-1535
• DOI: 10.1109/JSTARS.2024.3399600

Classic deep convolutional neural network (DCNN) models have demonstrated notable efficacy in segmenting remote sensing images. However, their ability to enhance the precision of water body detection, particularly for smaller ones amid intricate backgrounds, remains challenging. This article proposes the negative Laplacian filter (NLF) method as a solution, enhancing regional color contrast during preprocessing to capture more intricate details effectively. Furthermore, a novel approach employs a differential dual-encoding structure that encodes diverse spectra based on their spectral attributes. Lastly, leveraging prior insights from remote sensing, we introduce the weak label weight adjustment operation for refining predicted images in postprocessing stages. The proposed model significantly outperforms the comparison models on our remote sensing water body dataset.