Geoinformation-based landslide recognition using deep learning in subtropical regions: a case study of Ruijin City

Conducting regional landslide recognition is of great significance for the timely prevention and treatment of landslide disasters in subtropical areas. The purpose of this study is to recognize landslides based on the U-Net model by establishing landslide recognition indicators, taking Ruijin City as an example. The results indicate a negative correlation between entropy and the second moment in the recognition features of landslides and terraces; based on this relationship, the Landslide and Terrace Distinction Index (LTDI) is established. Through random parameter experiments, the optimal training parameters for the landslide recognition model were determined, which significantly improved the model’s recognition performance. Combined with optical remote sensing images and field verification, a total of 221 historical landslide areas were recognized by the landslide recognition model with the optimal parameters in Ruijin City. The F1 of the U-Net model integrated with LTDI reached 0.91, representing an increase of 0.13 compared to the model without LTDI, effectively reducing FP in terrace scenarios and enhancing the detection sensitivity for small-scale landslides. This study can provide an operational reference for landslide recognition in subtropical regions based on remote sensing images and deep learning techniques.