Integrating the Particle Swarm Optimization (PSO) with machine learning methods for improving the accuracy of the landslide susceptibility model

Landslide is one of the serious concerns due to which, the safety and sustainability of hilly areas across the globe, become vulnerable. Therefore, preparing of landslide susceptibility maps (LSMs) with the sound methods is a preliminary task for the safe and sustainable land use planning and design. In this research, Particle Swarm Optimization (PSO) was integrated with the pre-existing machine learning techniques such as Artificial Neural Network (ANN), Radial Basis Functional Neural Network (RBF-net), Reduced Error Pruning Tree (REPTree), Random Tree, Multivariate Adaptive Regression Splines (MARS) and M5tree to increase their efficiency and accuracy of prediction of landslide susceptibility in upper catchment of Meenachil river of Kerala. For the advancement of the ongoing research, a total of 189 landslide locations were analysed and datasets were obtained. To prepare LSMs, 70% of the total datasets was utilized for training and the rest 30% was used for the validation purposes. In this research, methods like: ROC, Precision, Proportion Incorrectly Classify, Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Taylor diagram were applied for the validation of the models. Based on the prior pieces of literature, a total of twelve landslide conditioning factors (LCFs) were chosen. However, none of them found to be possessing multi-collinearity. It is challenging to select features from a dataset through optimization. In this regards PSO is effective because, it is straightforward with efficient universal optimization techniques. The PSO algorithm has updated and optimized the weights of models, and as results, the efficiency of the used models in predicting landslide susceptibility has increased. Nearly, 13% of the study area is very highly susceptible to landslide. The area under the curve (AUC) value of the Random Tree-PSO integrated model is the highest, 86.38% for the training dataset and 88.05% for the validation dataset. According to the sensitivity analysis elevation is most sensitive factor (0.285) and curvature is very less sensitive factor (0.115). As a result, it can be concluded that, of all the models evaluated, it is the most suitable for predicting a landslide tragedy.