Landslide susceptibility mapping using CNN models based on factor visualization and transfer learning

Landslide susceptibility mapping (LSM) is an important means of preventing landslides. In recent years, the utilization of deep neural networks, specifically convolutional neural networks (CNN), has gained traction in LSM research. However, the structures of CNN employed in these studies are often shallow, limiting the effective utilization of CNN’s advantages. To enhance the accuracy of CNN-based landslide susceptibility analysis, this study introduces two novel factor visualization methods: Gramian Angular Summation Fields + (GASF +) and Binary Code. Furthermore, various efficient CNN models employed in object detection are incorporated into landslide susceptibility analysis through transfer learning. The study area chosen for investigation is Wanxian County, located in Chongqing, China. The findings indicate that transfer learning effectively improves model training speed and stability, with the combination of GASF + and ResNet yielding high accuracy in landslide susceptibility analysis, surpassing the traditional 2D Matrix-based CNN model by approximately 5.2%. Additionally, the study innovatively analyzes factor importance using a combination of Gram-Cam and image encoding, thereby enhancing model interpretability. The results indicate that human engineering activities significantly influence landslides in the study area. This study significantly expands the depth of LSM research based on CNN and provides new avenues for future investigations in this field.