LWAMNNet: A novel deep learning framework for surface water body extraction from LISS-III satellite images

• Published in: Earth science informatics Vol. 17; no. 1; pp. 561 - 592
• Main Authors: Nagaraj, R, Kumar, Lakshmi Sutha
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.02.2024; Springer Nature B.V
• Subjects: Complexity; Computation; Deep learning; Earth and Environmental Science; Earth Sciences; Earth System Sciences; Encoders-Decoders; Feature extraction; Fresh water; Hydrologic cycle; Information Systems Applications (incl.Internet); Machine learning; Modules; Neural networks; Ontology; Performance measurement; Rainfall; Remote sensing; Satellite imagery; Simulation and Modeling; Space Exploration and Astronautics; Space Sciences (including Extraterrestrial Physics; Surface water; Water availability; Water bodies; Water resources management; Surface water dynamics; LISS-III Images; Lightweight DL model; Water body extraction
• ISSN: 1865-0473; 1865-0481
• DOI: 10.1007/s12145-023-01187-1

Fresh water is vital for all living creatures and maintains the hydrological cycle. Surface water bodies conserve freshwater and exhibit dynamic changes yearly due to high/low rainfall and over/underutilization. Therefore, extracting water bodies and determining their extent is imperative for effective water resource management. Water body extraction using Water Indices (WI) and Machine Learning (ML) face threshold selection and feature optimization challenges, respectively. This paper proposes a Lightweight Attention-based Multiscale Neural Network (LWAMNNet) for surface water body extraction from Linear Imaging Self Scanning‒III (LISS-III) remote sensing images. The LWAMNNet is an encoder-decoder architecture designed using a modified residual block in both the encoder and decoder to extract high-level features. Feature Extraction Module (FEM) is sandwiched between encoder and decoder to extract global contextual features. The LWAMNNet replaces convolutions with depthwise separable convolutions to reduce computation complexity. In the decoder, the attention module is incorporated to provide attention to fused features (i.e., combined deep features with spatial encoder features) at different scales. The LWAMNNet effectively extracts different-sized water bodies with non-linear boundaries. The proposed LWAMNNet qualitatively and quantitatively outperforms other DL models in performance metrics (accuracy of 99.5%) and computation complexity (in terms of trainable parameters and time). Additionally, the water extent of five major reservoirs in south India was determined annually from 2016 to 2019. Also, the reason for water dynamics is analyzed with the help of rainfall and water availability data provided by the Indian Metrological Department (IMD).