Automatic relict landslide identification using LiDAR and machine learning

Abstract

The objectives of this Study are twofold: (1) to evaluate the applicability of Machine Learning (ML) and Deep Learning (DL) methods for pixel-based identification of relict landslides based on Light Detection and Ranging (LiDAR) data, and (2) to examine the importance of input features to the model’s performance. The input features include elevation, slope gradient, slope curvature, slope aspect, slope roughness and aerial photographs of year 1963. To achieve the objectives, three traditional ML models – Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR) were compared with one DL model – the Convolutional Neural Network (CNN) with U-Net architecture. The four models were trained and tested on a 6-band composite dataset derived from LiDAR data and historical aerial photographs. The results demonstrated that the DL model, CNN U-Net, outperformed the traditional ML models in terms of precision scores and visual coherence. Among the ML models, RF achieved the highest accuracy but exhibited a noticeable saltand- pepper effect. While SVM and LR models produced smoother results but did not match the performance of CNN U-Net. Furthermore, this Study identified the three most important features: slope gradient, surface roughness, and elevation contributing to model accuracy. These features are directly related to the physical conditions of relict landslides and their inclusion significantly enhanced the models' classification capabilities. This Study underscores the potential of applying high-resolution remote sensing data alongside Artificial Intelligence (AI) techniques to the landslide hazard assessment and risk management. It paves a way to create a more efficient and systematic approaches for landslide identification, susceptibility mapping and prediction, leading to informed decision making of risk mitigation strategies.