Application of variable importance tool in improving the efficiency of multilayer perceptron model for landslide susceptibility

A Landslide is a natural process that turns into a hazard when it interacts which human settlement. The slope failure is caused by the interaction between the geologic, geomorphic, and hydrologic forces, which works on a scale that varies from seconds to thousands of years. The interaction between the various forces increases the curiosity among the researchers, which led them to involve as many as possible to compute the landslide susceptibility. This increase in the magnitude of the data requires high computation ability, which is not ubiquitous in every academic institute. The misinterpretation of the term big data-led people to believe that model is not reliable until a large amount of the data trains it. It is found that the redundancy in the data set caused by the high correlation among the landslide causing factors reduces the model's accuracy. The neural network's variable importance (vi) tool prioritizes the factors based on their weightings in calculating the correct output. The tool uses the loss function to define the factor's importance in the model. In this study, the Pithoragarh region of the Uttarakhand state of India is taken as the study area, and the region is prone to landslide events. The predisposing factors such as Elevation, stream, Slope, Road, NDVI, Curvature, Lineaments, Lithology, LULC, Geomorphology, Aspect, and Soil depth are considered to train the model. The entire work is carried out in ArcMap and RStudio. In this study, the model agnostic and model-specific methods of variable importance are used. The outcome suggested that topographic factors are more important in the model's training than the anthropogenic, geologic, and geomorphic factors. Also, the system running time has been reduced by 18.2 per cent. The outcome of this work will be useful, especially in the natural hazard's studies, in which a great magnitude of the data is involved in the processes. The vi tool will reduce the data requirement for the model tuning, which will lead to better, faster, and accurate results.