Sensitivity analysis workflow to address landslide susceptibility model variability

Abstract

Contemporary landslide susceptibility modeling commonly utilizes machine learning (ML) approaches and algorithms. These modeling approaches typically use landslide inventories as a primary component, along with a conditioning feature selection method. Although model results may be sensitive to variability in input data, resulting effects are seldom investigated in detail. We leveraged a susceptibility approach based on logistic regression (LR) to assess the sensitivity of model results to landslide inventory size. Using an inventory of 1054 landslides, we evaluated sample size at 50-landslide increments (to the maximum possible value of 1050), generating 1000 random inventory subsets for each sample size (i.e., Monte Carlo simulation) to quantify variation in model performance due to randomized train and holdout selection. Initial visual inspection of susceptibility map outputs produced using these model results revealed extreme overprediction. The spurious models indicate a false correlation between model accuracy and realism. To address this, our sensitivity analysis workflow utilizes a performance metric correction factor and stricter feature selection to promote consistent, realistic, and reproducible models. The workflow presented here quantifies model variability and assesses the sensitivity of susceptibility maps to changing model inputs.