Evaluation of Trans-Regional Landslide Susceptibility of Reservoir Bank Based on Transfer Component Analysis

It is crucial to create a "migratable" landslide susceptibility model considering the time-consuming process of recording landslides. However, a sample set of all known landslide data areas must be used in forecasting the susceptibility of unsampled regions adequately. In this paper, we attempt to develop a TCA-CNN model to enable trans-regional landslide susceptibility evaluation, which is based on the adaption domain of transfer learning by integrating the transfer component analysis (TCA) with deep learning convolutional neural network (CNN). The spatial database of landslides is constructed by extracting 11 environmental factors of the reservoir bank area, then the Mianhuatan reservoir area without samples is then predicted using the susceptibility model from the Chitan reservoir area with samples. The results show that: (1) The maximum mean discrepancy (MMD) of data from different study areas treated by TCA decreased significantly (0.022), and the data is approximately identically distributed; (2) The trans-regional prediction accuracy by the TCA-CNN model is 0.854, which is higher than that of the CNN model (0.791). And it can be verified that the proportion of landslide frequency falling into the high/extremely-high susceptibility interval is the highest (89.1%) in the historical landslide; (3) The area under the receiver operating characteristic (ROC) curve of the TCA-CNN model is 0.93, which is higher than that of the CNN model (0.90). It's obvious that the TCA-CNN model can effectively use the samples data of the modeling area to realize the susceptibility evaluation of the unsampled area. Compared with the traditional machine model, TCA-CNN model has higher and more stable prediction accuracy and stronger generalization ability in cross-region prediction.