下载中心
优秀审稿专家
优秀论文
相关链接
首页 > , Vol. , Issue () : -
摘要

2008年汶川8.0级地震触发了大量的崩塌滑坡地质灾害,导致强震区震后地质灾害频发,因其对生命和财产的巨大威胁而广泛关注,利用遥感等技术快速提取滑坡信息,对于减少灾害造成的损失具有重要的现实意义。本文提出一种迁移学习方法,从自然场景数据集中学习特征,迁移到滑坡提取中。该方法首先在ImageNet上预训练ResNet网络,然后输入滑坡区影像样本,将预训练网络及参数迁移至LinkNet上,最终实现滑坡提取。通过对2013-2015年三景影像的汶川地震震后滑坡提取实验进行分析及验证,结果显示,相较于传统支持向量机和其他深度学习方法,本文提出的迁移学习方法有较优的提取精度,有利于后续研判及决策。
Objective: Landslides, as natural disasters, result from various factors. They usually come together with catastrophic damages and casualties. A big earthquake can trigger plenty of landslides. Therefore, landslide extraction is important to provide information timely for after-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 visual interpretation of aerial photographs or high-resolution remote sensing images, involving vast manpower. There are several kinds of landslide extraction methods, such as pixel-based methods, which have relatively low accuracy; object-oriented methods, whose parameters need to be decided subjectively. With deep learning developing in image semantic segmentation, precise and automatic remote sensing image binary classification becomes possible. Many researchers study the use of deep learning for landslide extraction in different areas. However, relatively small amount of landslide data can easily cause overfitting of the model. Transfer learning—transferring knowledge from the source domain to the target domain—can moderate this problem. It utilizes knowledge in the source domain to improve performance in the target task. A transfer learning deep network is designed to improve the accuracy of landslide extraction. Method: First, three GF-1 images from 2013 to 2015 in the research area are processed successively by geometric correction, registration and image fusion, to obtain 2 meters resolution 4 bands (red, green, blue and near-infrared) images. Then, a proper network is designed. The encoder of ResNet which is trained on ImageNet is chosen as our encoder. The decoder of LinkNet is then chosen as our decoder whose residual structure and bypass links improve performance. Besides, bypass links in decoder remedy spatial information lost in max-pooling in the encoder while residual structure enables network to well learn complex features. After pre-training ResNet network on ImageNet, we adjust the number of input channels of first convolution layer to 4 and drop the last fully connected layer, then form our network with decoder. Finally, we input remote sensing landslide images to fine-tune our model. Result: When testing different network depths. The network doesn’t always perform better while depth increases. There is relevantly peak performance in ResNet50, so we choose it as our encoder. Then, we compare our method to SVM, our network without a pre-training encoder, improved U-Net and a mainstream transfer learning method AlbuNet. It shows that deep learning methods perform better than SVM, while transfer learning methods perform better than deep learning methods trained on landslide images. On average, the proposed method is 17.16%, 18.58% and 17.4% higher in precision, recall, and F1 measure than SVM; 2.98%, 6.35%, 4.61% higher in precision, recall and F1 measure than improved U-Net; and 0.9%, 1.98%, 1.48% higher in precision, recall and F1 measure than the AlbuNet. Conclusion: It is recommended to select encoder of ResNet50 combined with decoder of LinkNet to form a landslide extraction network with higher accuracy than other ResNet encoders of different depths and transfer learning network AlbuNet. Transferring knowledge learned from ImageNet can also improve the performance of the landslide extraction deep learning network. The proposed method is convenient for follow-up landslide risk assessment, disaster investigation, disaster warning and decision-making. Keywords: landslide extraction, transfer learning, ImageNet, GF-1