Performance assessment of artificial neural network using chi-square and backward elimination feature selection methods for landslide susceptibility analysis

In the machine learning models, it is desirable to remove most redundant features from the data set to reduce the data processing time and to improve accuracy of the models. In this paper, chi-square (CS) and backward elimination (BE), which are well-known feature selection methods, were used for the optimum selection of input features/factors for training artificial neural network (ANN) for landslide susceptibility modeling. Initially, seventeen landslide affecting factors were considered for the ANN model which were reduced to twelve and eleven based on the ANN optimized by CS (CSANN) and BE (BEANN), respectively. Accuracy (ACC), Kappa Index, root mean square error (RMSE), and area under the receiver operating characteristic (AUROC) curve were used to evaluate and validate performance of the models. Results show that both the feature selection methods (CS and BE) improved significantly performance of the hybrid BEANN and CSANN models in comparison to single ANN model. Results indicated that performance of the BEANN model (AUROC 0.963; ACC 91.31) is the best in comparison to CSANN (AUROC 0.950; ACC 89.80) and ANN (AUROC 0.949; ACC 76.40) models in the accurate prediction of landslide susceptible areas/zones. Therefore, it is reasonable to state that the BE is more effective feature selection method than the CS in improving performance of the ANN model and thus, it can be used for better landslide susceptibility analysis for the landslide management of the area.

Automated summary

This study utilized chi-square and backward elimination methods to select optimal input features for training an artificial neural network model, with performance evaluated using various metrics. The results indicated that the BE model outperformed in accurately predicting landslide-prone areas.