Landslide susceptibility mapping using multiscale sampling strategy and convolutional neural network: A case study in Jiuzhaigou region

Landslides are one of the most widespread natural disasters and pose severe threats to people, properties, and the environment in many areas. Landslide susceptibility mapping (LSM) has proven useful in designing landslide mitigation strategies for reducing disaster risk and societal and economic losses, which are essential for land use planning, hazard prevention, and risk management. Recent efforts for improving accuracies of LSM have focused on the utilization of convolutional neural network (CNN) in some image-related tasks, however, due to the inconsistency of data representation, CNN-related studies need to be further explored. In this study, a CNN-based approach for LSM was proposed and experimentally applied in a Jiuzhaigou region where a catastrophic earthquake taken place on 8 August 2017, in Sichuan, China. To address the issue of data representation in the CNN model, we proposed a multiscale sampling strategy which to our knowledge is novel in LSM. In this way, the multiscale training samples (i.e., small scale, medium scale and large scale) were generated from the selected eleven landslide causative factors. The success-rate curve (SRC) and prediction-rate curve (PRC) were applied to validate the LSM results, and three conventional machine learning algorithms, i.e., logistic regression, multi-layer perceptron (MLP) neural network and radial basis function (RBF) neural network, were attempted for comparison. Specifically, MLP neural network achieved the best performance among three machine learning methods, with the area under the SRC (AU-SRC) and PRC (AU-PRC) values of 81.18% and 82.84%, respectively. Nevertheless, the AU-SRC and AU-PRC values of CNN-based approach reached to 97.45% and 88.02%, which were about 16% and 6% higher than that of the MLP neural network, respectively. The present study demonstrated both the excellent goodness-of-fit and strong prediction ability of CNN-based approach for LSM, which also showed the effectiveness and feasibility of the proposed multiscale sampling strategy. Additionally, present study revealed that the spatial data close to the landslide location might be more suitable to predict the probability of the landslide occurrence. Finally, we expect that the deep learning method based on multiscale data representation will advance our ability to assess the landslide susceptibility and raise the awareness of landslide disasters.