Consideration of optimal pixel resolution in deriving landslide susceptibility zoning within the Sydney Basin



Publication Details

Palamakumbure, D., Flentje, P. & Stirling, D. (2015). Consideration of optimal pixel resolution in deriving landslide susceptibility zoning within the Sydney Basin. Computers and Geosciences, 82 13-22.


This paper discusses the progress of the landslide susceptibility mapping in the wider Sydney Basin area to facilitate engineering and geological studies and land-use zoning; using induced decision trees. This study investigates the effect of the basic unit of this spatial modelling work (pixel resolution) on the accuracy of the modelling outcome, and reports on the effectiveness of using See5 pruned decision trees to model the landslide susceptibility of the Sydney Basin. Landslide susceptibility was determined from the landslide confidence value derived from the Laplace ratio of the rule based predicted classes. The modelling work has been carried out at 2 m, 5 m, 10 m, 15 m, 20 m, 25 m, 30 m, and 40 m pixel resolutions for a trial area within the Sydney Basin. Ten different GIS based datasets derived from the same original datasets have been used each time as landslide causative factors. The optimum tree pruning parameters for each pixel resolution were identified by analysing the behaviour of misclassification errors. Performance of the models at different pixel resolutions was compared using ROC curves and five-fold cross validation accuracy. High density ALS elevation point clouds and large scale datasets allowed model development at a higher resolution (2 m) but the decision tree model at 10 m resolution performed better than the rest. The ratio between the square root of the mean landslide area of the inventory and the area covered by a single pixel has been developed as a worthwhile quantitative measurement of the adequacy of the model resolution. The validation results of the final modelling outcome show that landslide susceptibility descriptors fulfil the requirements of the LRM guidelines. The model has a conservative success of 90% according to the field validation and a cross validation accuracy of 92%.

Please refer to publisher version or contact your library.