Evaluating landslide hazard in Hunza-Nagar watershed basin, through GIS-based, statistical and machine learning techniques.

The Hunza-Nagar districts in northern Pakistan, nestled in the central Karakoram range, are known for their distinct climate and geology. However, the area is geologically unstable due to steep mountains, active faults, seismic zones, and sheared rock masses, leading to frequent landslides. This research aims to assess the comparative efficacy of three cutting-edge deep machine learning techniques (DMLTs), namely, the Universal Network U-Net, InceptionV3+, and DeeplabV3+, alongside three bivariate statistical models: Weight of Evidence (WofE), Intuitionistic Fuzzy Divergence (IF-D), and Frequency Ratio (FR), for Landslide Susceptibility Mapping (LSM) in Hunza-Nagar, North Pakistan. Initially, 148 landslide locations were marked, and 10 conditional factors were chosen to create a Landslide Inventory Map (LIM). The validation of landslide susceptibility models was assessed using several metrics: Area under the Curve (AUC), Density of Landslide Distribution (DLA), and the Seed Cell Area Index (SCAI). The evaluation of landslide susceptibility maps, produced through statistical models and DMLTs, revealed notable Prediction Rate Curves (PRC). Specifically, the WofE model achieved an 85 % PRC, the FR model reached 82 %, and the IF-D model also achieved an 85 % PRC. Furthermore, when estimating the performance of DMLTs using PRC, DeeplabV3+ established an 82 % success rate, InceptionV3+ reached 79 %, and U-Net achieved 80 %. In conclusion, the analysis designates that the WofE model, with an 85 % PRC and a D-value of 3.6, and the IF-D model, boasting an 85 % PRC and a D-value of 2.4, exhibited the highest prediction accuracy and classification ability.