|
See What You've Been Missing with Subpixel Image Classification
By Kelly Roberts and Gene Roe Just like the rockets that carry the "birds" into space, the interest in using satellite imagery is finally beginning to take off, thanks in large part to the launch press that IKONOS (Space Imaging, Thornton, Colo.) and others have been generating. And the push to educate the general public about the ability of these new satellites to produce highly detailed pictures of the Earth, is beginning to pay off as well. Of course, much of the public's initial interest extends merely to finding their house in one of the scenes, or on one of the Web sites such as Terraserver or Globexplorer.
Over on the technical side, the remote sensing community has been using satellite imagery for decades, but the potential for commercial success of high-risk ventures undertaken by Space Imaging, EarthWatch (Longmont, Colo.) and ORBIMAGE (Dulles, Va.), lies in the degree to which the much larger GIS community is willing to embrace the use of the new, high-resolution imagery. Unleash the Real Value of Imagery
There is no question that one-meter scenes can provide a very informative background/basemap for many GIS applications. The real value of imagery, however, is unlocked with powerful, geographic imaging software that can find patterns, changes and phenomena which are not easily seen, or in many cases, are not visible at all to the human eye. This is when the use of satellite imagery becomes exciting, at least in terms of solving some very challenging, real-world problems.
In many applications, the key to the software's ability to analyze imagery comes from the use of color sensors in the satellites. This results in so-called multispectral images, where each pixel is characterized by a specific color signature, a set of numbers that represents the combination of the reflectance values of materials in that pixel. The satellite's sensor records the digital values for the different intensities at a given wave length of light. These numbers create each pixel's spectral signature (see Figure 1).
Remote sensing software can automatically search for, exploit and/or classify the satellite imagery according to the digital spectral properties of the pixels. This unleashes the real power of using satellite imagery-the ability to "remotely sense" changes in crops, land cover and land use, the environment, and any other spatial phenomena detected through changes in the spectral properties of the materials of interest (MOI). The Mixed Pixel
If every pixel in a satellite image were comprised of one type of material, i.e., if the spectral signature of a pixel had a one-to-one correspondence to a particular MOI, classifying materials by using remotely sensed data would be simple.
Unfortunately, this is not the case. In fact, the natural world is just the opposite because most pixels consist of a number of materials, resulting in the vast majority of the pixels in a scene being mixed, i.e., the spectral signature recorded by the satellite's sensors is the combination of the spectral properties of all materials covered by that pixel.
Further complicating the process is the fact that, due to the fractal world in which we live, mixed-pixel behavior is in all imagery, regardless of its resolution. Fractal behavior (the fact that the shape of an object in nature is derived from many smaller, internal components that have the same shape, such as a cloud) insures that, despite increases in the resolution, you will still see material variability within a pixel. Traditional Supervised Classifiers
A number of standard, supervised classification methodologies exist in the science of remote sensing. These are in the public domain and available in a number of remote sensing software packages. Most of these techniques, such as Maximum Likelihood and Minimum Distance classifiers, assume that a pixel is made up of a single material, treating the mixed or combined spectral signature for each pixel as being representative of the MOI.
In a supervised classification mode, the analyst provides the computer with pixels of known identity, called training pixels. The locations of the training-set pixels should be based on ground truth, whenever possible. The computer uses the spectral characteristics of the training pixels to identify other pixels in the image with similar characteristics. Choosing these training pixels is the key to the success of any supervised classification method. A Subpixel Approach
To overcome the inherent limitations of traditional supervised classifiers, and to tackle the mixed pixel problem, Applied Analysis Inc. (Billerica, Mass.), working closely with ERDAS Inc. (Atlanta, Ga.), chose a subpixel approach to resolve these issues. The collaboration resulted in the development of the IMAGINE Subpixel Classifier, a fully integrated add-on module to ERDAS IMAGINE, that can detect specific materials that make up as little as 20 percent of a pixel.
The design of the IMAGINE Subpixel Classifier is based on the premise that the signature of each pixel is made up of two spectral components-MOI and everything else, or what is collectively termed "background." The one exception to this is in the rare case that a pixel is found to be pure, meaning its signature matches that of the MOI, i.e., it does not have a background.
In order to perform a supervised classification of a multispectral image using the IMAGINE Subpixel Classifier, the software first determines which one spectrum is most common to all of the training-set pixels. This is taken as the signature of the MOI.
Once the IMAGINE Subpixel Classifier knows what it is looking for, essentially every other spectrum in the scene that differs from the MOI is assumed to represent a possible background material signature. The IMAGINE Subpixel Classifier then analyzes each pixel and subtracts varying types and percentages of background spectra, with the goal of identifying a remaining spectrum that matches the desired MOI signature.
The remaining spectrum is compared to the desired signature. If it matches within an acceptable tolerance, you have found a pixel that contains what you are looking for, and you also know the percentage of MOI in that pixel. If it does not match, then the pixel does not contain a detectable level of the MOI. Finding Impervious Land Cover
As anyone in local government will quickly tell you, identifying and, more importantly, accurately quantifying impervious land cover can be a very labor intensive (and costly) process. However, the information becomes critical input for urban land-use master planning, watershed analysis, storm water management, and in the overall assessment of urban growth.
With the cost of satellite imagery decreasing-most notably that of Landsat Thematic Mapper-a growing number of people are investigating the use of remotely sensed data as an alternative to more traditional, hands-on techniques.
Of course, one of the key challenges with this remote sensing approach is that urban areas contain a large number of mixed pixels due to the highly variable nature of the urban landscape. As noted earlier, this will be true whether you are working with 30-meter pixels, as is the case with Landsat TM, or a high resolution four-meter IKONOS image.
One of the few proven and powerful ways to "unmix" these urban pixels and determine the quantity of impervious surface, is with the IMAGINE Subpixel Classifier. In a recent, detailed impervious land-use study, the IMAGINE Subpixel Classifier was applied to a Landsat TM scene containing a portion of downtown Charleston, S.C. (Ji and Jensen 1999). In order to address the diversity of urban features in the scene, the investigators decided to develop multiple subpixel signatures-one for buildings and three for different pavement types.
These signatures were then used, in multiple passes, to identify which pixels contained the materials of interest, and more importantly, to quantify the percentage of impervious surface found in all pixels, down to as little as 20 percent ( see Figure 2, noting the small white areas on the east coast of Charleston, S.C.).
These output image maps were next combined with a single image map from a layered classification procedure, which identified difficult-to-classify, high-density urban pixels (Ji and Jensen 1999) (see Figure 3, noting that the small white areas from Figure 2 are now filled in with the 90-100 percent impervious class). A layered classification map was created using rules regarding the brightness of the pixels. These formed the basis of a decision tree to guide the selection of highly impervious pixels.
To assess the accuracy of the final output, the authors used two 1:40,000-scale NAPP (National Aerial Photography Program) color infrared aerial photographs. Once scanned and registered with the Landsat TM image, 200 randomly generated points were sampled to act as ground truth for assessing the overall accuracy of the study.
The output from the layered classifier was assigned to the 90 to 100 percent impervious class and then merged with the four outputs from the IMAGINE Subpixel Classifier to form the final impervious land cover map.
These results were then compared to the land-cover types manually derived for the ground truth sample points. The IMAGINE Subpixel Classifier was found to have an overall accuracy of 96.5 percent in detecting the presence or absence of urban imperviousness in a Landsat TM scene. More Spectral Imagery on the Horizon
With the anticipated launch of a host of new spectral sensors over the next few years, the number one constraint on the growth of the remote sensing industry, i.e., data availability, will soon be overcome. Perhaps the most exciting aspect of this news is that a number of these new sensors will be capable of sensing and collecting many more narrowly defined bands of radiation.
Although not a replacement for subpixel analysis, these so-called hyperspectral sensors offer the promise of much more precise discrimination of material signatures which, in a number of ongoing research projects, have been demonstrated to be capable of identifying oil seeps, mineral deposits, toxic gas plumes and phytoplankton, to name but a few.
Armed with these powerful hyperspectral sensors, the impending increase in the supply of spectral imagery, and the improving sophistication of analysis tools, the remote sensing industry is poised to make a significant contribution to the growth of the geotechnologies marketplace, as well as the future health and well being of our planet. About the Author:
Kelly Roberts is the GIS manager for ViewPoint Engineering. She may be reached by phone (+1 781-933-9366) or via e-mail ([email protected]). Gene Roe is the business development manager for Applied Analysis, Inc. He may be reached by phone (+1 978-663-6828) or via e-mail ([email protected]). Visit these Web sites for more information about the IMAGINE Subpixel Classifier: www.discover-aai.com and www.erdas.com. References:
Ji, Minhe and Jensen, John R. Effectiveness of Subpixel Analysis in Detecting and Quantifying Urban Imperviousness from Landsat Thematic Mapper Imagery. Geocarto International, Vol. 14, No. 4, December 1999. Be Careful When Ordering Satellite Imagery
When satellite imagery is first collected and recorded, it is in its "raw" form. Various pre-processing steps must be specified and applied, usually by the data provider, before the imagery is ready for any kind of digital processing or exploitation, including classification.
Pre-processing includes correcting for systematic, radiometric, and geometric errors. These corrections involve shifting pixels into proper alignment and removing artifacts such as banding, striping, and dropped lines or pixels in the scene.
The type of process used for these corrections is often a matter of personal preference. Therefore, be sure to investigate the pre-processing options available before settling on the defaults of the data provider. In other words, you need to find out what is being done to the data before you buy it.
When using the IMAGINE Subpixel Classifier, for example, the type of image orientation and the type of resampling used to project the imagery parameters are important. Path-oriented imagery is preferred over map-oriented imagery, and nearest-neighbor resampling is preferred over cubic convolution (often the default) for geometric correction. Path-oriented and nearest-neighbor resampled imagery is preferred because it retains purer pixels in terms of the digital numbers compared to other methods. Original digital numbers can provide better results especially when the material of interest occupies less than 90 percent of a pixel.
Back
|
