Exploiting multiview temporally-diverse hyperspectral imagery for activity pattern analysis

• Published in: Proceedings of SPIE, the international society for optical engineering Vol. 13455; pp. 134550D - 134550D-27
• Main Authors: Chan, Allison, Whelsky, Amber, Qiu, Frank, Van Omen, Alan, vander Laan, John, Ziemann, Amanda, Anderson, Dylan
• Format: Conference Proceeding
• Language: English
• Published: SPIE 28.05.2025
• ISBN: 1510686991; 9781510686991
• ISSN: 0277-786X
• DOI: 10.1117/12.3053254

Understanding activities occurring at a location of interest (e.g., commercial bakeries, gas processing facilities) requires the detection, characterization, and interpretation of complex and often ambiguous activity-related signatures. Hyperspectral imagery (HSI) is phenomenologically well-suited for detecting the physical indicators associated with these signatures due to its high spectral resolution, enabling precise material identification and analysis that surpasses traditional imaging technology capabilities. HSI has already proven effective in applications such as detecting effluent plumes, analyzing aging chemical residues, and identifying paint on vehicles, making it a natural fit for activity pattern analysis. This research presents a new vector-first framework for utilizing multi-look, temporally-diverse HSI for activity pattern identification. For a set of multi-view, temporally-diverse HSI data, the approach first identifies regions of interest (ROIs) through three methodologies (target-based, anomaly-based, and clustering-based), and then associates these ROIs into “activity patterns” through spectral and spatial similarity criteria. Finally, Dynamic Bayesian Networks are used to map these associated activity patterns to process-based observables to understand and interpret the overall activities. Using an airborne longwave infrared (LWIR) hyperspectral dataset over the Los Angeles region from 2017, we demonstrate the entire workflow from ROI extraction to multi-ROI activity pattern characterization with quantified confidence in activity interpretation at several distinct locations. This dataset contains the varied collection geometries expected in upcoming commercial spaceborne HSI, as well as temporally-varying activity patterns. The preliminary success seen in this workflow relies on the integration of expected, possible, and unexpected signatures, low-confidence detections, and subject matter expertise into a deeper understanding of on-ground activity.