Active learning for designing detectors for infrequently occurring objects in wide-area satellite imagery

Generating ground truth to design object detectors for large geographic areas covered by hundreds of satellite images poses two major challenges: one algorithmic and the other rooted in human-computer interaction considerations. The algorithmic challenge relates to minimizing the human annotation burden by collecting only those ground truth samples that are likely to improve the classifier. And the human-computer interaction challenge relates to the temporal latencies associated with scanning all the images to find those ground truth samples and eliciting annotations from a user. We address the algorithmic challenge by using the now well-known concepts from Active Learning, albeit with a significant departure from how Active Learning has traditionally been presented in the literature: we present a human-operated active learning framework, rather than relying on previously collected fully labeled datasets for simulated experiments. And, we address the human-computer interaction challenge by using a distributed approach that relies on multiple virtual machines working in parallel to carry out randomized scans in different portions of the geographic area in order to generate the active-learning based samples for human annotation. We demonstrate our wide-area framework for two infrequently occurring objects over large geographic areas in Australia. One is for detecting pedestrian crosswalks in a region that spans 180,000 sq. km, and the other is for detecting power transmission-line towers in a region that spans 150,000 sq. km. Using randomly selected unseen regions for measuring detector performance, the crosswalk detector works with 92% precision and 72% recall, and the transmission-line tower detector with 80% precision and 50% recall.