Efficient Hyperspectral Target Detection and Identification With Large Spectral Libraries

Abstract

Numerous hyperspectral algorithms have been developed to detect both full and sub-pixel solid target materials. Target signatures are obtained from spectral libraries that contain both target and non-target materials. When the library is large and contains many potential targets, it is inefficient to run an individual detector for each material of interest. Additionally, such an approach produces numerous false alarms (i.e., multiple detections per pixel) due to spectral similarity among targets. In this paper, we present an efficient approach for detecting multiple targets within large spectral libraries while mitigating false alarms. We first group spectrally similar materials within the library into a hierarchy of clusters. From each cluster containing a target material, a single detector is obtained. Each detector represents multiple library spectra, so an identification step is needed for detected pixels. Detected pixels are modeled as a mixture between their local in-scene background and candidate library spectra. Candidates are chosen from adjacent library clusters. The candidate model providing the best fit is chosen to report. Use of local background spectra provides a physically meaningful mixing model that adapts to detected pixels. Clustering the library reduces the computational complexity of modeling detected pixels. We demonstrate detection and false alarm mitigation performance of our proposed algorithm with a real hyperspectral dataset.

Spectral variability modeling with variational auto-encoders for hyperspectral target analysis

“…Such a system can effectively be used to substantially reduce FAs when compared to using a detector bank alone. 11 A library is denoted by a set of N lib spectra S = {s 1 , s 2 , . .…”

Physics-guided neural networks for hyperspectral target identification

“…[23][24][25][26][27][28] For large spectral libraries, the sensing problem is sometimes framed as a two-step process consisting of target detection and identification; we note that spectral library sizes often inflate significantly when including multiple representations of targets with high variability in their spectra (e.g., due to morphology). [29][30][31][32] In a combined detection and identification procedure, candidate target pixels are first detected, and then target identification traditionally consists of comparing a region of interest (ROI) consistent of detected pixels against the spectral library, 33,34 often using tools such as linear regression. 35 While these approaches work well for target detection, the target identification problem -in the face of very large (and continuously growing) spectral libraries -invites reframing the identification step as a multi-class classification problem; deep learning is a natural fit for tackling this challenge 5,36 as neural networks can learn nonlinear relationships between the signature and the class label, potentially offering more accurate identification.…”

Anomaly Detection Based on Tree Topology for Hyperspectral Images

“…The abundant information provided by HSIs makes hyperspectral remote sensing a valuable technology with strong comprehensiveness and broad application prospects [7], [8], [9]. The research on target detection is one of the most important directions of hyperspectral remote sensing [10], [11], [12], exhibiting excellent performance and unique advantages in many civil and military fields [13], [14], [15]. With the rapid development of remote sensing, the quality of captured observational data has substantially improved [16], [17].…”