A Bottom-up Approach to Vegetation Mapping of the Lake Tahoe Basin Using Hyperspatial Image Analysis

Increasing demands on the accuracy and thematic resolution of vegetation community maps from remote sensing imagery has created a need for novel image analysis techniques. We present a case study for vegetation mapping of the Lake Tahoe Basin which fulfills many of the requirements of the Federal Geographic Data Committee base-level mapping (FGDC, 1997) by using hyperspatial Ikonos imagery analyzed with a fusion of pixel-based species classification, automated image segmentation techniques to define vegetation patch boundaries, and vegetation community classification using querying of the species classification raster based on existing and novel rulesets. This technique led to accurate FGDC physiognomic classes. Floristic classes such as dominance type remain somewhat problematic due to inaccurate species classification results. Vegetation, tree and shrub cover estimates (FGDC required attributes) were determined accurately. We discuss strategies and challenges to vegetation community mapping in the context of standards currently being advanced for thematic attributes and accuracy requirements. Introduction Advances in the techniques and technology of hyperspatial image analysis are beginning to narrow the discrepancy between the fields of ground-based forestry and terrestrial remote sensing. With the need for increasingly larger scales of study to understand landscape level processes, pressure has been mounting for more accurate outputs from remote sensing given the large expense associated with field campaigns. However, to date, remote sensing has not approached the degree of accuracy and precision in measuring vegetation that an investigator on the ground achieves. In recent years, the U.S. Forest Service has been developing standards for vegetation classification and mapping using remote sensing imagery (Franklin et al., 2000; USDA, 2002). This paper details novel techniques by which most of the requirements of base-level mapping can be fulfilled using a combination of A Bottom-up Approach to Vegetation Mapping of the Lake Tahoe Basin Using Hyperspatial Image Analysis Jonathan A. Greenberg, Solomon Z. Dobrowski, Carlos M. Ramirez, Jahalel L. Tuil, and Susan L. Ustin hyperspatial image analysis including individual plant mapping, automated image segmentation, and vector-raster querying techniques. A base-level map must contain the following information: FGDC physiognomic classifications of order, class, and subclass; floristic classifications of cover types, dominance types and alliances; total vegetation, tree, and shrub cover classes in increments of 10 percent; and mean tree diameter classes ranging from 0 to 50 inches. The minimum accuracies of these attributes are 80 percent for physiognomy, 65 percent for floristics, 65 percent for cover class, and 65 percent for mean tree diameter class. The minimum mapping unit (MMU) for a base-level map is defined as “the smallest polygon feature to be mapped at a given map level” (Warbington et al., 2002). We note the use of the term “polygon”: the concept of vegetation mapping units is based on the spatial extent of soil orders at different scales (Warbington et al., 2002) and were historically performed by manual digitization of aerial photographs outlining patches of vegetation and soil. We hereafter refer to these polygon mapping units as “patches” consistent with FGDC terminology. While few base-level maps have been produced to date, coarser thematic-scale maps commonly employ one of two approaches to produce maps: (a) pixel-based classifications of mediumand coarse-scale imagery (ground resolution 1 m, e.g., CALVEG, Parker and Matayas, 1979), and (b) object-oriented analysis (OOA) of (typically) hyperspatial imagery (ground resolution 1 m, e.g., Lobo et al., 1998). Pixel-based classifications of medium and coarse-scale imagery typically rely on training data of field plots, which have been classified to a given physiognomic or floristic level. Techniques such as maximum likelihood and classification and regression trees (CART) are then used to produce maps with the appropriate classification. Higher taxonomic levels can be mapped by merging the classes into the appropriate coarser floristic or physiognomic level, or by reclassifying the field data into the coarser classes and then re-training the classifier. OOA is a relatively new technique for mapping which consists of two steps: (a) generation of a vector layer of vegetation patches by automated image segmentation algorithms (e.g., Baatz et al., 2003) and (b) classification of patches using spectral and textural data from raster pixels that fall within these patches. OOA techniques are garnering more attention for vegetation PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING May 2006 581 Jonathan A. Greenberg is at the NASA Ames Research Center, MS 242-4, Moffett Field, CA 94035 ([email protected]) and formerly at CALSPACE, University of California at Davis. Solomon Z. Dobrowski, Jahalel L. Tuil, and Susan L. Ustin are with CALSPACE, University of California at Davis, 1 Shields Ave., Davis, CA 95616. Carlos M. Ramirez is with the USDA Forest Service, Region 5, State and Private Forestry – Forest Health Protection, 3237 Peacekeeper Way, Suite 207, McClellan, CA 95652. Photogrammetric Engineering & Remote Sensing Vol. 72, No. 5, May 2006, pp. 581–589. 0099-1112/06/7205–0581/$3.00/0 © 2006 American Society for Photogrammetry and Remote Sensing HR-05-025.qxd 4/10/06 2:52 PM Page 581