Forested landslide detection using LiDAR data and the random forest algorithm: A case study of the Three Gorges, China

• Published in: Remote sensing of environment Vol. 152; pp. 291 - 301
• Main Authors: Chen, Weitao, Li, Xianju, Wang, Yanxin, Chen, Gang, Liu, Shengwei
• Format: Journal Article
• Language: English
• Published: New York, NY Elsevier Inc 01.09.2014; Elsevier
• Subjects: Accuracy; Algorithms; Animal, plant and microbial ecology; Applied geophysics; autocorrelation; Biological and medical sciences; case studies; China; Classification; Earth sciences; Earth, ocean, space; Exact sciences and technology; Feature selection; Fundamental and applied biological sciences. Psychology; General aspects. Techniques; Internal geophysics; Landslide mapping; Landslides; LiDAR; Random forest; remote sensing; Slopes; Surface layer; surveys; synthetic aperture radar; Teledetection and vegetation maps; Texture; The Three Gorges; Topographic analysis; Training; vegetation cover; Far East; Asia; China; China, People's Rep., Sichuan Prov., Changjiang R., Three Gorges Dam; China, People's Rep; Feature selection; Topographic analysis; Random forest; The Three Gorges; Landslide mapping; LiDAR; news; algorithms; detection; mountains; surveys; Shadow; remote sensing; vegetation; Extract; dams; forests; Random graph; techniques; landslides; digital terrain models; Lidar; Plant cover; Feasibility; Synthetic aperture radar
• ISSN: 0034-4257; 1879-0704
• DOI: 10.1016/j.rse.2014.07.004

The Three Gorges region of central western China is one of the most landslide-prone regions in the world. However, landslide detection based on field surveys and optical remote sensing and synthetic aperture radar (SAR) techniques remains difficult owing to the dense vegetation cover and mountain shadow. In the present study, an area of Zigui County in the Three Gorges region was selected to test the feasibility of detecting landslides by employing novel features extracted from a LiDAR-derived DTM. Additionally, two small sites—Site 1 and Site 2—were selected for training and were used to classify each other. In addition to the aspect, DTM, and slope images, the following feature sets were proposed to improve the accuracy of landslide detection: (1) the mean aspect, DTM, and slope textures based on four texture directions; (2) aspect, DTM, and slope textures based on aspect; and (3) the moving average and standard deviation (stdev) filter of aspect, DTM, and slope. By combining a feature selection method and the RF algorithm, the classification accuracy was evaluated and landslide boundaries were determined. The results can be summarized as follows. (1) The feature selection method demonstrated that the proposed features provided information useful for effective landslide identification. (2) Feature selection achieved an improvement of about 0.44% in the overall classification accuracy, with the feature set reduced by 74%, from 39 to 10; this can speed up the training of the RF model. (3) When fifty randomly selected 20% of landslide pixels (PLS) and 20% of non-landslide pixels (PNLS) (i.e., 20% of PLS and PNLS) were utilized in addition to the selected feature subsets for training, the test sets (i.e., the remaining 80% of PLS and PNLS) yielded an average overall classification accuracy of 78.24%. The cross training and classification for Site 1 and Site 2 provided overall classification accuracies of 62.65% and 64.50%, respectively. This shows that the random sampling design (which suffered some of the effects of spatial auto-correlation) and the proposed method in this present study contribute jointly to the classification accuracy. (4) Using the Canny operator to delineate landslide boundaries based on the classification results of PLS and PNLS, we obtained results consistent with the referenced landslide inventory maps. Thus, the proposed procedure, which combines LiDAR data, a feature selection method, and the RF algorithm, can identify forested landslides effectively in the Three Gorges region. •The aspect, DTM, and slope textures based on aspect direction are newly introduced.•Feature selection enhances the classification accuracy and reduces the feature set.•Forested landslides can be detected using a LiDAR DTM and supervised classification.