Comparison of optimized data-driven models for landslide susceptibility mapping

Locations prone to landslides must be identified and mapped to prevent landslide-related damage and casualties. Machine learning approaches have proven effective for such tasks and have thus been widely applied. However, owing to the rapid development of data-driven approaches, deep learning methods that can exhibit enhanced prediction accuracies have not been fully evaluated. Several researchers have compared different methods without optimizing them, whereas others optimized a single method using different algorithms and compared them. In this study, the performances of different fully optimized methods for landslide susceptibility mapping within the landslide-prone Kermanshah province of Iran were compared. The models, i.e., convolutional neural networks (CNNs), deep neural networks (DNNs), and support vector machine (SVM) frameworks were developed using 14 conditioning factors and a landslide inventory containing 110 historical landslide points. The models were optimized to maximize the area under the receiver operating characteristic curve (AUC), while maintaining their stability. The results showed that the CNN (accuracy = 0.88, root mean square error (RMSE) = 0.37220, and AUC = 0.88) outperformed the DNN (accuracy = 0.79, RMSE = 0.40364, and AUC = 0.82) and SVM (accuracy = 0.80, RMSE = 0.42827, and AUC = 0.80) models using the same testing dataset. Moreover, the CNN model exhibiting the highest robustness among the three models, given its smallest AUC difference between the training and testing datasets. Notably, the dataset used in this study had a low spatial accuracy and limited sample points, and thus, the CNN approach can be considered useful for susceptibility assessment in other landslide-prone regions worldwide, particularly areas with poor data quality and quantity. The most important conditioning factors for all models were rainfall and the distances from roads and drainages.