Effect of time and space partitioning strategies of samples on regional landslide susceptibility modelling

Assessment of the spatial probability of future landslide occurrences for disaster risk reduction is done through landslide susceptibility modelling. In this study, we investigated the effect of time and space partitioning strategies of samples on the performance of regional landslide susceptibility models on macro-scale mapping in the state of Mizoram, India, covering 21,087 km 2 area. We used landslide inventory data of 2014 and 2017 periods consisting of 1205 and 2265 landslides, respectively, to train and test the models with four sampling strategies such as spatial, temporal, temporal (size constrained) and temporal (geographic constrained). We used five commonly inherited models such as multiclass weighted overlay (MCWO), information value (IV), weights of evidence (WoE), logistic regression (LR) and artificial neural network (ANN) to evaluate the effect of sampling strategies on the model performance for regional landslide susceptibility mapping. Validation of model performance was done using receiver operating characteristic (ROC) curve. Traditional spatial sampling strategy applied to landslides in 2014 with a random split in 70:30 proportion provided a high performance of all the five models but failed to predict landslides in 2017. The landslide incidences in 2017, when used for model validation either entirely or in different split conditions (both size and geographic constrained), provided consistent performance, even though the testing sample size is large or have a different spatial disposition, if the training was carried out with non-linear susceptibility models such as LR and ANN using landslide incidences in 2014. Results show the importance of sample selection during validation of landslide susceptibility models on a regional scale.