|
|
Technology Introduction to Hyperspectral Remote Sensing Semi-Automatic Terrain Segmentation (TERSEG) Hyperspectral remote sensing exploits the fact that all materials reflect, absorb, and emit electromagnetic energy (photons) in ways characteristic of their molecular makeup. By measuring the intensity of these emissions, as a function of spectral wavelength, it is possible to obtain a unique spectral signature for any given material. For example, chlorophyll, which is indicative of green vegetation, will absorb energy in one very narrow part of the spectrum which differentiates it from other objects in the scene. A typical vegetation spectra is shown in figure 1. Hyperspectral sensors collect data in a portion of the electromagnetic spectrum that span wavelengths ranging from the visible (0.4 - 0.7 micro-meters) to near infrared (< 2.4 micro-meters). Data are collected in hundreds of narrow contiguous "bands" that are nominally about 10 nano-meters wide. For airborne based sensors, spatial resolution or pixel size varies from meters to tens of meters and is primarily a function of flight altitude. Data consist of "hypercubes" (or simply "cubes") where the X and Y values correspond to pixel locations and Z values correspond to the reflectance of materials in each band acquired by the sensor. An example data cube is illustrated in figure 2. Practitioners of the method are often interested in identifying the composition of objects on the surface of the earth. For example, in natural resource investigations, the principle objective might be to map surface distributions of specific minerals or rock types. Or, another possibility, would be to use hyperspectral data to identify plant types in ecosystems analyses. Other applications focus on discriminating between objects in a scene. These objectives typically arise in military and intelligence applications where the intent might be to assess target identity. The common theme in all of these applications is the need to assess whether or not a given spectral response can be assigned to a particular class. Many analysis paradigms have been derived to address this need. One currently in vogue is called the "matched filter." This filter takes a "spectral prototype" (which is the desired spectral profile of a material of interest obtained from either laboratory measurements or directly from the image) and, on a pixel-by-pixel basis, computes the cross-correlation between the spectral prototype and the test spectra under investigation. The amplitude of the resulting computation is then compared to the value of a threshold (the threshold is determined statistically which is based in a probability of false alarm) and a statistical test is invoked which either accepts or rejects the hypothesis that the test spectra belongs in the same class as the prototype. Given these results, it is possible to construct maps where pixels that pass the test are color coded in such a manner that differentiates them from other objects in the scene. Unfortunately there is a problem with applying the matched filter to hyperspectral data. Because the spatial resolution of a pixel is often greater than the physical size of objects in the scene, it is possible to have contributions to the test spectra that arise from several reflecting surfaces. This phenomenon, called a "mixed pixel" also occurs when a pixel crosses an object boundary. The creates a situation where the test spectra is hardly unique and hopes of matching this spectra to a spectrally "pure" prototype are greatly diminished. An example of a mixed pixel is illustrated in figure 3. If the pixel’s ‘footprint’ covers more than one material type (in this case vegetation and the mineral calcite) then a mixed spectra would be observed by the imaging spectrometer. Usually it is assumed that the mixed spectra is a linear combination of the pure or ‘endmember’ spectra that is weighted by the fraction of its geometric cross section. It is the analysts job to determine whether or not the desired target, in this case the vegetation, exists in the observed mixture. One approach to address this problem is to implement a matched filter that incorporates the effects of this mixing by rejecting interference from contributions to the observed spectra that are not a part of the spectral prototype. This filter, called the "Matched Subspace Detector" or MSD has this unique property. Unlike the matched filter, which has only one input (the spectral prototype), the matched subspace detector has a minimum of two inputs, i.e., the spectral prototype of the target and the spectral prototype of one or more interferences that are expected to be in the scene. For example, if the detection problem was to find the vegetation in a scene where the background consists of the calcite, the spectral prototype of the vegetation would be the target and the spectral prototypes of the calcite would be the interference. Then, if the vegetation was located in an area dominated by calcite (i.e., limestone) the interference suppression capability of the MSD would greatly enhance our ability to detect the vegetation. In fact, one desirable property of the MSD is that it is invariant to the amount of interference in the signal. That is, if the pixel was dominated by interference, the MSD would still be capable of extracting the vegetation’s spectral signature and make a positive identification. Semi-Automatic Terrain Segmentation (TERSEG) Data Fusion Corporation |
