A comparative study of statistical methods and machine learning algorithms for prediction of landslides in Mizoram state of India through analysis of causative factors using Geo-informatics

Abstract The landslide has been one of the most severe and significant natural hazards in the study area, Mizoram, which has rolling hills and deep valleys in almost every landform. A comparative study of landslide hazards in the area was conducted using various statistical analytic techniques and machine learning algorithms. The statistical method includes- Frequency Ratio (FR), Analytic Hierarchical Process (AHP), Shannon’s Entropy (SE), and Weight of Evidence (WOE), while the machine learning algorithms methods comprise basic classifiers such as Gradient Boosting Decision Tree (GBDT), Random Forest (RF), and Extreme Gradient Boosting (XGB), and hybrid classifiers using the Logistic Regression (LR) methods viz., GBDT + LR, RF + LR, XGB + LR. The study aims to find out the collinearity of various parameters of landslide-inducing factors and analyse their weight for most contributing factors to least contributing factors. It also aims to develop the Landslide Hazard Zonation (LHZ) map using various parameters weights and layer stacking by weighted sum overlay in a GIS software environment. The generated LHZ map was separated into five classes viz., low, moderate, high, very high, and severe. For statistical analysis, validation of the zonation maps was done by using past landslide inventories. Classification of the number of past landslides point data in each class of the zonation map was done to validate the accuracy of the zonation map. More than 65 per cent of Landslide point data falls in the High to Severe zone in the classification for FR, AHP, and SE which was considered to be in the positive validate zone, whereas only 60 per cent of Landslide point data falls in the High to Severe zone for WOE which was considered to be inadequate and undesirable for applicable LHZ map. For machine learning algorithms, a buffer zone of a 50m radius was created for the application of the seeding technique for preparing landslide inventory. More than 10000 landslide seeds cells and non-landslide cells were taken in which 80% and 20% train-test split was conducted. A series of metrics such as accuracy, precision, recall, f- f-measure, Area Under (receiver operating characteristic) Curve (AUC), kappa index, mean absolute error (MAE), and root mean square error (RMSE) was used to evaluate the accuracy and performance of the seven models. Based on the AUC curve, the XGB model having the highest AUC value (0.9039) was identified as the most efficient model among the machine learning models. It was found that an improvement of more than 15% accuracy was shown by the machine learning models compared to the statistical approach. The results suggest that the machine learning method is propitious for an application in landslide estimation in the study area.