Landslide Inventory Mapping With Bitemporal Aerial Remote Sensing Images Based on the Dual-Path Fully Convolutional Network

This article presents a novel dual-path full convolutional network (DP-FCN) model for constructing a landslide inventory map (LIM) with bitemporal very high-resolution (VHR) remote sensing images. Unlike traditional methods for drawing LIM, the proposed DP-FCN directly draws LIMs from the bitemporal aerial images with VHR through a trained deep neural network without generating the change magnitude map. Thus, the proposed approach can effectively reduce the effects of pseudo changes caused by phenological differences rather than landslide events. The proposed DP-FCN model contains two modules, namely, deep feature extraction, and joint feature learning networks. Deep feature extraction aims to reduce redundancy while extracting the high-level deep features from bitemporal images. Joint feature learning establishes the relationship between the deep features of bitemporal images and the ground reference map. Experiments on the real datasets of the landslide sites in Lantau Island of Hong Kong, China, demonstrate the feasibility and superiority of the proposed approach in drawing LIM with VHR remote sensing images. Moreover, compared with the results obtained by the state-of-the-art algorithms, the proposed DP-FCN method achieves the best performance in terms of accuracy for landslide inventory mapping.