A multi-criteria landslide susceptibility mapping using deep multi-layer perceptron network: A case study of Srinagar-Rudraprayag region (India)

Abstract Landslide susceptibility analysis is vital for understanding precursors, associated risk, and early rescue operations. In the past, numerous attempts have been made to develop landslide susceptibility maps using various tools. This research work aims to generate landslide susceptibility maps (LSM) by developing a deep multi-layer perceptron network named as DMLP-LSM and to demonstrate it over Srinagar-Rudraprayag region, a landslide prone region recognized by the Geological Survey of India (GSI). Initially, training datasets are generated using the analytical hierarchy process (AHP) enabled multi-criteria analysis, which involves information from various landslide causative factors. In a first, one dimensional line of sight (1D LOS) velocity, one of the critical inputs and prominent indicator for landslide susceptibility, is generated by processing 60 Sentinel-1A scenes using multi-temporal interferometric synthetic aperture radar (MT-InSAR) algorithm. The 1D LOS velocity map, generally used for quantitative monitoring slow moving landslides, is used here as a thematic layer. The thematic layers and LSMs were fed to the proposed DMLP-LSM network as input and output, respectively. The developed model reached an accuracy of 91.47% with precision, specificity and ROC_AUC score as 92.11%, 95.55% and 0.9072, respectively. It further identified nearly 15 square km area as susceptible to landslide. The efficacy of the proposed model is compared with two well-known machine learning models for LSM preparation, namely support vector machine (SVM) and random forest (RF). In terms of accuracy, DMLP-LSM and RF had similar performance, but better than SVM. The velocity maps derived from MT-InSAR processing played a prominent role in the generation of LSM, supporting model training and performance. The proposed model achieved highest precision and specificity, and its performance would also improve LSM preparation when using larger datasets. The obtained results are promising, and the trained model can help in the rapid generation of LSMs.