Automatic detection of landslides from satellite images using a range of training events

Landslides in remote or uninhabited regions can be undocumented, leaving gaps in landslide inventories which are a key input for hazard and risk assessments. This can lead to landslide events being missing from research studies, and contribute to a bias in the events used for training of machine learning models.

In this work we use satellite images, terrain information, and labelled examples of landslides to train a convolutional neural network (U-Net), for the purpose of adding previously undocumented and new landslides to inventories. This model segments the input images and highlights the pixels it labels as landslides.

Our work focusses on landslides with a range of types and triggers, so that the model is exposed to a variety of training data. We describe the key properties of the landslides in the training set, and discuss the implications for future uses of the trained model.