Study Of Landslide Hazards Evaluation Based On Multi-source Remote Sensing Data

Longmen Mts. faults belt is a dangerous geological hazard area in China. It is just the place that had taken place the Ms 8.0 Wenchuan earthquake in 12 May 2008 and the epicenter is located at Yingxiu town, Wenchuan county. This paper focuses on landslide hazards caused by Wenchuan earthquake in a 48-square-meters study area in Weizhou town, Wenchuan County.Firstly, landslide inventory data and susceptibility factors extracted from remote sensing data are become assessment material. Before thinking better of the real condition in study area, this paper chooses eight kinds of susceptibility factors as thematic data, such as elevation, slope, aspect, vegetation cover, distance from road, distance form river, lithology and land use. Practicably, identified landslide inventory thematic data from 0.5m resolution airborne remote sensing imagery; extracted elevation, slope, aspect and river net from DEM data; computed vegetation cover; classified land use and distinguished road form 2.5m resolution SPOT-5 fusion imagery; using ASTER data to detect the earth's surface lithology,which needs object-based multi-resolution segmentation and analyses at different segmentation level.Secondly, susceptibility factors thematic data can be formatted as the basis input data for logistic regression model and support vector machine (SVM) model to predict hazard probability. The study result shows that logistic regression model can reflect the samples' characters well, and the SVM model owns nice predicating ability. Therefore, a new model named logistic regression-weighted support vector machine (LR-WSVM) is used to establish a hazards evaluation system and lastly to acquire landslide hazards evaluation result.Finally, this study employ ROC curve to compare the predicting ability of these models. Through the compared results we can conclude that the area under the ROC curve provided by LR-WSVM obviously better than the others. Besides, it is also classify every susceptibility factors at given grades and evaluate these landslide contribution ratios. Then choosing the maximum value in grades as the marked representation of susceptibility factors, and according these to compare the landslide influence degrees. We can find that slope and vegetation cover are two major susceptibility factors in this study.In a word, this paper acquires susceptibility factors thematic data from remote sensing data, which provides a new thinking for related thematic data extraction. The LR-WSVM method developed by this paper evidently improves landslide prediction accuracy, and presents real utility worthiness.