Introducing a novel multi-layer perceptron network based on stochastic gradient descent optimized by a meta-heuristic algorithm for landslide susceptibility mapping.

The main objective of the current study was to present a methodological approach that combines Information Theory, a neural network and meta-heuristic techniques so as to generate a landslide susceptibility map. Specifically, the methodology involved three important tasks: Classifying the landslide related variables, weighting them and optimizing the structural parameters of the neural network. Shannon's entropy index was used to estimate for each landslide related variable the number of classes which maximized the information coefficient, whereas the Certainty Factor method was used to weight the variables. A Neural Network, a (NN) which uses stochastic gradient descent (SGD), the structural parameters of which are optimized by a Genetic Algorithm (GA), was implemented to generate the landslide susceptibility map. A well defined spatial database which included 380 landslides and fourteen related variables (elevation, slope, aspect, plan curvature, profile curvature, topographic wetness index, stream power index, stream transport index, land use cover, distance to road, distance to faults, distance to river, lithology and soil cover) were considered for implementing the NN-SGD-GA model, in the Yanshan County located in Shangrao Municipality, in the north-eastern of Jiangxi province, China. To validate the predictive power of the novel model, a Logistic Regression (LR) and Random Forest (RF) model were used for comparison. The results showed that the NN-SGD-GA model achieved the highest prediction accuracy (88.10%), followed by the RF (86.26%) and the LR (85.82%) models. Furthermore, by analyzing the validation data, concerning the spatial distribution of landslides and the susceptibility index, the proposed model showed an area under curve value of 0.8212, followed by the RF (0.8124) and the LR (0.8020) models. Finally, the proposed model showed the highest relative landslide density value of 65.09, followed by the RF (62.51) and the LR (61.76) models, when using the validation dataset. The novelty of our approach is the usage of an intelligent way to select and classify the most appropriate prognostic variables and also the implementation of an evolutionary wrapper automatic procedure that efficiently generates prediction models with reduced complexity and adequate generalization capacity. Overall, the proposed model can be successfully used for landslide susceptibility mapping as an alternative spatial investigation tool.