Factors Affecting Machine Learning-Based Landslide Detection for the 2015 Lefkada Earthquake

Accurate methods to efficiently detect landslides are needed to generate landslide inventories following natural disasters, such as earthquakes. This study leverages a detailed three-dimensional inventory of more than 700 landslides that occurred during the Mw 6.5 Lefkada earthquake on November 17, 2015, to quantify the factors that affect the accuracy of machine learning algorithms in detecting landslides. Several machine learning algorithms are trained using spectral features from satellite imagery and geometric features from digital elevation models. The selection of features is the most critical consideration for successful landslide detection, and the postevent slope and preevent slope are the two most essential features in the Lefkada event. At the same time, the 10 best features result in a robust landslide detection model that is similar to the model performance using 92 features. The size and geospatial distribution of training samples are also critical. Secondary factors affecting detection accuracy are the resolution of input data and machine learning algorithms. Other factors, such as segmentation parameters, the spatial distance between nonlandslide and landslide training samples, and the geometry of the training sample are less critical. Geospatial distribution influences the training sample size required to generate a reliable landslide detection model. Wider geospatial distribution of training samples results in better landslide detection ability.