Eric Truslow

Detection Algorithms in Hyperspectral Imaging Systems: An Overview of Practical Algorithms

Hyperspectral imaging applications are many and span civil, environmental, and military needs. Typical examples include the detection of specific terrain features and vegetation, mineral, or soil types for resource management; detecting and characterizing materials, surfaces, or paints; the detection of man-made materials in natural backgrounds for the purpose of search and rescue; the detection of specific plant species for the purposes of counter narcotics; and the detection of military vehicles for the purpose of defense and intelligence. The objective of this article is to provide a tutorial overview of detection algorithms used in current hyperspectral imaging systems that operate in the reflective part of the spectrum (0.4 - 24 μm.) The same algorithms might be used in the long-wave infrared spectrum; however, the phenomenology is quite different. The covered topics and the presentation style have been chosen to illustrate the strong couplings among the underlying phenomenology, the theoretical framework for algorithm development and analysis, and the requirements of practical applications.

Performance Prediction of Matched Filter and Adaptive Cosine Estimator Hyperspectral Target Detectors

Many applications of hyperspectral remote sensing involve the detection of subpixel targets for search and rescue or defense and intelligence operations. The design and potential capabilities of these systems depends on their target detection performance. Therefore, it is important to have tools that reliably predict the performance of target detection systems under different realistic situations. The purpose of this paper is to present a hyperspectral target performance prediction model for the widely used matched filter (MF) and adaptive cosine estimator (ACE) detectors. We use a replacement signal model for resolved and subpixel targets and a finite probability mixture of t-elliptically contoured distributions ( t-ECDs) for the background. A major contribution of this paper is the development of a robust analytical and numerical approach to determine the output distribution of ACE for mixtures of t-ECDs. The proposed technique can be a very useful tool for evaluating target detection performance for highly complex backgrounds.

Performance evaluation of cluster-based hyperspectral target detection algorithms

Detection of targets in background clutter using hyperspectral imaging sensors, is a problem of great practical interest [1]. This paper addresses some practical problems related to the adaptive estimation of clutter models and their effects on the performance of matched-signature detection algorithms. More specifically, we compare clutter estimation algorithms using spatially-local adaptation or spectral clustering to deal with the nonstationarity of hyperspectral backgrounds.

Comparison of hyperspectral change detection algorithms

There are a multitude of civilian and military applications for the detection of anomalous changes in hyper-spectral images. Anomalous changes occur when the material within a pixel is replaced. Environmental factors that change over time, such as illumination, will affect the radiance of all the pixels in a scene, despite the materials within remaining constant. The goal of an anomalous change detection algorithm is to suppress changes caused by the environment, and detect pixels where the materials within have changed. Anomalous change detection is a two step process. Two co-registered images of a scene are first transformed to maximize the overall correlation between the images, then an anomalous change detector (ACD) is applied to the transformed images. The transforms maximize the correlation between the two images to attenuate the environmental differences that distract from the anomalous changes of importance. Several categories of transforms with different optimization parameters are discussed and compared. One of two types of ACDs are then applied to the transformed images. The first ACD uses the difference of the two transformed images. The second concatenates the spectra of two images and uses an aggregated ACD. A comparison of the two ACD methods and their effectiveness with the different transforms is done for the first time.

Distributed adaptive spectral and spatial sensor fusion for super-resolution classification

A distributed architecture for adaptive sensor fusion (a multisensor fusion neural net) is introduced for 3D imagery data that makes use of a super-resolution technique computed with a Bregman-Iteration deconvolution algorithm. This architecture is a cascaded neural network, which consists of two levels of neural networks. The first level consists of sensor networks: two independent sensor neural nets, namely, a spatial neural net and spectral neural net. The second level is a fusion neural net, which contains a single neural net that combines the information from the sensor level. The inputs to the sensor networks are obtained from unsupervised spatial and spectral segmentation algorithms that can be applied to the original imagery or imagery enhanced by a proposed super-resolution process. Spatial segmentation is obtained by a mean-shift method and spectral segmentation is obtained by a Stochastic Expectation Maximization method. The decision outputs from the sensor nets are used to train the fusion net to a specific overall decision. The overall approach is tested with an experiment involving a multi-sensor airborne collection of LIDAR and Hyperspectral data over a university campus in Gulfport MS. The success of the system in utilizing sensor synergism for an enhanced classification is clearly demonstrated. The final class map contains the geographical classes as well as the signature classes.

Wavelength calibration correction for ground radiance spectra in LWIR hyperspectral imagery

Accurate retrieval of surface emissivity from long-wave infrared (LWIR) hyperspectral imaging data is necessary for many scientific and defense applications. Emissivity estimation consists of two interwoven steps: atmospheric compensation (AC) and temperature-emissivity separation (TES). AC uses an atmospheric estimate to convert the at-aperture radiance to ground radiance. Using the ground radiance, TES produces a temperature and emissivity estimate. TES algorithms require an accurate atmospheric model, and assumes that emissivity spectra for solids are smooth, compared to atmospheric features. A high-resolution atmospheric model is band-averaged to the sensors spectral response function (SRF). Characterization and maintenance of the SRF is difficult, and errors cause rough emissivity estimates. We propose a method where spectra with smooth reflective emissivities are used to correct errors from the SRF. In-Scene AC (ISAC) methods can be used to find accurate estimates of the band-averaged atmospheric upwelling and transmission, but not the downwelling radiance which is needed for TES. Typical TES methods use a model for the downwelling radiance and an assumed SRF, which will differ from the true SRF causing unnaturally rough emissivity estimates. While ISAC estimates include the true SRF it is difficult to separate the SRF from these measurements. Instead of estimating the SRF directly, our method uses smooth low emissivity materials to produce a correction for the downwelling radiance that matches the true band-averaged values. We demonstrate this technique using simulated data.

Distributed adaptive framework for multispectral/hyperspectral imagery and three-dimensional point cloud fusion

Abstract. A proposed framework using spectral and spatial information is introduced for neural net multisensor data fusion. This consists of a set of independent-sensor neural nets, one for each sensor (type of data), coupled to a fusion net. The neural net of each sensor is trained from a representative data set of the particular sensor to map to a hypothesis space output. The decision outputs from the sensor nets are used to train the fusion net to an overall decision. During the initial processing, three-dimensional (3-D) point cloud data (PCD) are segmented using a multidimensional mean-shift algorithm into clustered objects. Concurrently, multiband spectral imagery data (multispectral or hyperspectral) are spectrally segmented by the stochastic expectation–maximization into a cluster map containing (spectral-based) pixel classes. For the proposed sensor fusion, spatial detections and spectral detections complement each other. They are fused into final detections by a cascaded neural network, which consists of two levels of neural nets. The success of the approach in utilizing sensor synergism for an enhanced classification is demonstrated for the specific case of classifying hyperspectral imagery and PCD extracted from LIDAR, obtained from an airborne data collection over the campus of University of Southern Mississippi, Gulfport, Mississippi.

Performance evaluation of a remote-sensing chemical identification algorithm

The remote detection and identification of gaseous chemical plumes is an important problem with many military and commercial applications. In this paper, we consider the performance of a chemical identification system that consists of a detector bank and a model averaging algorithm. Using standard detection metrics and an identification metric, we compare the two algorithms and show that cascading the two algorithms can lead to superior performance.

Metrics for the comparative evaluation of chemical plume identification algorithms

The detection of chemical agents with hyperspectral longwave infrared sensors is a difficult problem with many civilian and military applications. System performance can be evaluated by comparing the detected gases in each pixel with the ground truth for each pixel using a confusion matrix. In the presence of chemical mixtures the confusion matrix becomes extremely large and difficult to interpret due to its size. We propose summarizing the confusion matrix using simple scalar metrics tailored for specific applications. Ideally, an identifier should determine exactly which chemicals are in each pixel, but in many applications it is acceptable for the output to contain additional chemicals or lack some constituent chemicals. A performance metric for identification problems should give partially correct results a lower weight than completely correct results. The metric we propose using, the Dice metric, weighs each output by its similarity with the truth for each pixel, thereby giving less importance to partially correct outputs, while still giving full scores only to exactly correct results. Using the Dice metric we evaluated the performance of two identification algorithms: an adaptive cosine estimator (ACE) detector bank approach, and Bayesian model averaging (BMA). Both algorithms were tested individually on real background data with synthetically embedded plumes; performance was evaluated using standard detection performance metrics, and then using the proposed identification metric. We show that ACE performed well as a detector but poorly as an identifier; however, BMA performed poorly as a detector but well as an identifier. Cascading the two algorithms should lead to a system with a substantially lower false alarm rate than using BMA alone, and much better identification performance than the ACE detector bank alone.

Hyperspectral performance prediction of the adaptive cosine estimator

The adaptive cosine estimator is a popular and effective algorithm for detecting materials in hyperspectral images. To predict the performance of this algorithm in real hyperspectral scenes, a statistical model using a mixture of multivariate t-distributions for the background and a Gaussian distribution for the target is utilized. In this paper, two methods for finding the response of the adaptive cosine estimator (ACE) and Beta-detector when applied to a statistical model. To verify that the proposed techniques work as expected, t-distribution and F-distribution quantiles are computed and compared to standard values. Finally, a preliminary validation with Monte Carlo simulation based on real hyperspectral data is presented. We build on previous work for the matched filter and extends it to use two more detectors.