Landslide Susceptibility Modeling by Interpretable Neural Network

Landslides are notoriously difficult to predict because numerous spatially and temporally varying factors contribute to slope stability. Artificial neural networks (ANN) have been shown to improve prediction accuracy. However, traditional ANNs are uninterpretable, complex black box models. This makes it difficult to extract mechanistic information about landslide controls in the modeled region or trust the outcome in this high-stakes application. Herein we present the first application of an interpretable additive neural network to landslide susceptibility modeling. We introduce a new additive ANN optimization framework, as well as new dataset division and outcome interpretation techniques uniquely suitable for modeling applications with spatially dependent data structures such as landslide susceptibility. We refer to our approach which features full interpretability, high accuracy, high generalizability and low model complexity as superposable neural network (SNN) optimization. We validate our approach by training models to assess landslide susceptibility in three different regions of the easternmost Himalaya that are highly susceptible to landslides. The interpretable neural network models generated by the SNN outperform physically-based stability and statistical models and achieve similar performance to state-of the-art deep neural networks while offering insight regarding the relative importance of landslide control factors. The SNN models found the product of slope and precipitation and hillslope aspect to be important primary contributors to high landslide susceptibility in the studied regions. These identified controls suggest that strong slope-climate couplings, along with microclimates, play dominant roles in landslide occurrences of the easternmost Himalaya.