A Deep Learning Method for Mapping Glacial Lakes from the Combined Use of Synthetic-Aperture Radar and Optical Satellite Images

Abstract

Glacial lakes (GLs), a vital link between the hydrosphere and the cryosphere, participate in the local hydrological process, and their interannual dynamic evolution is an objective reflection and an indicator of regional climate change. The complex terrain and climatic conditions in mountainous areas where GLs are located make it difficult to employ conventional remote sensing observation means to obtain stable, accurate, and comprehensive observation data. In view of this situation, this study presents an algorithm with a high generalization ability established by optimizing and improving a deep learning (DL) semantic segmentation network model for extracting GL contours from combined synthetic-aperture radar (SAR) amplitude and multispectral imagery data. The aim is to use the high penetrability and all-weather advantages of SAR to reduce the effects of cloud cover as well as to integrate the multiscale and detail-oriented advantages of multispectral data to facilitate accurate, quantitative extraction of GL contours. The accuracy and reliability of the model and algorithm were examined by employing them to extract the contours of GLs in a large region of south-eastern Tibet from Landsat 8 optical remote sensing images and Sentinel-1A amplitude images. In this study, the contours of a total 8262 GLs in south-eastern Tibet were extracted. These GLs were distributed predominantly at altitudes of 4000–5500 m. Only 17.4% of these GLs were greater than 0.1 km2 in size, while a large number of small GLs made up the majority. Through analysis and validation, the proposed method was found highly capable of distinguishing rivers and lakes and able to effectively reduce the misidentification and extraction of rivers. With the DL model based on combined optical and SAR images, the intersection-over-union (IoU) score increased by 0.0212 (to 0.6207) on the validation set and by 0.038 (to 0.6397) on the prediction set. These validation data sufficiently demonstrate the efficacy of the model and algorithm. The technical means employed in this study as well as the results and data obtained can provide a reference for research and application expansion in related fields.

ACFNet: A Feature Fusion Network for Glacial Lake Extraction Based on Optical and Synthetic Aperture Radar Images

“…The evaluation results of these U-Nets are detailed in Table 1. In addition to the four fusion methods and ACFNet mentioned in Section 3, we also trained and evaluated Wu's model [46] and two other classical semantic segmentation models (SegNet and DeepLabV3+). Similar to ACFNet, Wu's model was proposed to extract glacial lakes based on optical and SAR images.…”

“…This study did not improve the network specially and just used SAR images. Wu et al proposed a model based on U-Net for glacial lake extraction using a combination of Landsat 8 optical images and Sentinel 1 SAR images [46]. Their research showed that the addition of SAR features helps to identify glacial lakes.…”

“…In this study, to effectively use the spectral characteristics of optical images and the geometric characteristics of SAR images to accurately extract glacial lakes, we propose an atrous convolution fusion network (ACFNet) that sufficiently fuses the features of optical and SAR data using atrous convolutions. Furthermore, we compare the performance of ACFNet to the following: four fusion models in which data fusion occurs at input, encoder, decoder, and output; Wu's model [46]; and two typical methods in semantic segmentation (SS), field SegNet [47] and DeepLabV3+ [48]. These models were all trained and evaluated based on glacial lake data distributed throughout the Himalaya (Figure 1).…”

Comparing Methods for Segmenting Supra-Glacial Lakes and Surface Features in the Mount Everest Region of the Himalayas Using Chinese GaoFen-3 SAR Images

“…The quality and quantity of training data are highly important [34,35] In general, the use of satellite images for the monitoring of glacial lakes has been mostly concerned with conventional manual digitization, edge detection, image segmentation methods, and object-oriented classification, rarely adopting some advanced deep-learning methods. Compared with other methods, deep-learning algorithms have a high ability in terms of feature extraction and autonomous learning; they possess a large number of hidden layers and can support a higher level of data abstraction and prediction.…”

“…An attempt has been made recently to map glacial lakes using deep-learning models. Wu et al, 2020 presented an improved U-Net-based, deep-learning, semantic-segmentation network model for extracting the unconnected small glacial lake outlines using a combination of Sentinel-1 ground-range-detected (GRD) SAR images and Landsat-8 multispectral imagery data [34]. Qayyum et al, 2020 explored the potential of PlanetScope optical imagery for the mapping of glacial lakes in the different years over the Hindu Kush, Karakoram, and Himalayan region.…”