A semi-automatic landslide detection model combining spatial statistical analysis and change detection

Landslide inventory mapping is a critical component of landslide susceptibility analysis and prediction. The mapping process has been carried out based on field surveys and comparisons of aerial or satellite imagery, which are both time-consuming and labor-intensive. Therefore, recent studies have utilized artificial intelligence models to identify landslide locations. However, the accuracy of these approaches remains limited due to dense vegetation, the low spectral resolution, and seasonal spectral variations in forested regions. Consequently, there have been efforts to enhance the accuracy of landslide inventory mapping through the integration of landslide conditioning factors.

The objective of this study is to enhance landslide detection through the utilization of Sentinel-2 satellite imagery prior to and following landslide occurrences, in conjunction with landslide conditioning factors. The analysis is divided into two phases: a change detection phase and a post-processing phase. In the change detection phase, Sentinel-2 L2A images from before and after landslide events were analyzed using a multi-layer perceptron model, with changes in NDVI and surface reflectance across bands 2 to 12. In the post-processing phase, the frequency ratio technique was applied to calculate the conditioning factor grades. These grades were then used to weight the result of the change detection phase. The conditioning factors encompassed effective soil depth, timber age, elevation, slope, geological lithology, and land cover. To validate and compare the results, the area under the curve (AUC) was computed based on receiver operating characteristic (ROC) curves. The model's training and validation were carried out using data from Jecheon-si, a region that experienced a high incidence of landslides in 2020. In addition, the model's performance was evaluated using the data from the study area. The proposed integrated approach integrates change detection using satellite imagery with landslide conditioning factors to enhance the accuracy of landslide detection models. The proposed model is expected to contribute to the enhancement of landslide hazard management and prevention by providing more reliable detection techniques.

 

Acknowledgement

This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government(MSIT) (RS-2023-00222563)