Spectrum Software for Texas Gap Analysis
Nick C. Parker, Raymond W. Sims, Yonglun Lan, and Miguel Cano
Due to the size of Texas and the variability of conditions present, it sustains very complex and diverse vegetation. More than 90% of the land is privately owned, and access to the land for surveying vegetation is limited. A project of this magnitude, involving the analysis of satellite imagery, requires specialized computer hardware and software and highly trained personnel for processing and analyzing remotely sensed imagery.
To cope with these problems, TX-GAP has sought to optimize its capabilities by adopting the most efficient tools to automate the process to reduce time and costs. A straightforward, methodological approach has been developed for the analysis of land cover and the production of high-quality, low-cost maps in a relatively short time frame. The progress in TX-GAP has been made possible by applying the hyperclustered Landsat Thematic Mapper (TM) imagery of the Multi-Resolution Land Characteristics Consortium (MRLC) (Loveland and Shaw 1996) and the software program Spectrum (Khoral Research, Inc.), designed specifically as the analysis software for hyperclustered scenes (Benjamin et al. 1996).
MRLC Imagery and Spectrum
The MRLC represents a novel strategy established among U. S. government agencies that combined their efforts to obtain and produce standardized geographic information. This innovative partnership has considerably reduced the cost of imagery through the cooperative purchase and redistribution among MRLC members (Jennings 1996). A significant achievement of the MRLC consortium has been the production of preprocessed satellite images. In addition to the common operations performed to the Landsat TM multispectral data (radiometric and geometric corrections), these scenes are spectrally classified following a special unsupervised classification approach developed by Kelly and White (1993) resulting in a 240-class "hyperclustered" data set. A detailed description of this consortium and their project goals may be found in Benjamin et al. (1996), Campbell (1996), and Loveland and Shaw (1996) (refer also to the MRLC Web page: http://www.epa.gov/grd/mrlc).
The hyperclustering algorithm identifies 240 clusters of data, grouping sets of individual pixels having a similar spectral signature across six of the seven bands. In the resulting image, individual pixel values are representative of the mean values of the clusters produced across the six bands (240 cluster values). These clusters are linked to a statistical codebook which permits calculations to explore spectral properties of the hyperclustered scene emulating the original (raw) multispectral data. Specific clustering procedures have been reported elsewhere (Benjamin et al. 1996, Kelly and White 1993).
Spectrum is a special image visualization and analysis program developed specifically for categorization of these hyperclustered scenes. Its design and capabilities provide means for the direct interpretation of the spectral pattern observed on the scene, supported by ancillary information or ground-truthing. The use of Spectrum interpretation capabilities are explained in Benjamin et al. (1996) and Myers et al. (1995).
TX-GAP has been one of the pioneers to apply these new tools to land cover mapping. Since its initial stages, TX-GAP personnel worked together with Spectrum software developers to test the program and suggest enhancements to better adapt it to the methodology and specific requirements of Gap Analysis. Some of the suggestions pertained to the correction of programming errors, loading and reading point-location files, and other changes which have made Spectrum a more viable option for landscape-level mapping projects (Sims and Hammer 1996).
TX-GAP Land Cover Analysis Approach
In general terms, the land cover map is generated by digital classification of satellite imagery supported by field surveys and ancillary information. Accuracy assessment will involve a statistical comparison of subset samples from the classified scene to ground observations. The full process is illustrated by Figure 1.
Specifically, our digital analysis is based on the direct interpretation of MRLC hyperclustered TM scenes using Spectrum. Ground control points are used to discriminate among the different pixel cluster values on the scene, individually or by groups (following spectral pattern). After a specific polygon or set of polygons have been identified, all other polygons having the same cluster values are automatically categorized by the program.
For the field survey, we begin planning a tentative itinerary, marking potential sample sites based on a small-scale map of state highways, existing vegetation maps, or a visual examination of the spectral pattern present on the scene (displayed in Spectrum using true-color simulation capability). We do not follow any specific sampling design. Our goal is to attempt the most representative sample of the major vegetation types present in the study area (the area covered by each scene) and enough replication to detect potential variations in spectral patterns of a vegetation type across the scene.
Sampling size is not fixed; intensity depends upon the image analysts perception and confidence in the progress of the classification process. As scene interpretation progresses, field surveys are restricted to sites with a larger proportion of unclassified pixel clusters or where we are in doubt about the preliminary classes defined. Field trips conclude when 70 to 80% of the scene has been classified (for example, a total of about 50 points per scene have been necessary for the extensively agricultural Texas Panhandle, while more than 100 points have been required in the more biologically diverse Trans-Pecos region).
Sampling points are selected and georeferenced on-site along a preestablished route, considering they are observable at the scale of satellite imagery (30 x 30 m pixels) and that there will be an error factor introduced by the Global Positioning System (GPS) unit (+/- 50 m). A priority is to select vegetation patches sufficiently large and homogeneous that represent characteristic patterns in the image. We also attempt to maximize existing contrast among vegetation types to support our delineation of spectral patterns.
Because of the problem that Texas land ownership confers to site accessibility, sample points are usually within 200 m of selected roads. In addition to GPS coordinates, the distance and bearing from the road to the target communities are also recorded.
Field data includes a list of dominant and co-dominant species observed and the percentage of total cover they represent. Other described characteristics include soil texture and color, slope, aspect and comments on any environmental features that might help to differentiate a particular vegetation, or help to delineate the target community on the image. Videos or photographs are taken of the site, and a drawing is made to support the overall description. The sketch information turns out to be very important; most of the time this drawing provides the basic reference to check if the pixel clusters labeled on the satellite image represent a spatial pattern similar to that observed in the field.
Two different situations commonly occurring during the land cover classification process will illustrate the methodological approach using Spectrum. The first situation is when a spectral pattern is clearly recognizable on the satellite image, and this pattern corresponds to the field data. When this occurs, we use the GPS location (UTM coordinates) as a guide and select a pixel from the polygon area representing the target vegetation observed in the field. Here the "Zoom Window" function is used to magnify a section of the image, allowing precise selection of individual pixels of interest. Then, we use the "Class Operation" function of Spectrum to attribute the corresponding vegetation type to the pixel, and the "Legend" function to assign a specific color. All other pixels with the same values as the ones selected are automatically labeled across the scene. Additional pixels are then iteratively selected or deleted using the "Cluster Operation" function. Sometimes an atypical pixel is selected, leading to an erroneous pattern (relative to our knowledge of the region or the site description recorded in the data sheet). The pixel incorrectly labeled is then eliminated from this class. The process is finished when the labeled pattern resembles the spatial pattern observed in the field.
Second, if the spectral pattern appears inconsistent or incorrect for interpretation of some area within an image, a ground control point is selected and related to a single pixel. The patterning produced by highlighting all pixels in the same spectral cluster is then checked. A decision is made based on field data and the analysts knowledge of the region. When an erroneous patterning is produced or when in doubt, we may try using a different ground-control point, working through the labeling process in an iterative fashion.
When a field point, or a set of field points, is used to interpret different areas across a satellite image, they progressively outline those vegetation classes that have no apparent spectral pattern by aggregating the vegetation types that have been labeled. In addition, the "Spectral Response Curve" function may be used to support the interpretation by comparing graphs of those pixels that potentially belong to the same class as the one originally selected for a target vegetation type. Another tool that can also be used "on-the-fly" is the "Display" function, which emulates different band combinations and enhances particular spectral patterns.
The digital classification process is iteratively performed with the field survey. Field trips at this stage also serve to informally review progress and our confidence in the classification process by checking preliminarily labeled sections in the field. This process continues until we are satisfied with the amount of image classified (70-80%), and then we apply the "Auto-Classify Operation" function. This special feature automatically completes the classification process for a partially interpreted scene, based on the spectral similarities between labeled and unlabeled pixels. We apply this function after we are satisfied that we labeled all potential vegetation categories we want to represent for that particular scene. If we are unsatisfied with the results of Auto-Classify, we can simply go back to where we left off and continue with the labeling process.
Finally, the interpreted image with all pixel clusters labeled for the land cover classes of interest is produced and saved as a binary file. The file can be transferred to a geographic information system for further refinement and editing. For example, in the Lubbock area, we reassigned pixels that were incorrectly classified (juniper erroneously mapped as occurring in the High Plains, or orchards mislabeled as riparian). In this process, we used a set of grids with elevation information, ecoregion areas, and buffer zones for streams. A simple decision model was created to transfer pixels labeled as riparian and located outside of stream buffer zones to orchards. The same model was used to transfer pixels labeled as juniper (Juniperus spp.) in the High Plains to mesquite (Prosopis glandulosa) (Figure 2).
The availability of Landsat TM imagery through the MRLC has considerably reduced the costs for TX-GAP. Hyperclustered TM scenes have avoided the need to get involved in complex and time-consuming preprocessing and spectral classification procedures. This has been advantageous for TX-GAP, allowing the employment of ecologists lacking experience in correcting and manipulating remotely sensed raw imagery. In addition, hyperclustered data sets have reduced requirements for computer storage and consequently made our land cover analysis easier and more expedient. The use of Spectrum and the value-added MRLC data have allowed more resources to go into collecting field data.
The selection of quality ground-points during field verification (e.g., controlled by size and contrast among vegetation types sampled) has proved useful for identifying patterns on the hyperclustered images. In addition, the use of a GPS has provided confidence that clusters selected from the hyperclustered TM scene really represent what is found on the ground at the sample points.
Although no formal accuracy assessments have been implemented yet, the iterative process between field and laboratory has permitted us to progressively check the developing classified maps. Preliminary observations have shown a high correlation between the maps under preparation and the vegetation types observed on the field. Overall, Spectrum has proven to be a quick and effective tool for TX-GAP personnel mapping land cover.
We thank Robert Bradley, Carlos Villalobos, and Kay White for reviewing this manuscript. This is publication No. T-10-111 of the College of Agricultural Sciences and Natural Resources, Texas Tech University.
Benjamin, S., J.M. White, D. Argiro, and K. Lowell. 1996. Land cover mapping with Spectrum. Pages 279-288 in J.M. Scott, T.H. Tear, and F.W. Davis. Gap Analysis: A landscape approach to biodiversity planning. American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland.
Campbell, P. 1996. MRLC update and new rules for TM access: The Landsat program management agreement. GAP Analysis Bulletin 5:11-12.
Jennings, M.D. 1996. Satellite imagery, pattern delineation and cooperative mapping of land cover (Last revision: 02/96). Pages 1.1-1.5 (Chapter: Imagery) in National Biological Survey Gap Analysis Program. Handbook, Version 1 (1994). National Biological Survey, Idaho Cooperative Fish and Wildlife Research Unit, University of Idaho, Moscow, Idaho.
Kelly, P.M. and J.M. White. 1993. Preprocessing remotely sensed data for efficient analysis and classification. Pages 24-30 in Applications of Artificial Intelligence 1993: Knowledge-based systems in aerospace and industry. Proceedings: SPIE 1993. Orlando, Florida (a reprint also appeared in the appendix of the first version of the Gap Analysis Handbook).
Loveland T. R. and D. M. Shaw. 1996. Multi-resolution land characterization: Building collaborative partnerships. Pages 79-85 in J.M. Scott, T.H. Tear, and F.W. Davis. Gap Analysis: A landscape approach to biodiversity planning. American Society for Photogrammetry and Remote Sensing, Bethesda, Maryland.
Myers W., G. Thelin, A. Rasberry, S. Benjamin, J. Hood, P. Etzler, J. Majure, J. Brakebill, P. Green and J. Findley. 1995. SPECTRUM - Satellite image interpretation with automated delineation: A workshop-based assessment of SPECTRUM software. GAP Analysis Bulletin 4:10-13.
Sims, R. and R. Hammer. 1996. Status of Spectrum software (Notes). GAP Analysis Bulletin 5:49.