Performance analysis of a U-Net landslide detection model

Image segmentation algorithms are a type of image classifier that assigns a label to each individual pixel in an image. U-Nets, initially developed for the analysis of biomedical images and now widely used in a variety of fields, are an example of such algorithms. It has been shown that U-Nets are specially interesting when working with small training datasets and combined with data augmentation techniques.

In this study, we used satellite images with labelled landslide masks from known events to train a U-Net to identify areas of potential landslide. These landslide masks are time-consuming to create, resulting in a small initial training set. Even when working with U-Nets, the success of machine learning and AI tools depends on the availability and quality of training data, as well as the algorithm settings during the training process. Tuning machine learning models to achieve the best performance possible from limited amounts of data is important to generate trustworthy results that can be used to advance the knowledge of landslide events around the world.

Here, we show the differences in algorithm performance as we use different types of data augmentation and model parameters. We also explore and assess the effects on performance of options such as including different satellite bands, terrain information and alternative colour band representations.