• remote sensing
• target detection
• forward modeling
• hyperspectral imaging
• background characterization
• anomaly detection

In recent years, the literature of hyperspectral target detection has seen the growth of a well-defined research topic known under the name of difficult target detection. Despite the extensive studies carried out over the years, the large variety of detection algorithms developed may still not be enough to adequately solve the difficult target detection problem. An extensive analysis of the diverse implications of real-world application is hence required to avoid the performance degradation typically observed in operational scenarios.
The main goal of this thesis is to investigate into the weak links in the difficult target detection processing chain, to develop ad hoc hyperspectral image processing methods, and to quantitatively evaluate the end-to-end performance improvements. Within this goal, two main aims have been identified and achieved by the present work.
First, the underlying problem of background characterization in difficult target detection has been examined. Three well-known background models have been investigated as regards the main limitations arising in difficult target detection applications. Novel solutions aimed at improving background modeling and estimation procedures have been accordingly proposed, and the actual payoff in terms of detection performance has been evaluated by means of simulated and real hyperspectral data.
Secondly, the impact of atmospheric, viewing, and illumination condition uncertainty on the detection process has been investigated. Physics-based approaches aimed at compensating for the radiation transfer effects have been examined as regards their operational applicability in difficult target detection scenarios. The new insights achieved have been substantiated by a thorough experimental analysis performed on real hyperspectral data.
In the light of the results obtained, this thesis has shown that investigation into the limitations of theoretically valid models and algorithms arising in operational scenarios, as well as into their impact on the detection process, can undoubtedly help identify the weak links in the detection processing chain. The ad hoc hyperspectral image processing methods derived have allowed the weak links identified to be strengthened, with improved difficult target detection performance as the ultimate outcome.