Landslide Recognition and Prediction Using Spaceborn Multispectral Data

Landslides are severe environmental hazards in mountains areas. Nowadays, the threat of landslides to public safety has become more pronounced as a result of the burgeoning development in landslide-prone areas, the increase of deforestation in hilly areas, and the increase of regional precipitation caused by global climate change.
Traditional landslide risk assessment requires immense physical power to assemble different in-situ data such as identification of landslide location, land-cover classification and surface geometry. Such traditional data collection technique is very time consuming, and thus impossible to be applied for the regional scale assessment.
Remote sensing techniques, therefore, are the solutions for providing fast and up-to-data landslide assessments. This thesis focuses on the applications of multispectral remotely sensed data for landslide recognition and landslide prediction. Wollongong of Australia was chosen as a test bed for these analyses.

For landslide recognition analysis, three change detection techniques were employed, which were image differencing, bi-temporal linear data transformation and post-classification comparison. For the first two change detection methods, a new landslide identification procedure was developed by integrating surface change information of greenness, brightness and wetness. During the image differencing, the three surface change components were derived from Vegetation Indices (VIs). Four different surface change composites were generated using different VIs; each composite contained three surface change bands which were greenness, brightness and wetness.
For bi-temporal linear data transformation, multitemporal Kauth-Thomas (MKT) transformation was adopted for providing the three types of surface change information.

In the landslide recognition analysis, the best landslide 1 mapping performance is yielded by the image differencing method using brightness and wetness components of Kauth-Thomas transformation and NDVI. Its omission error (i.e. the percentage of actual landslide pixels which were not detected) and commission error (i.e. the percentage of change pixels identified which were not landslide) are 14.4% and 3.3%, respectively, with a strong agreement (KHAT = 88.8%).

For landslide forecast study, a three-layer multistage decision-making structure was developed, whereby the image was segmented based on vegetation status (i.e. Normalised Difference Vegetation Index), surface temperature and surface moisture content (i.e. Normalised Difference Water Index) sequentially. Since the image was acquired half-month prior to the landslides, therefore, excellent prediction results are not expected. Nevertheless, a framework of the improved landslide prediction algorithm was illustrated in details and the experiments indicated a good potential for further exploration.