Glacial lake extraction is essential for studying the response of glacial lakes to climate change and assessing the risks of glacial lake outburst floods. Most methods for glacial lake extraction are based on either optical images or synthetic aperture radar (SAR) images. Although deep learning methods can extract features of optical and SAR images well, efficiently fusing two modality features for glacial lake extraction with high accuracy is challenging. In this study, to make full use of the spectral characteristics of optical images and the geometric characteristics of SAR images, we propose an atrous convolution fusion network (ACFNet) to extract glacial lakes based on Landsat 8 optical images and Sentinel-1 SAR images. ACFNet adequately fuses high-level features of optical and SAR data in different receptive fields using atrous convolution. Compared with four fusion models in which data fusion occurs at the input, encoder, decoder, and output stages, two classical semantic segmentation models (SegNet and DeepLabV3+), and a recently proposed model based on U-Net, our model achieves the best results with an intersection-over-union of 0.8278. The experiments show that fully extracting the characteristics of optical and SAR data and appropriately fusing them are vital steps in a network’s performance of glacial lake extraction.
Optical images-based edge detection in Synthetic Aperture Radar images
