Date of this Version
Published in the proceedings of the Third International Conference on Geospatial Information in Agriculture and Forestry held in Denver, Colorado, 5 -7 November, 2001. CD-ROM, 1 disk.
The Multi-Resolution Land Characteristics (MRLC) consortium was initiated in early 1990s to address the need for consistently developed national and regional land cover data (Loveland and Shaw, 1996). Through this consortium, a 1992-vintage National Land Cover Dataset (NLCD) was developed for the conterminous United States (Vogelmann et al., 2001), and a second generation National Land Cover Dataset (NLCD 2000) will be developed using 2000-vintage Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images and ancillary data. The 2000 NLCD will consist of a suite of data layers relevant to many applications, including a tree canopy density layer describing percentage of tree canopy cover within each 30 m pixel. As a continuous variable, this tree canopy density layer is proposed in addition to a land cover classification to better characterize subtle variations of tree canopy and to meet the increasing needs for continuous measures of land cover components (DeFries et al., 1995).
Previous efforts to estimate tree canopy density as a continuous variable have utilized linear spectral mixture analysis (SMA) or linear regression techniques (e.g. Iverson et al., 1989, Zhu and Evans, 1994, DeFries et al., 2000). Other techniques such as physically based models and fuzzy logics have also been explored but are probably premature for use over large areas (e.g. Li and Strahler, 1992, Baret et al., 1995, Maselli et al., 1995). A major disadvantage of SMA is that it cannot predict tree canopy density directly, because tree canopy is not a spectral end-member (Roberts et al., 1993). Both linear SMA and linear regression use linear models to approximate the relationships between spectral signal and canopy density. However, such relationships are often very
complex and highly variable, especially over large areas (e.g. Ray and Murray, 1996). This is partly due to multiple scattering effects and the highly spatially variable spectra of tree canopy and other surface materials (Borel and Gerstl, 1994). This problem may be partially alleviated using non-linear regression techniques. However, many nonlinear regression techniques require prior knowledge on the nonlinear form of a relationship (Gallant, 1987), which may be spatially variable and often unknown for land cover analysis. The regression tree technique, however, may be appropriate for this purpose because it could potentially approximate complex relationships using a set of linear models, which were found more accurate than a single linear regression model (Huang and Townshend, 2001). Therefore, we propose a strategy for deriving tree canopy density at intermediate spatial resolutions using this technique. We tested its applicability over large areas in three study areas located in Virginia, Utah and Oregon of the United States.
The overall approach of the proposed strategy consists of three key steps: deriving reference data from high resolution images, calibrating canopy density models using the derived reference data, and extrapolating the developed models spatially using 30 m resolution images (figure 1). Considering the extremely high cost of intensive fieldwork over large areas, deriving reference data from high-resolution images was common in previous studies (e.g. DeFries et al., 1997). In this study we used 1 m Digital Orthophoto Quadrangle (DOQ) images for reference data development and 30 m ETM+ images for model extrapolation.