Self-Supervised Representation Learning Enhances Broad Area Search in Multi-Temporal Satellite Imagery

We describe the advancement of the classical anomaly detection paradigm to a task-relevant change detection problem. Modern machine learning methods support the development of more sophisticated multi-temporal satellite image analysis. Here, we look at the problem of detecting and distinguishing anthropogenic change from natural change over broad regions around the globe. These tasks are well-suited for machine learning algorithms, however, the creation of large scale annotated satellite imagery datasets with sufficient spatial and temporal resolution is expensive. In this paper, we explore utilizing spatiotemporal self-supervised learning which leverages the natural chronology of the data collection to train generalizable feature extractors for various downstream tasks. This approach is shown to boost downstream performance (+10% F1 score, precision, recall), and reduce training time by 80% for the broad area search problem using multi-temporal Sentinel-2 imagery.