 |
 |
 |
 |
| |
|||||
|
|
|
 Analyzing Hyperspectral Data By Edward Ashton, Brian Fianagan, and Sherry Olson
New data-analysis algorithms are able to convert vast amounts of data generated by hyperspectral sensors into useful products. In the past, this has been an extremely labor-intensive effort. Hyperspectral data is collected by sensors at hundreds of distinct wavelengths. The corresponding spectral signature or fingerprint provides a direct indicator of the material being imaged and therefore has obvious applications for battlefield surveillance, hazardous waste identification, agriculture, and prospecting (mineral identification). Three MITRE programs are developing new methods to process hyperspectral data: anomaly detection, target identification, and spectral library tools. Anomaly detection is used to identify peculiarities in an image when the scene under surveillance is unknown. Target identification is employed when the spectral signature of a material is known, and requires access to reference spectral signatures. Novel feature extraction and hierarchical data structures are used as tools to construct databases for more efficient use of spectral reference libraries. Anomaly Detection and Illumination Suppression Anomaly detection is commonly used when nothing is known about the scene under surveillance, but we wish to identify anything that might be out of place; for instance, a vehicle in a forest clearing or an oil or chemical spill in an otherwise uncontaminated area. Anomaly detection algorithms flag any suspicious areas in a scene and pass those cues to an image analyst, a spectral analysis system, or another sensor. The idea behind anomaly detection is simple. A mathematical model of the background is formed based on a series of observations taken before or during the surveillance operation. As each incoming spectrum is received, it is compared to the model's prediction. The output of the anomaly detector at each pixel in a scene is then a probabilistic measure of that pixel's deviation from its predicted value. In this way, pixels that fail to conform to the background model's expectations are flagged for further analysis. In order for hyperspectral data to be useful for either anomaly detection or material identification, it is necessary to remove the effects of varying illumination from a scene. We have developed a two-step process to accomplish this. First, the global path radiance in the scene is estimated and removed. Path radiance is the energy that is due entirely to photons that have been reflected back to the sensor by the atmosphere. Once the path radiance has been removed, the data is passed through a transform that isolates the energy component related to illumination from the components related to spectral shape, or "color". The illumination component is then discarded.  (A) One band from a hyperspectral scene. Much of the bottom half of the image is lost through cloud shadowing. (B) The same scene after illumination suppression. Shadowing effects are almost entirely gone, and the details of the bottom half of the image are now visible. (C) A subsection of (A), where a feature is marginally visible in the center of the image, but shadowing makes it impossible to distinguish. (D) The same subsection, but here the feature is clearly identifiable as a utility pole. Moreover, the individual utility lines are visible despite the fact that these wires occupy only a small fraction of the pixels they intersect. Target Identification Target identification is employed when we know the spectral signature of a material of interest. One application of this, commonly used for environmental monitoring, is the detection of specific chemical effluents in a gas plume. A traditional method used for target identification is to process the data using a matched filter based on a target spectrum of interest. All pixels whose spectra match the target spectrum (to a specified level of confidence) are marked as potential targets. The underlying assumption with this approach is that the pixels containing the target are "pure" that is, the material we are interested in fills the entire pixel and is not mixed with any background material. Unfortunately, a gas plume may be transparent, so any pixel containing the plume will be a mixture of the plume and the ground beneath it. The pixel spectra in this case will not resemble the target spectrum unless the background material can somehow be suppressed. Building on a manually intensive orthogonal filter algorithm, a new approach was devised in the ESL to automatically characterize and suppress the background spectra underneath the plume. Starting with the target spectrum, the algorithm iteratively processes the scene and selects the spectrum that is most unlike the target spectrum and any previously selected background spectra. It then generates a filter that matches the target spectrum but is orthogonal (statistically independent) to all the identified background spectra. The resulting filter, when applied to the scene, identifies only the plume pixels and suppresses all other pixels. Spectral Library Tools Ultimately, it would be useful to identify the materials in every pixel in a scene; however, doing so would require access to representative spectral signatures for every known material. A database holding all these spectra could conceivably contain millions of spectral signatures. The goal of the Smart Spectral Libraries research project at MITRE is to develop methods to characterize, organize, and search spectral library databases for implementation in operational systems. To date, we have developed three techniques to characterize spectral signatures. The first technique, wavelet analysis, provides a compact description of the general shape of the signature. The second technique, binary encoding, when applied to regions of the curve, converts the signature to a simple two-level representation for each region. That provides a coarse measure of the major features embedded in the signature. The final technique, feature analysis, measures fine detail about all the significant features in the signature, including position, width, depth, and asymmetry. The results of the signature characterization are then used to group signatures into a hierarchical structure. An unknown spectrum is identified by comparing it to a very general class of signatures and then to increasingly smaller and more tightly clustered groups of signatures. In this way one can rapidly traverse a large database and identify the spectral signature most similar to the unknown spectrum. That provides a significant advantage in speed over an exhaustive search of the entire database. Summary Hyperspectral data analysis offers a tool for deriving useful information for many difficult problems. However, there are challenges associated with processing of hyperspectral data that make it unwieldy to use in an operational environment, including the sheer volume of data and the computer-intensive nature of the data analysis. The three programs discussed above (anomaly detection, target identification, and spectral library tools) illustrate a few of the analytical methods being investigated by MITRE for faster, more accurate processing of hyperspectral data. For all of these methods our goal is the same, to bring practical and timely exploitation of hyperspectral data one step closer to reality. For more information, please contact Sherry Olson using the employee directory. |
Solutions That Make a Difference.® |
|
|
