Creating A Training Data Set For A Convolutional Neural Network Classifier To Map Landslides

Though often associated with steep areas experiencing heavy rainfall, defining the likelihood that a landslide may occur in an area within a particular interval of time is more challenging. Unfortunately, these probabilities are the main quantitative tool required for decision making and accurate landslide risk assessments. Thus, most of the existing landslide assessments are primarily based on the occurrence of past landslides within a particular area, which are often subjective and based on limited inventories of past failures.

Recent advances in high resolution, short repeat observations of Earth and automated identification through convolutional neural network (CNN) classifiers may finally allow us to characterize and monitor landslides over broad areas in near real-time. However, these classifiers require tens to hundreds of thousands of training samples to make accurate predictions, posing a significant challenge to their use for landslide detection. In this contribution, we exploit high-resolution mapping of earthquake-generated landslides to create a large training data set for the CNN classifiers. Earthquakes cause unstable ground to shake, resulting in large numbers of spatially clustered landslides, whose time of failure is known precisely. By exploiting the geophysical process of earthquakes, we are able to collect many well-mapped optical samples in spatially restricted areas. Our training data is based on cloudless, four-band ortho tile imagery at a resolution of 5m / pixel from Planet Lab's Planetscope constellation. We have been able to create a pipeline to extract image chips from the monthly mosaics of the locations corresponding to the data sets. We are currently using the pipeline to create a set of image chips based on the highest quality data sets. The highest quality data sets will be used to train the classifier through different climatic zones to create a classifier for detecting where and when landslides occur across the surface of Earth.