Unsupervised Anomaly Detection in Remote Sensing Imagery: Advancing Algorithms for Operational Monitoring

Abstract

Monitoring is an essential activity that various research, government, and industry groups use in their routine operations to extract key insights from regions of interest (ROIs). In remote sensing imagery, detecting anomalies - unexpected objects or significant events - is crucial for environmental management and protection, as well as surveys and surveillance. This need has become more urgent with increasing monitoring demands across diverse operational contexts, from climate-driven hazard events that require consistent long-term observation, to time-critical scenarios demanding immediate insights. This thesis presents two new unsupervised anomaly detection methods in remote sensing, SHAZAM and ERX, each designed for distinct but complementary operational monitoring contexts.

SHAZAM is a self-supervised change monitoring method for hazard detection and mapping, addressing the increasing impact of climate-driven hazardous events. Evaluated on four datasets that contain various hazards (bushfires, burned regions, extreme and out-of-season snowfall, deforestation, algal blooms, drought, and floods), SHAZAM achieved F1 score improvements of between 0.066 and 0.234 compared to similar learning-based methods, while remaining extremely lightweight with only 473K parameters.

While SHAZAM enables regular monitoring of ROIs for hazards, satellite-based anomaly detection is constrained by revisit times and low spatial resolution. ERX addresses this limitation through real-time anomaly detection for hyperspectral line scanning, offering immediate insights onboard smaller platforms such as drones. ERX processed the dataset with the highest number of bands (108 bands) 9 times faster than the next-best algorithm onboard a Jetson Xavier NX. ERX showed a 29.3% AUC improvement on the most challenging dataset, and achieved an AUC of 0.941 on drone-collected data without geometric corrections (a 16.9% improvement over existing algorithms).