Parallel algorithms for fast subpixel detection in hyperspectral imagery

Abstract

We present parallel algorithms for fast subpixel detection of targets in hyperspectral imagery produced by our Hyperspectral Airborne Tactical Instrument (HATI-2500). The HATI-2500 hyperspectral imaging system has a blue-enhanced visible-near-IR (VNIR) and a full short-wave IR (SWIR) range response from 400 to 2500 nm. It has an industry-leading spectral resolution that ranges from 6 nm down to 1.5 nm in the VNIR region. The parallel detection algorithm selected for processing the hyperspectral data cubes is based on the adaptive coherence/cosine estimator (ACE). The ACE detector is a robust detector that is built upon the theory of generalized likelihood ratio testing (GLRT) in implementing the matched subspace detector to unknown parameters such as the noise covariance matrix. Subspace detectors involve projection transformations whose matrices can be efficiently manipulated through multithreaded massively parallel processors on modern graphics processing units (GPU). The GPU kernels developed in this work are based on the CUDA computing architecture. We constrain the detection problem to a model with known target spectral features and unstructured background. The processing includes the following steps: 1) scale and offset applied to convert the data from digital numbers to radiance values, 2) update the background inverse covariance estimate in a line-by-line manner, and 3) apply the ACE detector for each pixel for binary hypothesis testing. As expected, the algorithm is extremely effective for homogeneous background, such as open desert areas; and less effective in mixed spectral regions, such as those over urban areas. The processing rate is shown to be faster than the maximum frame rate of the camera (100 Hz) with a comfortable margin.