ENSEMBLE LEARNING-BASED AUTOMATIC DETECTION OF LANDSLIDE AREAS FROM AERIAL PHOTOGRAPHS

Abstract

Landslides pose a significant threat to human life and property worldwide. Japan, with its vulnerability to these natural disasters, records a high incidence of landslides. The Geospatial Information Authority of Japan employs experts to visually examine aerial photographs before and after landslide events, a costly and time-consuming approach that can limit accuracy. This study aims to aid in mitigating the damage caused by landslides through accurate and efficient mapping and prediction. An Ensemble U-Net model integrating three U-Nets has been proposed to predict landslide areas from aerial photographs. Comparative analysis with a single U-Net model revealed that the Ensemble model significantly outperformed the single model in all accuracy measures, including precision, recall, and F1-score. The ensemble model's average intersection over union (IoU) value of 0.80 also indicated a stronger agreement between the predicted outcome and ground truth than the single U-Net model. Visual analysis of prediction results further demonstrated the superiority of the ensemble model in aligning closely with the ground truth, thereby reducing misidentification and missed detections. The proposed Ensemble U-Net model's potential to enhance the accuracy and efficiency of landslide mapping seems promising.