Hyperspectral imagery provides the ability to detect targets that are smaller than the size of a pixel. They provide this ability by measuring the reflection and absorption of light at different wavelengths creating a spectral signature for each pixel in the image. This spectral signature contains information about the different materials within the pixel; therefore, the challenge in subpixel target detection lies in separating the target's spectral signature from competing background signatures. Most research has approached this problem in a purely statistical manner. Our approach fuses statistical signal processing techniques with the physics of reflectance spectroscopy and radiative transfer theory. Using this approach, we provide novel algorithms for all aspects of subpixel detection from parameter estimation to threshold determination.

Characterization of the target and background spectral signatures is a key part of subpixel detection. We develop an algorithm to generate target signatures based on radiative transfer theory using only the image and a reference signature without the need for calibration, weather information, or source-target-receiver geometries. For background signatures, our work identifies that even slight estimation errors in the number of background signatures can severely degrade detection performance. To this end, we present a new method to estimate the number of background signatures specifically for subpixel target detection.

At the core of the dissertation is the development of two hybrid detectors which fuse spectroscopy with statistical hypothesis testing. Our results show that the hybrid detectors provide improved performance in three different ways: insensitivity to the number of background signatures, improved detection performance, and consistent performance across multiple images leading to improved receiver operating characteristic curves.

Lastly, we present a novel adaptive threshold estimate via extreme value theory. The method can be used on any detector type - not just those that are constant false alarm rate (CFAR) detectors. Even on CFAR detectors our proposed method can estimate thresholds that are better than theoretical predictions due to the inherent mismatch between the CFAR model assumptions and real data. Additionally, our method works in the presence of target detections while still estimating an accurate threshold for a desired false alarm rate.