Transfer learning method for landslide extraction from GF-1 images after the Wenchuan earthquake

Landslides are natural disasters that are driven by various factors and often leave catastrophic damages and casualties. A huge earthquake can trigger plenty of landslides. Therefore, landslide extraction is critical to provide timely information for post-disaster decision making. Remote sensing is a convenient tool for landslide information acquisition. However, landslide features are so intricate that landslide extraction mainly relies on a visual interpretation of aerial photographs or high-resolution remote sensing images, which requires vast manpower. Several landslide extraction methods are available today, including pixel-based methods, which have relatively low accuracy, and object-oriented methods, whose parameters need to be decided subjectively. With the continuous development of deep learning in image semantic segmentation, a precise and automatic remote sensing image binary classification becomes possible. Many researchers have investigated the use of deep learning for landslide extraction in different areas. However, a relatively small amount of landslide data can easily lead to model overfitting. Transfer learning, where knowledge is transferred from the source domain to the target domain, can alleviate this problem by using knowledge in the source domain to improve performance in the target task. A transfer learning deep network is then designed to improve the accuracy of landslide extraction. First, three GF-1 images taken from 2013 to 2015 in the research area were processed successively by geometric correction, registration, and image fusion to obtain 4 bands of images (red, green, blue, and near-infrared) with a resolution of 2 m. Second, a proper network was designed. The encoder of ResNet that was trained on ImageNet was chosen as the encoder, and the decoder of LinkNet, whose residual structure and bypass links can improve performance, was selected as the decoder. The bypass links in the decoder can also address the spatial information loss in max-pooling in the encoder, and the residual structure allows the network to learn complex features. After pre-training the ResNet network on ImageNet, we adjusted the number of input channels of the first convolution layer to 4, drop the last fully connected layer, and then form our network with the decoder. We eventually inputted remote sensing landslide images to finetune our model. When testing different network depths, the network does not always perform better as the depth increases. We chose ResNet50 as our encoder given its peak performance. Afterward, we compared our method with SVM. Without a pre-training encoder, our network improves U-Net and a mainstream transfer learning method, AlbuNet, thereby suggesting that deep learning methods outperform SVM, whereas transfer learning methods outperform deep learning methods that are trained on landslide images. The proposed method outperforms SVM in terms of precision, recall, and F1 measure by 17.16%, 18.58%, and 17.4%, respectively, outperforms the improved U-Net by 2.98%, 6.35%, and 4.61%, and outperforms AlbuNet by 0.9%, 1.98%, and 1.48%. The ResNet50 encoder combined with the LinkNet decoder should be selected to form a landslide extraction network with a higher accuracy compared with the transfer learning network AlbuNet and other ResNet encoders of different depths. Transferring knowledge learned from ImageNet can also improve the performance of the landslide extraction deep learning network. The proposed method can be used conveniently for follow-up landslide risk assessment, disaster investigation, disaster warning, and decision making. © 2023 National Remote Sensing Bulletin