Hyper-parameter optimization for enhanced machine learning-based landslide susceptibility mapping

Landslides pose a substantial threat to life and property, and landslide susceptibility mapping is crucial for effective disaster management. Machine learning (ML) techniques can efficiently generate landslide susceptibility maps (LSMs) to identify high-risk areas. However, the performance of ML models relies on the careful tuning of hyper-parameters. This study focuses on hyper-parameter optimization (HPO) techniques to enhance the accuracy and reliability of ML-based landslide susceptibility mapping. The study compares different HPO methods like grid search (GS), random search (RS), Bayesian optimization (BO), hyperband, and iterative race (iRace), with a particular emphasis on introducing the iRace optimization technique in landslide susceptibility mapping studies. Different ML models like CART, SVM, RF, XGBoost, and LightGBM were used to explore the influence of the HPO techniques. The ML-HPO techniques are assessed using metrics like AUC, accuracy, $\kappa$ , precision, recall, and F1-score, utilizing data from the northeastern Indian states. The best ML-HPO combinations for each state are Arunachal Pradesh (GS-LightGBM ), Assam (iRace-RF and RS-RF), Manipur (GS-XGBoost), Meghalaya (BO-RF), Mizoram (iRace-RF), Nagaland (Hyperband-RF), Sikkim (BO-RF), and Tripura (BO-XGBoost). Results suggest GS, iRace, and BO are effective HPO techniques. The final LSM of northeast India integrates the susceptibility map generated using the best ML-HPO combinations for each state. The map can enable effective mitigation strategies and land-use planning, ultimately reducing the impact of landslides in the region.