Combining spectral matching and anomalous change detection for target rediscovery in hyperspectral images

Abstract

In surveillance applications, tracking a specific target by means of subsequent acquisitions over the monitored area is of great interest. Multitemporal HyperSpectral Images (HSIs) are particularly suitable for this application. Multiple HSIs of the same scene collected at different times can be exploited to detect changes using anomalous change detection (ACD) techniques. Moreover, spectral matching (SM) is a valuable tool for detecting the target spectrum within HSIs collected at different times (target rediscovery-TR). Depending on the monitored area and the specific target of interest, TR can be a challenging task. In fact, it may happen that the target has spectral features similar to those of uninteresting objects in the scene and the use of SM techniques without additional information can generate too many misleading detections. We introduce a new TR strategy aimed at mitigating the number of alarms encountered in complex scenarios. The proposed detection strategy combines the SM approach with the unsupervised ACD strategy. We focus on rediscovery of moving targets in airborne HSIs collected on the same complex area. False alarms mitigation is achieved by exploiting both the target spectral features and the temporal variations of its position. For this purpose, SM is performed only on those pixels that have undergone changes within multiple acquisitions. Results obtained applying the proposed scheme on real HSIs are presented and discussed. The results show the effectiveness of the fusion of spectral and multitemporal analysis to improve TR performance in complex scenarios