NAU-Net: A New Deep Learning Framework in Glacial Lake Detection

Abstract

Glacial lake mapping is essential for understanding the response of glacial lakes to climate change and the risk assessment of glacial lake outburst floods (GLOFs). Given that glacial lakes have little area compared with background objects, extracting glacial lakes high precisely is still challenging. Recently, U-Net, a deep learning (DL) method, has shown great potential in glacial lake extraction, due to its elaborate encoder-decoder structure and powerful skip connections. However, the skip connections transmit a lot of information irrelevant to glacial lakes from the low-level appearance features to the high-level semantic features, leading to inefficient utilization of the low-level features. In this letter, we propose a normalized difference water index (NDWI) attention U-Net (NAU-Net) for pixel-wise glacial lake segmentation, which utilizes NDWI as a spatial attention to highlight the signals of water regions in low-level feature maps, thereby making the network pay more attention to glacial lakes. Compared with the classical networks and the state-of-the-art non-DL methods, our NAU-Net shows better performance on glacial lake extraction. The source code can be found at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/JinxiaoWang/NAU-Net</uri> .