Mapping Complex Landslide Scars Using Deep Learning and High-Resolution Topographic Derivatives from LiDAR Data in Quebec, Canada

Abstract

This study evaluates deep learning (DL) models, particularly ResU-Net with attention mechanisms, for mapping landslides in Quebec, Canada, utilizing high-resolution digital elevation model (HRDEM) data and its seven derivatives (slope, aspect, hillshade, curvature, ruggedness, surface area ratio, and max difference from mean). Three scenarios were considered to assess the effectiveness of various features in landslide segmentation: training the model on all features, each feature individually, and on slope and hillshade. Model performance on individual features was significantly poor, while the model trained with hillshade and slope outperformed the model using all seven features, particularly in F1-score (improved by 8% for rotational landslides and 11% for retrogressive landslides) during validation. Furthermore, for the test dataset, model performance on all seven features was compared against slope and hillshade. As a result, for rotational landslides, slope and hillshade achieved F1-scores of 0.68 and 0.93 for rotational and retrogressive landslides, respectively, while the same metrics using all features were 0.61 and 0.83, respectively. This suggests hillshade and slope provide the most relevant information and reduce computational complexity. Overall, the findings enhance our understanding of HRDEM derivatives and emphasize the importance of feature selection in optimizing model performance and reducing computational complexity.

Keywords

• remote sensing
• rotational landslide
• retrogressive landslide
• machine learning