# Landslide Susceptibility Modeling by Interpretable Neural Network

**Authors:** Khaled Youssef, Kevin Shao, Seulgi Moon, Louis-Serge Bouchard  
**arXiv:** [2201.06837](https://arxiv.org/abs/2201.06837)  
**Date:** 2022-01-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2201.06837v3/

## Key points

- Landslides are a major natural hazard that is difficult to predict
- Artificial neural networks (ANN) have shown promise in improving prediction accuracy
- Researchers introduce a new approach called superposable neural network (SNN) optimization
- SNN models outperformed physically-based and statistical models
- SNN models achieved similar performance to state-of-the-art deep neural networks
- Slope and precipitation, as well as hillslope aspect, were important contributors to high landslide susceptibility
- Strong slope-climate couplings and microclimates play a significant role in landslide occurrences
- SNN models can identify contributing features and extract their functions from the data
- The easternmost Himalaya region is particularly susceptible to landslides due to various factors such as steep slopes, extreme precipitation events, flooding, and earthquakes
- The researchers generated a landslide inventory for this region using manual delineation and a semi-automatic detection algorithm
- The SNN model utilized 15 single features including elevation, slope mean annual precipitation, distance to channels and faults
- The study provides valuable insights into the contributing factors of landslides in the easternmost Himalaya region.

## Summary

Landslides are a major natural hazard that is difficult to predict due to the complex and varying factors that contribute to slope stability. However, artificial neural networks (ANN) have shown promise in improving prediction accuracy. In this study, researchers introduce a new approach called superposable neural network (SNN) optimization, which combines interpretability, high accuracy, high generalizability, and low model complexity. To validate their approach, the researchers trained SNN models on landslide inventory data from three different regions in the easternmost Himalaya. The SNN models outperformed physically-based and statistical models and achieved similar performance to state-of-the-art deep neural networks. The analysis revealed that the product of slope and precipitation as well as hillslope aspect were important contributors to high landslide susceptibility. This highlights the significance of strong slope-climate couplings and microclimates in landslide occurrences. The researchers also conducted validation tests using synthetic datasets which demonstrated the ability of the SNN models to identify contributing features and extract their functions from the data. The SNN optimization algorithm used in this study is described in detail. The easternmost Himalaya region is particularly susceptible to landslides due to steep slopes, extreme precipitation events, flooding, and frequent earthquakes. The researchers generated a landslide inventory for this region by combining manual delineation with a semi-automatic detection algorithm which included 2289 mapped landslides with varying sizes across three specific regions within: N-S region (Dibang), E-W region (Lohit), and NW-SE region (range front). By testing the SNN model on these regions with different environmental conditions they aimed to identify universal or distinct controls of landslides as well as similarities or differences in feature contributions across these regions. The SNN model utilized 15 single features including aspects such as elevation, slope mean annual precipitation and distance to channels and faults which were selected based on previous studies that examined landslide susceptibility. They also compared its performance against individual classifiers for each feature as well as a physically based slope stability model (SHALSTAB). In conclusion this study provides valuable insights into the contributing factors of landslides by demonstrating the effectiveness of SNN models in predicting landslide occurrences with high accuracy and interpretability while highlighting the innovative approach of using superposable neural network optimization for assessing landslide susceptibility in the easternmost Himalaya region.