Rapid landslide detection can give timely information for emergency responses when group-occurring landslides occurred. However, it is frequently difficult to quickly acquire sufficient data for landslide detection in a short period. Transfer learning harnesses the knowledge of landslide detection from the source domain to the target domain with little labeled data. Graph neural networks (GNN) explicitly models global or local relationships by constructing a graph structure where nodes represent pixels and edges represent connections, thereby improving segmentation consistency. Here, we proposed a deep learning model integrated the attention mechanism, multiscale connections, and GNN to capture contextual information and extract the important features for landslide detection. The proposed method was first pretrained in the large-scale dataset, then transferred and fine-tuned the parameters in the two case studies: 2013 Niangniangba rainfall-induced landslides in China and 2018 Hokkaido coseismic landslides in Japan. We examined the feasibility of the proposed model and studied how much impact the scale of the target domain would have on the landslide detection. The controlled experiments reported that our proposed method could achieve the best F1-score in the data-rich condition. Our results also reveal that the deep learning models with transfer learning in data-limited conditions can perform closely to those in data-rich conditions. The fine-tuning model updated parameters in the target domain besides gaining knowledge from the source domain; hence, performance was improved significantly in a new region despite having little new data. Our approach demonstrates a potential way to improve landslide detection assessment, particularly in areas where landslides are extremely difficult to label.
扫码关注我们
求助内容:
应助结果提醒方式:
