Recognition and Mapping of Landslide Using a Fully Convolutional DenseNet and Influencing Factors

The recognition and mapping of landslide (RML) is an important task in hazard and risk research and can provide a scientific basis for the prevention and control of landslide disasters. However, traditional RML methods are inefficient, costly, and not intuitive. With the rapid development of computer vision, methods based on convolutional neural networks have attracted great attention due to their numerous advantages. However, problems such as insufficient feature extraction, excessive parameters, and slow model testing have restricted the development of this technology. This research proposes a new RML framework based on a new semantic segmentation network termed the fully convolutional DenseNet (FC-DenseNet). In this network, the features extracted from each layer are repeatedly used in a dense connection, and the parameters are controlled by a bottle-neck structure. Meanwhile, the structure of the encoder-decoder solves the problem of the slowness of model testing. Finally, the landslide influencing factors are added, which enriches the training data. To verify the effectiveness of the proposed method, we focused on several deep networks for comparison and analysis. The results show that FC-DenseNet can better recognize the boundary and interior of landslides, and there are fewer missing and excessive recognition results. The kappa value of the new method is 94.72% in Site 1, which is 6% and 4% higher than that of U-Net and ResU-Net, respectively, and 94.56% in Site 2, which is 6% and 3% higher than that of U-Net and ResU-Net, respectively, indicating that FC-DenseNet has great potential in RML applications.