Volume No. 12, 2003/2004
Land Cover
The state Gap Analysis projects each have developed approaches to image interpretation. In Hawaii, most of the land area has been surveyed at some time, and detailed vegetation classifications are available for more than a dozen significant areas covering portions of each island.
GAP led the development of the Multi-Resolution Land Cover Consortium (MRLC), and most GAP state projects enjoy access to the MRLC archive (Hegge et al. 2001). In most cases this includes Landsat Thematic Mapper 7 images available for each path/row representing three seasons. Under the MRLC these images are preprocessed to standardize geographic location as well as correct for terrain displacement and atmospheric reflection. Additionally, the EROS Data Center now makes MRCL images available that have been corrected to “at-satellite radiance values” (Homer and Hegge, EROS Data Center/Raytheon). The process also employs the Sun-Earth and Earth-radiometer distances at the time the image was taken to compensate for the radiometric distortion effects of the Earth’s atmosphere, making images taken on different revolutions more comparable. The Hawaiian entry for MRLC “at-satellite radiance” images had not been populated prior to the HI-GAP effort, and we were able to partner with EROS Data Center to select and process scenes for each path/row representing seasonality as well as completing a cloud-free mosaic using Landsat TM images from 12/99-12/02. These scenes represent a consistent data set on which the HI-GAP spectral decision tree classification was implemented.
Several image interpretation methods were tested for the HI-GAP application. Classification and Regression Tree Analysis was considered for its objectivity and statistical strength (De’ath and Fabricius 2000, Hansen et al. 1996, Lawrence and Wright 2001), but this approach requires a significant investment in field data collection, and the majority of the land area in Hawaii has been previously surveyed. Hawaiian vegetation systems are relatively well-studied and have been mapped and classified several times previously. Detailed vegetation maps are available for significant portions of many of the Hawaiian Islands. Additionally, previous and concurrent land cover research has led to a significant spectral signature library for Hawaiian native and invasive vegetation types. Research at The University of California, Santa Barbara has focused on employing aspects of spectral signatures to perform classifications on AVRIS hyperspectral imagery (Roberts et al. 1998, Serrano et al. 2000). But AVRIS imagery is not available statewide, and Landsat 7 TM data does not have sufficient spectral resolution to enable a library/signature-based approach under the conditions and needs of HI-GAP. A more ecologically driven classification approach has been developed at Duke University’s Nicholas School of the Environment, where radiometric enhancements are employed to enable classification based on “natural” variables such as level of vegetation or soil exposure (Khorram et al. 1992). Also, three recent case studies developed for mapping impervious surfaces from Landsat 7 ETM were consulted for their possible applicability in land cover mapping approaches for HI-GAP (Yang et al., in review).
After extensive research on different methods for land cover classification in Hawaii, the HI-GAP team chose to employ an ecologically driven spectral decision tree approach to land cover classification. The approach is based on the application of ERDAS Imagine’s Knowledge Engineer software platform. Knowledge Engineer was selected because it provided the hierarchical structure to perform image classification and offered a good platform for storing and analyzing spectral properties. Knowledge Engineer files were developed for each image and then integrated into a central Knowledge Engineer to produce the final classification.
The first stage in this process is removing the ocean and clouds by masking the image. The remaining areas of water are the first branch of our decision tree classification. These areas have strong absorption of near infrared light and therefore very low values in Landsat band 4. We are able to use this “natural” or “ecological” property to form the basis of a decision. If the values recorded for band 4 are below a defined cutoff point, then we expect those cells to represent standing water and classify those areas accordingly. In addition we are able to clearly identify areas of industrial or urban land cover as having very high reflectance in certain raw bands and can build this principle into a spectral decision tree classification.
Many of the vegetation types in Hawaiian forests cannot be distinguished clearly from the information available in raw TM bands for a variety of reasons, ranging from complex topography to small-scale mosaics of adjacent vegetation types within a limited geographic area. We employ two techniques to address this natural complexity. First, vegetation is known to have a low reflectance in Landsat band 3 (0.63-0.69 nm) and a high reflectance in Landsat band 4 (0.76-0.90 nm). As a result, vegetation indices have been designed to isolate this spectral feature and distinguish the amount of vegetation in an area. We use the standard Normalized Difference Vegetation Index (NDVI) to build the first branch in our decision tree separating areas of high biomass from areas of low biomass as indicated on the left in Figure 1. Using treatments such as the NDVI, Principal Component Analysis (PCA), and Tasseled Cap, we are able to find “cut points” or variables at which we can build branches for our classification tree illustrated in Figure 1.
Figure 1. Image Classification Decision Tree. Landsat images used are shown in rhomboid boxes, treatments applied are described in rectangular boxes, and splits into branches are in diamond-shaped boxes.
The tasseled cap treatment is particularly valuable in the Hawaiian High Islands ecosystem because of its utility in revealing brightness, greenness, and wetness. These variables are strong identifiers for Hawaiian vegetation communities, since their distributions are closely tied to moisture availability, exposure, and nutrient availability (Pratt and Gon 1998, Wagner et al. 1999).
Establishing the spectral decision tree within the Knowledge Engineer in one area enables the analyst to test and refine the decision trees in similar areas. When applying a decision tree to a new scene, it has often been found that only minor adjustments are needed to apply it in different places―or the same place under different conditions. However, when scenes are used from different seasons, we find significant adjustments are required in the classification, particularly in areas of grasslands and invasive shrubs, where changes in greenness due to “green-up” phenology are substantial. Being aware of seasonal variation and its effects on particular vegetation types enabled the HI-GAP team to adapt the classification to take advantage of seasonality differences in the tropics.
The spectral decision tree classification has been implemented on the Big Island of Hawaii, Maui, Lanai, and Molokai with positive results. The Big Island of Hawaii was the first island that was classified using this methodology, because it provided a wide variety of vegetation types with which to develop the decision tree methodology. The results for the southern forested regions of the Big Island were assessed using field point data gathered during helicopter surveys. The preliminary results indicate a 90% accuracy level for this region. Spectral properties for specific vegetation types were taken from the Big Island and applied to Maui and Molokai. The spectral values needed small adjustments to achieve the desired results. Results from the first draft have been reviewed and approved by various partners. Based on the work from the first draft on Maui and Molokai it is clear that spectral values gathered from the Big Island can be applied to other islands as initial hypotheses and then adjusted according to ancillary data and expert knowledge.
The minimum mapping unit for the HI-GAP project is 90 m 2. The methodology described above has provided accurate results at this scale where it has been tested. The same methodology was recently tested on a small area of vegetation on the Big Island of Hawaii. The results from this test indicate the methodology can be applied at higher resolutions than 90 m 2. The results from this study produced a detailed vegetation map with 30-meter pixel resolution. The significance of these results indicates the methodology being developed can be applied to small areas and is capable of producing detailed vegetation classifications of these areas.
Establishing a model of initially large, then progressively refined vegetation classes is an approach that matches the current hierarchical vegetation classification system developed for the Hawaiian Islands. Broad elevation, moisture, and physiognomic categories are initially established, within which canopy dominants are used to identify alliances and associations (Pratt and Gon 1998). This allows for specific analysis to be confined to geographic subregions on the basis of elevation as well as moisture, stratifying and thereby separating complex signature sets that might otherwise be indistinguishable on signature alone. Several such broad subdivisions can be readily defined and are relevant to ecological considerations such as species ranges and ecophysiology along gradients. Typical examples are wet windward vs. dry leeward, coastal/lowland vs. montane, and closed vs. open/sparse canopy variants of a given dominant canopy species.
A broader geological age gradient is also apparent across the archipelago from the youngest island, Hawaii (less than half a million years) to the oldest of the main islands, Kauai, at the opposite end of the chain (over 5 million years). There are also well-documented floristic differences among the islands due to their isolation from one another, which further complicates vegetation classification. This isolation factor results in the need for adjustments in decision criteria for spectral signatures appropriate to one island when applied to another. At the same time, key dominant species in the canopy are consistent from one end of the archipelago to the other, so an initial hypothesis of applicability of decision criteria from an island to a neighboring island may be valid but requires testing and adjustment for each island.
Finally, although there are over 150 described vegetation associations in the Hawaiian Islands, their distribution in space is strongly correlated with elevation, moisture, and soil, so that in a given class of elevation, moisture, and formation there is a manageably small set of associations that are typically present, especially in landscapes dominated primarily by native vegetation types. The situation is greatly complicated in alien-dominated or mixed native-alien vegetation, where an unstable disturbance-driven mosaic greatly increases the number of possible vegetation associations that can be present.
The Knowledge Engineer approach to developing a classification tree enabled the HI-GAP team to readily test refinements and alternative classification decisions and to analyze the effects of alternative approaches on results. It employs the objectivity and repeatability of CART but augments statistical outcome with the strength of local knowledge and the ability to refine and adapt the decision tree as information becomes available. Results to date indicate the methodology described here provides a new approach to mapping tropical land cover. Where available knowledge is limited, the upper levels of the classification hierarchy can be initially characterized, and as ground-truthing or surveys provide more extensive knowledge on the vegetation composition in the study region, the decision criteria can be further refined without discarding previous work, greatly enhancing the utility of foundation efforts.
De'ath, G., and K.E. Fabricius. 2000. Classification and Regression Trees: A powerful yet simple technique for ecological data analysis. Ecology 81:3178-3192.
Hansen, M., R. Dubayah, and R. Defries. 1996. Classification Trees: An alternative to traditional land cover classifiers. International Journal of Remote Sensing 17:1075-1081.
Hegge, K. 2001. Multi-Resolution Landcover Consortium Handbook. EROS Data Center, U.S. Geological Survey, Sioux Falls, South Dakota.
Khorram, S., K. Siderelis, H. Cheshire, and Z. Nagy. 1992. Mapping and GIS development of land use and land cover categories for the Albemarle-Pamlico drainage basin. U.S. Environmental Protection Agency Project # 91-08. March 1992. Raleigh, North Carolina.
Lawrence, R.L., and A. Wright. 2001. Rule-based classification systems using Classification and Regression Tree (CART) Analysis. Photogrammetric Engineering and Remote Sensing 67:1137-1142.
Pratt, L.W., and S.M. Gon III. 1998. Terrestrial ecosystems. In S.P. Juvik and J.O. Juvik, editors. Atlas of Hawaii, 3rd ed. University of Hawaii Press, Honolulu.
Roberts, D., M. Gardner, R. Church, S. Ustin, G. Scheer, and R.O. Green. 1998. Mapping chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models. Remote Sensing of the Environment 65:267-279.
Serrano, L., S. Ustin, D. Roberts, J. Gamon, and J. Penuelas. 2000. Deriving water content of chaparral vegetation from AVIRIS data. Remote Sensing of the Environment 74:570-581.
Wagner, W.L., D.R. Herbst, and S.H. Sohmer. 1999. Manual of the flowering plants of Hawaii. B.P. Bishop Museum, Special Publication 97. University of Hawaii Press and Bishop Museum Press, Honolulu.
Yang, L., C. Huang, C. Homer, B. Wylie, and M. Coan. In review. An approach for mapping large-area impervious surfaces: Synergistic use of Landsat 7 ETM+ and high spatial resolution imagery. Canadian Journal of Remote Sensing.
Colin Homer, Kent Hegge, Dar Roberts, Pat Halpin, and Shannon McElvaney.
Return to Table of Contents