A Novel Remote Sensing Landslide Semantic Segmentation Method: Using CycleGAN-Based Change Detection Algorithms

The study of landslide segmentation using remote sensing images is now focused on change detection and deep learning semantic segmentation algorithm. Deep learning-based semantic segmentation algorithms often require a considerable amount of pixel-level labeled training data. In the study of landslide segmentation research, too high cost of labeling is a barrier to develop deep learning methods. Although the change detection method does not require training data, it is necessary to gather the pre-event and post-event remote sensing images. When change detection is applied to landslide segmentation, it is a big challenge to obtain pre-event remote sensing images. To address this issue, a cycleGAN-based change detection technique for remote sensing landslide semantic segmentation has been developed. First, we used landslide images and non-landslide images to train the cycleGAN model. Then, we applied the cycleGAN approach to produce a non-landslide image from the landslide remote sensing image. Finally, using change detection method, landslide regions are segregated. In addition, we evaluated the suggested technique on the Bijie remote sensing landslide dataset, and got 0.845 accuracy, 0.404 recall, and 0.184 mIoU, demonstrating the method’s viability.