Improvement of the predictive performance of landslide mapping models in mountainous terrains using cluster sampling

Landslide predictive performance is expected to vary with different sampling techniques, such as landslide random and cluster sampling. Current advancements in remote sensing technologies and machine learning (ML) have enhanced landslide prediction performance. The Himalayan Mountain range in Pakistan poses an unadorned threat to the ecosystem and valley population because of landslide occurrence. The present study explores, and tests alternative sampling technique based on spatial pattern characterization in the wake of increased landslide prediction efficacy, rather than a renowned random technique for training and testing sampling. Thereupon, landslide inventory data with 17 geo-environmental factors (i.e. topographic, hydrological and seismic factors) were determined. Landslide cluster patterns were confirmed by the Nearest Neighbor Index (NNI) method and after getting the cluster patterns, the predicted performance of landslide sampling was tested using ML and statistical methods. Advanced ML algorithms including Random Forest (RF), Extreme Gradient Boosting (XGBoost), Naive Bayes (NB), K-nearest Neighbors (KNN) and statistical methods including Weight-of-Evidence (WofE) and Logistic Regression (LR) were used and validated. The landslide-prone district of Azad Jammu and Kashmir (Neelum Valley), Kashmir Himalayas, Pakistan, was selected as a case study. Prediction performance rates are high with area under the curve (AUC) ranging from 0.802 to 0.912; accuracy (ACC) ranges from 0.78 to 0.89, and kappa ranges from 0.50 to 0.68 with cluster sampling technique, whereas the performance was low with random sampling technique, with AUC ranges from 0.768 to 0.895; ACC ranges from 0.74 to 0.86 and kappa ranges from 0.48 to 0.64. The descending order of accuracy of the six algorithms was XGboost, RF, KNN, NB, LR and WofE. Our results confirmed that the landslides followed cluster patterns in the study area, and ML algorithms with cluster training samples positively affected landslide susceptibility prediction with a statistically significant difference. The outcomes support the hypothesis that using landslides spatial natural existence, as training samples, instead of random concepts, improves the prediction ability; and highlights that alternative landslide partitioning technique could be a practicable and robust choice for landslides prediction modelling.