Different sampling strategies for predicting landslide susceptibilities are deemed less consequential with deep...

Area of Science:

• Geosciences and Artificial Intelligence
• Geotechnical Engineering
• Machine Learning in Earth Sciences

Background:

• Landslide susceptibility mapping is crucial for hazard assessment and mitigation.
• The choice of sampling technique can influence the predictive accuracy of landslide models.
• Advancements in artificial intelligence offer new possibilities for improving landslide susceptibility paradigms.

Purpose of the Study:

• To evaluate the predictive performance of different landslide sampling techniques.
• To compare the effectiveness of logistic regression (LR), neural network (NNET), and deep learning neural network (DNN) models.
• To assess the impact of sampling strategies on landslide susceptibility mapping accuracy.

Main Methods:

• Utilized four sampling techniques: landslide scarp centroid, landslide body centroid, scarp region samples, and landslide body samples.
• Employed logistic regression (LR), neural network (NNET), and deep learning neural network (DNN) models for analysis.
• Tested models using 11 predictor variables (seismic, topographic, hydrological) in the 2018 Hokkaido Earthquake affected areas.

Main Results:

• Deep learning neural network (DNN) models demonstrated high and consistent predictive accuracy (AUC: 0.904 - 0.919) across all sampling techniques.
• Logistic regression (LR) and neural network (NNET) models showed greater variability in accuracy depending on the sampling method.
• The highest success rates were consistently achieved using samples from the landslide scarp area.

Conclusions:

• Deep learning neural network (DNN) models are robust and suitable for landslide susceptibility mapping, especially when landslide inventories lack detailed characterization.
• Sampling strategy significantly impacts predictive accuracy for LR and NNET models, highlighting the importance of careful data selection.
• Prioritizing landslide scarp samples can enhance the predictive performance of susceptibility models.