Landslide susceptibility modelling using deep-learning and machine-learning methods-A study from southern Western Ghats, India

Accurate and reliable landslide susceptibility modelling is essential for regional landslide risk assessment and mitigation. In the present study, a spatially explicit deep learning neural network (DNN) based landslide susceptibility model for Wayanad district of Kerala state in India has been developed and it is compared with conventional random forest machine-learning model. The Wayanad was severely affected by devasting landslides during 2018 extreme climatic event and absence of an accurate landslide susceptibility model caused extensive causalities. A geospatial database was developed using 263 previous landslides event locations and twelve geo-environmental variables such as lithology, soil texture, land use/ land cover, slope angle, slope aspect, topographic wetness index, distance from the road, distance from the streams, distance from the lineaments, convergence index, profile and plan curvatures were used for susceptibility modelling. The present investigation used search grid to obtain best hyper tuning parameters and training data is split for 5-fold cross validation. Subsequently the best model is chosen based on minimal log loss, RMSE and maximum area under curve values (AUC) value for final prediction. The results are validated with receiver operating characteristics curve with AUC (ROC-AUC) and confusion matrix-based parameters. Results showed that DNN outweigh RF model with 93.4% (AUC=0.934) accuracy over 91.1% in training frame. In validation section DNN obtained 88.7% accuracy while RF shows only 85%. It is also noted that DNN shows better spatial pattern in predicting the probability of landslide occurrence whereas a spatial overfitting is found in RF model. The proposed DNN model is trust worthy and can be used for future hazard mitigation and land use planning in the study area.