| Remote Sensing Applications: Demystifying the Myth - Part II
By Skip Maselli Remote Sensing Systems to Consider
Satellite imagery is almost always captured digitally. However, the Russians have their KVR-1000 system, which is high-resolution film-based satellite imagery. The French government manages the SPOT series of satellites, which employs a push-broom scanning system that "paints" a 3000- to 6000-pixel-wide swath as it orbits the Earth. SPOT acquires four bands in the visible and IR ranges. Landsat 7 acquires three visible bands and four IR bands, including one that is thermal. Space Imaging IKONOS imagery includes a one-meter panchromatic product and a four-meter multi-spectral product (blue, green, red, near-IR). The IRS-1C is a five-meter panchromatic system that is quite popular, especially when merged with multi-spectral imagery. Several companies resell and integrate the Indian IRS product, including the Carterra product line by Space Imaging. NASA and its partners also employ numerous systems ranging from 10-meter, multi-spectral systems to 5km-resolution systems. Other companies developing or already running high-resolution satellite imagery systems include Orbimage and EarthWatch. ImageSat's EROS systems (including the 1.8-meter panchromatic EROS A1) are also expanding their systems' capabilities. EROS A1 was launched on December 5, 2000. Both airborne and satellite imagery can be acquired from a number of third-party vendors. Other commercial airborne sensors include the ADAR 5500 by Positive Systems, and sensors by TASC.  Figure 11. Georectification of one image to another
Resources for Processing Imagery
While processing imagery can be fascinating, it is also time-consuming. The idea is to "get away from the imagery" as fast as possible. Get the information required in the format that it is needed, and then pull it all together within a GIS. This may mean digitizing new vectors, adjusting existing GIS layers, classifying land-cover categories, or extracting the locations or spatial extents of interpreted features. Very often, the imagery itself is not the final product and may never even be seen by the data consumers.
Image processing is the method by which remotely sensed data is processed. Most software (for most applications) is as easy as a web browser to operate. The following are among the most referenced and reviewed packages: ERDAS IMAGINE, ERMapper, PCI EASI/PACE, TNTMips by MicroImages, Intergraph Image Analyst, and IDL and ENVI by RSI. Other more GIS-oriented packages with nominal image processing capabilities include ESRI's ArcInfo (GRID) and ArcView Image Analysis, IDRISI, and a few others. Visit these companies' web sites and ask about their pricing. Talk to others who have experience in image processing. Avoid the software resellers at first, who may promote premium add-on modules or special functions that are not truly necessary to accomplish one's objectives.
Image processing includes the formatting, projection, enhancement and manipulation of imagery content in order to extract information. This information may be in the form of a vector data set, or it could be a thematic image. Formatting
Formatting implies file-type organization and compression. Various software products produce, display or translate compressed imagery. With advancing desktop workstations and remote sensing systems, 500MB to 100GB image inventories (or even single images) are not uncommon. Consequently, image-compression applications are becoming critical, especially as we begin "serving" imagery across the Web. LizardTech's MrSID and ERMapper's ECW are two primary dedicated-compression software packages. Many other compression formats are supported in image processing software. Image Rectification and Projection
Remote sensing data should have an associated geographic projection. Each pixel is simply assigned a coordinate position based upon its row and column within the image matrix. An image can be assigned a coordinate system by being registered to another image, to a vector dataset, or to a set of survey coordinates. In any case, there is a source coordinate system and a reference system. In Figure 11, we are registering a color IR image to a panchromatic image. Control points are used to georectify one image to the other. Results could be a rotated and rescaled image.
Changing a projection or rectifying an image requires an alteration of image content. All imagery is subject to distortions due to camera and sensor system interior functions, terrain, curvature of the Earth, and sensor position. The process of correcting digital imagery through rotation, translation and scaling of pixel positions requires a resampling of the data. Resampling alters pixel values in order to best represent the correction in position of an object in the image. The value may change by simply moving the old pixel (value) to a new pixel location, using simple techniques like Nearest Neighbor resampling. (Figure 12) More robust techniques result in a smoother appearance to the corrected image, but the pixel values are more severely altered. This alteration of pixel values (from their true original values) may have a negative impact on the spectral analysis of the image.
Rubber-sheeting is the stretching or warping of a single image of remotely sensed data in different areas to varying degrees, depending upon distribution of spatial errors in the image. The order of transformation in these types of rectification procedures is also important. The higher the order of transformation, the more ground control points are required, and the more forceful is the warping. If just a few ground control points are grossly inaccurate, a high-order transformation can actually warp "correct" areas of the image to "bad areas" in the image.
Ortho-rectification is not the same as semi-automated rectification or rubber-sheeting. It is a rigorous (and more expensive) photogrammetric process that requires input of stereo imagery, sensor phenomenology, and redundancy in various correction parameters. Ortho-rectification removes distortion due to sensor position, terrain elevation, Earth curvature, and other affects. True ground positions may be extracted from ortho-rectified images. Imagery Enhancement
Imagery enhancement is generally categorized into spectral or spatial applications. Spectral enhancement changes the pixel values in order to improve visual content, remove errors, and enhance subtle spectral signatures of like objects on the Earth. Spectral enhancement has a fundamental dependence on the statistics represented by the matrix of pixels that comprise the image. These statistics are based primarily on the image histogram. (Figure 13)
The image histogram is used to improve contrast and brightness in an image and other spectral properties. The relationship between histograms of multiple layers (different wavelengths) is also compared to reveal other information "hidden" in the image. The images in Figure 14 demonstrate how a histogram manipulation may enhance vegetation information on a golf course, or urban data within shadow areas.
Spatial data enhancement uses frequency information in an image to sharpen or smooth detail in the image. (Figure 15) The panchromatic image here shows a high-pass filter that makes edges crisper within an image. These edges can also be suppressed with a low-pass filter.
Other spatial enhancements actually merge high-resolution and lower-resolution datasets. Remote sensing applications use this approach to combine the high spatial resolution of one image with the high spectral resolution of its complementary imagery. Space Imaging uses these techniques to produce pan-sharpened color IR images. In the illustration above, we show a one-meter color IR image by Positive Systems Inc., sharpened with precisely registered and ortho-rectified one-half-foot-pixel aerial photography. The results are dramatic and produce a useful image for interpretation and classification. (Figure 16)
Both spatial and spectral enhancement applications alter pixel values, resolution, and the geospatial integrity of an image. It is important to consider this when engaging in remote sensing applications. Image Classification
Image classification categorizes each pixel (or combination of pixels from multiple bands) into a discrete value based upon the actual land-cover class represented in the pixel(s). As stated above, three MSI layers alone can have more than 16 million possible values per pixel area. In the case of four- or even seven-layer MSI images, the possible combinations are astronomical in number. While images can have from hundreds to hundreds of millions of possible values and combinations of values, image classification puts all the pixels into a small (finite) set of values. This forms a single-layer thematic image; that is, each pixel is classified into a finite number of themes. Because a thematic image is usually eight bits deep, a classified image could have up to 256 categories. However, there are usually many fewer categories. For example, all pixels might be classified as one of ten possible categories as shown in Figure 19.
Once the image-processing application has been "trained" to statistically understand land-cover categories, a classifier analyzes every pixel in the image and assigns it a new pixel value in the thematic image. The pixel number (grid code) is indicative of the land-cover class. The classified image shown below illustrates the difference between raw remote sensing data and a classified image. The thematic image can then be vectorized, resulting in a vector layer identical to those used in most GIS databases. The example shown here depicts ArcInfo polygon coverage derived by a raster-to-vector conversion algorithm run in ERDAS IMAGINE. A 75-foot buffer is then used to "cookie-cut" the imagery-derived vector coverage, and provides quantitative results for how much of each land-cover class may be impacted by a right-of-way development. (See Figures 17 through 24 for illustration thereof)
Enhancing imagery involves these primary generic functions: contrast stretching, histogram manipulations, band ratios, image algebra and convolutions, and spatial filtering. These tools have been packaged into easy-to-understand graphical user interfaces. This isn't a recommendation to buy first and then push a number of buttons in various patterns until one likes what comes out; this is precisely what our snake oil salesmen hope will be done. For help, here are some good educational sites to visit:
¥ The ASPRS Remote Sensing Core Curriculum at www.research.umbc.edu/~tben ja1/
¥ The Remote Sensing Tutorial at the Canadian Centre for Remote Sensing (CCRS) at www.ccrs.nrcan.gc.ca/ccrs/eduref/educate.html
¥ NASA's Remote Sensing Tutorial at http://rst.gsfc.nasa.gov/
¥ For military buffs, or those interested in that sort of thing, the FAS site has some interesting links that may be helpful at http://sun00781.dn.net/irp/imint/index.html.
The ability to distinguish between jack pine, white pine and quaking aspen is inherent in many imagery types, and is further exposed by image processing. It is also feasible to discern the spectral separation between peas and alfalfa crops from an imaging satellite 600km away. However, the best image-processing software will not support clear and crisp distinctions. Furthermore, even the best remote sensing expert cannot manipulate these components to achieve 100 percent accuracy. There is statistical confusion among the spectral properties of many land-cover categories. For example, bare soil is often confused with asphalt, roofs are confused with sand, etc. This misclassification of land cover varies in severity but is still an improvement over what the unaided eye can interpret. Furthermore, different wavelengths may provide the improved ability to distinguish among such confused categories.
The ability to employ technology and applications to one's best advantage resides in the quality of a company's staff, the imagery acquired, and the software and hardware components employed. Conclusion
Remote sensing applications are not just for printing another "pretty face in the crowd of glossy E-size plots." The data can be readily and economically managed in order to make quantitative contributions to one's GIS. The cost of implementing remote sensing ranges from $1,000 software packages to $2,500 add-on modules, to $10,000 professional systems, and even beyond. Imagery acquisition costs can range from $500 per square mile to $200 (for higher-resolution systems-one- to five-meter MSI) to seven cents per mile for Landsat or SPOT imagery. As far as spectral and spatial resolution is concerned, remember the adage that "twice as small can cost four times as much." More importantly, consider that the most important resource for remote sensing applications is one's staff.
The secrets have hopefully been revealed. So the next time a man in a pinstriped suit yells from a stage, "Step right up and witness the revolutionary remedy for everything from the common cold to the plague," be sure to ask about extra fees for cloud-free data. Of course, steer clear of the dancing bears that are drinking sarsaparilla in the moonlight. About the Author:
Skip Maselli is western director of regional operations for InfoTech Enterprises Inc., Tucson, Ariz. He may be reached via e-mail at [email protected]. Back |
