Comparing data-driven landslide susceptibility mapping methods and the impact of modifying the study area and sampling method in the Hunza Valley (Pakistan)

We compared artificial neural networks (ANN), frequency ratio (FR), logistic regression (LR), and weights of evidence (WoE) for landslide susceptibility mapping in the Hunza Valley, Pakistan. We applied different data manipulation techniques (e.g. feature masking and sampling) to analyze their effects on the model. The landslide inventory was collected using Google Earth satellite images. The corresponding potential causative factors were derived from a geological map, a digital elevation model, and satellite imagery data. We evaluated the models with receiver operating characteristics curves using cross validation. Using the validation data, ANN showed the best performance, followed by LR, WoE, and FR. All applied procedures achieved good and comprehensible results. However, the susceptibility patterns show substantial differences. Modifying the study area (e.g. excluding trivial areas, such as glaciers) and using different sampling strategies significantly impacts the susceptibility patterns in all models. We recommend the use of WoE and FR in large areas with few causative factors, despite their lower performance, as their models are more robust in areas with few observations compared to LR and ANN. ANNs unfold their potential for landslide susceptibility mapping only completely in areas with many non-linear correlated continues data sets, where they are superior to other methods. Modifying the study area and sampling technique can have a bigger impact on the final susceptibility model than using another data-driven landslide susceptibility mapping method.