Spectral anomaly detection with machine learning for wilderness search and rescue

In wilderness search and rescue missions, unmanned aerial vehicles (UAVs) may be deployed to collect high-resolution imagery which is later reviewed by a first responder. The volume of images and the altitude from which they are taken makes manually identifying potential items of interest, like clothing or other man-made material, a difficult task. For this reason, we created a program that automatically detects unusually-colored objects in aerial imagery in order to assist responders in locating signs of missing persons. The program uses the Reed-Xiaoli (RX) spectral anomaly detection algorithm to determine which pixels in an image are anomalous and then generates an "anomaly map" where brighter pixels signify greater abnormality. While the RX algorithm has previously been proposed for search and rescue missions, up until now it has not been evaluated in a high-fidelity setting with real responders and real equipment. We tested the program on 150 aerial images taken over the Blanco River area in Hays County, Texas after the May 2015 flooding and demonstrated the results at a workshop on flooding hosted by Texas A&M's Center for Emergency Informatics. Early feedback from responders suggests that RX spectral anomaly detection is a valuable tool for quickly locating atypically-colored objects in images taken with UAVs for wilderness search and rescue.