Narrow Results

Detection of gaseous effluent plumes from airborne platforms provides a unique challenge to the remote sensing community. The measured signatures are a complicated combination of phenomenology including effects of the atmosphere, spectral characteristics of the background material under the plume, temperature contrast between the gas and the surface, and the concentration of the gas. All of these quantities vary spatially further complicating the detection problem. In complex scenes simple estimation of a “residual” spectrum may not be possible due to the variability in the scene background. A common detection scheme uses a matched filter formalism to compare laboratory-measured gas absorption spectra with measured pixel radiances. This methodology can not account for the variable signature strengths due to concentration path length and temperature contrast, nor does it take into account measured signatures that are observed in both absorption and emission in the same scene. We have developed a physics-based, forward model to predict in-scene signatures covering a wide range in gas / surface properties. This target space is reduced to a set of basis vectors using a geometrical model of the space. Corresponding background basis vectors are derived to describe the non-plume pixels in the image. A Generalized Likelihood Ratio Test is then used to discriminate between plume and non-plume pixels. Several species can be tested for iteratively. The algorithm is applied to airborne LWIR hyperspectral imagery collected by the Airborne Hyperspectral Imager (AHI) over a chemical facility with some ground truth. When compared to results from a clutter matched filter the physics-based signature approach shows significantly improved performance for the data set considered here.

In this paper we assess the effect that clustering pixels into spectrally-similar background types, for example, soil, vegetation, and water in hyperspectral visible/near-IR/SWIR imagery, prior to applying a detection methodology has on material detection statistics. Specifically, we examine the effects of data segmentation on two statistically-based detection metrics, the Subspace Generalized Likelihood Ratio Test (Subspace GLRT) and the Adaptive Cosine Estimator (ACE), applied to a publicly-available AVIRIS datacube augmented with a synthetic material spectrum in selected pixels. The use of synthetic spectrum-augmented data enables quantitative comparison of Subspace-GLRT and ACE using Receiver Operating Characteristic (ROC) curves. For all cases investigated, Receiver Operating Characteristic (ROC) curves generated using ACE were as good as or superior to those generated using Subspace-GLRT. The favorability of ACE over Subspace-GLRT was more pronounced as the synthetic spectrum mixing fraction decreased. For probabilities of detection in the range of 50-80%, segmentation reduced the probability of false alarm by a factor of 3–5 when using ACE. In contrast, segmentation had no apparent effect on detection statistics using Subspace-GLRT, in this example.

Region-of-interest cueing by hyperspectral imaging systems for tactical reconnaissance has emphasized wide area coverage, low false alarm rates, and the search for manmade objects. Because they often appear embedded in complex environments and can exhibit large intrinsic spectral variability, these targets usually cannot be characterized by consistent signatures that might facilitate the detection process. Template matching techniques that focus on distinctive and persistent absorption features, such as those characterizing gases or liquids, prove ineffectual for most hard-body targets. High-performance autonomous detection requires instead the integration of limited and uncertain signature knowledge with a statistical approach. Effective techniques devised in this way using Gaussian models have transitioned to fielded systems. These first-generation algorithms are described here, along with heuristic modifications that have proven beneficial. Higher-performance Gaussian-based algorithms are also described, but sensitivity to parameter selection can prove problematical. Finally, a next-generation parameter-free non-Gaussian method is outlined whose performance compares favorably with the best Gaussian methods.

We have developed a hyperspectral deconvolution algorithm that sharpens the spectral dimension in addition to the more usual across-track and along-track dimensions. Using an individual three-dimensional model for each pixel's point spread function, the algorithm iteratively applies maximum likelihood criteria to reveal previously hidden features in the spatial and spectral dimensions. Of necessity, our solution is adaptive to unreported across-track and along-track vibrations with amplitudes smaller than the ground sampling distance. We sense and correct these vibrations using a combination of maximum likelihood deconvolution and gradient descent registration that maximizes statistical correlations over many bands. Test panels in real hyperspectral imagery show significant improvement when locations are corrected. Tests on simulated imagery show that the precision of relative corrected positions improves by about a factor of two.

A feature reduction technique is proposed for the hyperspectral (HS) data classification problem. The new features have been developed through a curve fitting step which fits specific rational function approximations to every spectral response curve (SRC) of HS image pixels. Then, the coefficients of the numerator and denominator polynomials of these fitted functions are considered as new extracted features. The method concentrates on the geometrical nature of SRCs and is utilizing the information that exists in sequence discipline — ordinance of reflectance coefficients in SRC — which has not been addressed by many other statistical analysis based methods. Maximum likelihood (ML) classification results show that the proposed method provides better classification accuracies compared to some basic and state-of-the-art feature extraction methods. Moreover, the proposed algorithm has the capability of being applied individually and simultaneously to all pixels of image.

Hyperspectral image resolution offers limited spectral bands within a continual spectral spectrum, creating one of the spectra of most pixels inside the sequence which contains huge volume of data. Data transmission and storage is a challenging task. Compression of hyperspectral images are inevitable. This work proposes a Hyperspectral Image (HSI) compression using Hybrid Transform. First the HSI is decomposed into 1D and it is clustered and tiled. Each cluster is applied with Integer Karhunen–Loeve Transform (IKLT) and as such it is applied for whole image to get IKLT bands in spectral dimension. Then IKLT bands are applied with Integer Wavelet Transform (IDWT) to decorrelate the spatial data in spatial dimension. The combination of IKLT and IDWT is known as Hybrid transform. Second, the decorrelated wavelet coefficients are applied to Spatial-oriention Tree Wavelet (STW), Wavelet Difference Reduction (WDR) and Adaptively Scanned Wavelet Difference Reduction (ASWDR). The experimental result shows STW algorithm using Hybrid Transform gives better PSNR (db) and bits per pixel per band (bpppb) for hyperspectral images. The comparison between STW, WDR and ASWDR with Hybrid Transform for Indian Pines, Salinas, Botswana, Botswana and KSC images is experimented.

An overview of three techniques developed by our group for imaging superficial tissue is presented. Firstly, a novel polarized light capillaroscope has been developed for imaging the microcirculation. The capillaroscope has been used to make in vivo measurements of sickle cell disorder sufferers with aim of monitoring the polymerization of sickled red blood cells. Secondly, hyperspectral imaging for measuring oxygen saturation is described. The accuracy of such measurements is affected by the non-linear relationship between scattering and absorption and it is demonstrated that polarization techniques can be used to make the relationship more linear, thus improving accuracy. Finally, the use of smart CMOS optical sensors for laser Doppler blood flowmetry is described. A 32 × 32 pixel imaging array with on-chip processing is described and the potential for full field laser Doppler blood flow imaging is demonstrated through measurement on blood flow of tissue before and after occlusion.

Detection of gaseous effluent plumes from airborne platforms provides a unique challenge to the remote sensing community. The measured signatures are a complicated combination of phenomenology including effects of the atmosphere, spectral characteristics of the background material under the plume, temperature contrast between the gas and the surface, and the concentration of the gas. All of these quantities vary spatially further complicating the detection problem. In complex scenes simple estimation of a “residual” spectrum may not be possible due to the variability in the scene background. A common detection scheme uses a matched filter formalism to compare laboratory-measured gas absorption spectra with measured pixel radiances. This methodology can not account for the variable signature strengths due to concentration path length and temperature contrast, nor does it take into account measured signatures that are observed in both absorption and emission in the same scene. We have developed a physics-based, forward model to predict in-scene signatures covering a wide range in gas / surface properties. This target space is reduced to a set of basis vectors using a geometrical model of the space. Corresponding background basis vectors are derived to describe the non-plume pixels in the image. A Generalized Likelihood Ratio Test is then used to discriminate between plume and non-plume pixels. Several species can be tested for iteratively. The algorithm is applied to airborne LWIR hyperspectral imagery collected by the Airborne Hyperspectral Imager (AHI) over a chemical facility with some ground truth. When compared to results from a clutter matched filter the physics-based signature approach shows significantly improved performance for the data set considered here.

In this paper we assess the effect that clustering pixels into spectrally-similar background types, for example, soil, vegetation, and water in hyperspectral visible/near-IR/SWIR imagery, prior to applying a detection methodology has on material detection statistics. Specifically, we examine the effects of data segmentation on two statistically-based detection metrics, the Subspace Generalized Likelihood Ratio Test (Subspace GLRT) and the Adaptive Cosine Estimator (ACE), applied to a publicly-available AVIRIS datacube augmented with a synthetic material spectrum in selected pixels. The use of synthetic spectrum-augmented data enables quantitative comparison of Subspace-GLRT and ACE using Receiver Operating Characteristic (ROC) curves. For all cases investigated, Receiver Operating Characteristic (ROC) curves generated using ACE were as good as or superior to those generated using Subspace-GLRT. The favorability of ACE over Subspace-GLRT was more pronounced as the synthetic spectrum mixing fraction decreased. For probabilities of detection in the range of 50-80%, segmentation reduced the probability of false alarm by a factor of 3–5 when using ACE. In contrast, segmentation had no apparent effect on detection statistics using Subspace-GLRT, in this example.

Region-of-interest cueing by hyperspectral imaging systems for tactical reconnaissance has emphasized wide area coverage, low false alarm rates, and the search for manmade objects. Because they often appear embedded in complex environments and can exhibit large intrinsic spectral variability, these targets usually cannot be characterized by consistent signatures that might facilitate the detection process. Template matching techniques that focus on distinctive and persistent absorption features, such as those characterizing gases or liquids, prove ineffectual for most hardbody targets. High-performance autonomous detection requires instead the integration of limited and uncertain signature knowledge with a statistical approach. Effective techniques devised in this way using Gaussian models have transitioned to fielded systems. These first-generation algorithms are described here, along with heuristic modifications that have proven beneficial. Higher-performance Gaussian-based algorithms are also described, but sensitivity to parameter selection can prove problematical. Finally, a next-generation parameter-free non-Gaussian method is outlined whose performance compares favorably with the best Gaussian methods.

We have developed a hyperspectral deconvolution algorithm that sharpens the spectral dimension in addition to the more usual across-track and along-track dimensions. Using an individual threedimensional model for each pixel's point spread function, the algorithm iteratively applies maximum likelihood criteria to reveal previously hidden features in the spatial and spectral dimensions. Of necessity, our solution is adaptive to unreported across-track and along-track vibrations with amplitudes smaller than the ground sampling distance. We sense and correct these vibrations using a combination of maximum likelihood deconvolution and gradient descent registration that maximizes statistical correlations over many bands. Test panels in real hyperspectral imagery show significant improvement when locations are corrected. Tests on simulated imagery show that the precision of relative corrected positions improves by about a factor of two.

Machine learning (ML) approaches as part of the artificial intelligence domain are becoming increasingly important in multispectral and hyperspectral remote sensing analysis. This is due to the fact that there is a significant increase in the quality and quantity of the remote sensing sensors that produce data of higher spatial and spectral resolutions. With higher resolutions, more information can be extracted from the data, which require more complex and sophisticated techniques compared to the traditional approaches of data analysis. Machine learning approaches are able to analyse remote sensing (RS) data more effectively and give higher classification accuracy. This review will discuss and demonstrate some applications of machine learning techniques in the processing of multispectral and hyperspectral remote sensing data. Future recommendations will also be given to highlight the way forward in the use of machine learning approaches in optical remote sensing data analysis.