Key Elements for Improving Pixel-Based Landslide Susceptibility Mapping

These key elements for improving pixel-based landslide susceptibility mapping include:

(1) Constructing new landslide explanatory factors with higher nonlinearity

Physics-based models such as ISSA are powerful tools for generating new nonlinear explanatory factors, such as the slope safety factor. These factors enhance model generalization by transforming inconsistent geological and lithological data across regions into a continuous, normalized safety factor. They also mitigate prediction uncertainty caused by incomplete landslide inventories by pre-selecting non-landslide samples.

(2) Considering the pixel spatial neighborhood

Model performance improves substantially when using local-area samples instead of single-pixel samples. With local-area samples, the AUC values of Hybrid Model I in the testing region increase significantly, reducing the extent of areas mistakenly classified as high susceptibility and improving low-susceptibility area recognition.

(3) Increasing the nonlinearity of the data-driven module

Hybrid Model II consistently outperforms Hybrid Model I across the training region, testing region, and a new region (Wushan County). However, the improvement from considering pixel spatial neighborhoods exceeds the improvement contributed by increasing nonlinearity. Hybrid Model I—while incorporating spatial neighborhoods—already offers strong predictive accuracy along with better practicality and interpretability.

(4) Reducing model parameters

Because landslide labels are limited, minimizing the number of trainable model parameters is essential for improving susceptibility mapping. Fewer parameters reduce the required number of labels while maintaining predictive performance. This can be achieved through classical statistical models like logistic regression, or through specific deep learning architectures such as CNNs, which inherently reduce parameters via weight sharing and local connectivity.

(5) Increasing data quantity and quality

Both the number of landslide records (labels) and the quality of the input features (explanatory factors) strongly affect LSM model performance.

Read the full paper: Improving pixel-based regional landslide susceptibility mapping

Wei, X., Gardoni, P., Zhang, L., Tan, L., Liu, D., Du, C., & Li, H. (2024). Improving pixel-based regional landslide susceptibility mapping. Geoscience Frontiers, 15(4), 101782.