Assessment of landslides susceptibility
Legorreta Paulin, Gabriel
MetadataShow full item record
Shallow landslides or slope failures have been studied from several points of view. In particular, numerous methods embedded in Geographic Information Systems (GIS) applications have been developed to assess slope stability. However, little work has been done on the systematic comparison of different techniques to outline advantages and limitations, due to the difficulties of assessing a model in both natural and theoretical conditions. The lack of study of the systematic comparison of landslide models at different cartographic scales, DEM resolutions, and sampling strategies not only compromises the reliability of the models, but also leads to the misuse of the models. In this research, assessment of landslide susceptibility models is modeled by using LOGISNET, an acronym for Logistic Regression, Geographic Information System, and Neural Network. LOGISNET is a project of which the main purpose is to provide government planners and decision makers with a tool and guidance to assess landslide susceptibility models. The system is fully operational for models employing a cartographic/hydrologic approach (SINMAP) and multiple logistic regression (MLR). The models are enhanced by algorithms that allow the user to include vertical variation of geotechnical properties in multiple soil layers. The algorithms--global per region, per point, and per layer--are based on cartographic abstraction rather than mechanical meaning. But they are in concordance with the different levels of data storage and retrieval of topographic, hydrologic, and soil variables for the models. The models are assessed and compared on theoretical, artificial and natural conditions. Theoretical artificial conditions are based on artificial DEMs whose surface is smoothed and inscribed to create stable and unstable conditions. Natural conditions are based on two study areas: the Highway 101 corridor in Del Norte County, California and Mt. Diablo State Park in Contra Costa County, California, USA. The result shows that the enhanced implementations of SINMAP by using the LOGISNET algorithms match the result obtained with SINMAP in some cases. The landslide prediction by the LOGISNET algorithm is unstable. Depending on extreme values, the LOGISNET algorithm results are biased toward better or worse predictions than SINMAP. Overall, SINMAP and its enhanced version in LOGISNET show low to moderate match with the landslide inventory map. The variable percentages of model versus inventory match depend on geotechnical parameters, DEM resolution, cartographic expression of the landslide, and the type of landslide. In theoretical and controlled conditions, SINMAP and its enhanced version in LOGISNET can match landslide areas in the inventory map between 15.30% and 70.73%. But in natural conditions, concomitants of cartographic and environmental problems such as DEM resolution, DEM interpolation, cartographic size of the landslide, mixed landslide process, and types of landslides compromise the efficiency of the model. In natural conditions, SINMAP and its enhanced version in LOGISNET are found to match the inventory map in only 0.54% to 18.72% of the cases. Given the knowledge of low performance of SINMAP and difficulties in cost and time needed to obtain local geotechnical parameters, this thesis questions the use of SINMAP and its enhanced version in LOGISNET as a successful model for landslide prediction in natural conditions. MLR is found to match the inventory map from 27.70% to 78.19% of the cases tested under theoretical and controlled conditions. However, this percentage decreases to from 23.72% to 48.62% in natural conditions. Although MLR tends to overpredict landslides and depends greatly on DEM resolution and sampling strategy, MLR is a more flexible technique to predict landslides. For instance, the MLR approach does not depend on the type of landslide as does SINMAP. Given the restriction to thirteen geotechnical variables imposed on MLR, MLR uses only two to four variables. Although the two to four variables may not be the optimum for the analysis, MLR obtains similar or better results than the SINMAP approach.