Improve unsupervised Learning-based landslides detection by band ratio processing of RGB optical images: a case study on rainfall-induced landslide clusters

Abstract

Automatic detection of the heavy rainfall-induced landslide clusters based on optical remote sensing images had low accuracy due to the interference of many features. This work proposes a method to improve the accuracy of landslide identification based on change detection of pre- and post-disaster satellite images. Firstly, by band ratio preprocessing, the interference features on landslide detection are eliminated. Then, by further optimizing the relevant parameters, the threshold-based Image Difference and Principal Component Analysis-based K-means unsupervised learning classification methods are used to perform change detection and landslide cluster identification. Finally, the causes of missing and errors in landslide detection by using the two methods are analyzed and discussed. The results show that band preprocessing of remote sensing images can significantly weaken the interference features of human activities on landslide detection. The verification metrics of the two methods for detecting landslides, except for a slight decrease in precision, recall, F1-score and Kappa values have all increased significantly. After band ratio preprocessing and parameters optimization, both change detection methods have significantly improved the performance of the rainfall-induced landslide clusters detection, with effects close to those of deep learning methods. These findings can provide a reference for landslide detection using optical remote-sensing images.