A Neural Network Method for Efficient Vegetation Mapping

This article describes the application of a neural netsessment, fire control, wildlife habitat characterization, and water quality monitoring. In this context, a number work method designed to improve the efficiency of map of federal and state agencies as well as private companies production from remote sensing data. Specifically, the with large land holdings currently use vegetation maps ARTMAP neural network produces vegetation maps of derived from satellite-based remote sensing (e.g., Asthe Sierra National Forest, in Northern California, using pinall and Veitch, 1993; Bauer et al., 1994; Cohen et al., Landsat Thematic Mapper (TM) data. In addition to 1995; Congalton et al., 1993; Franklin and Wilson, 1991). spectral values, the data set includes terrain and location In one such application domain, Region 5 (California) of information for each pixel. The maps produced by ARTthe U.S. Forest Service (USFS) has, for the past two deMAP are of comparable accuracy to maps produced by cades, used Landsat sensor imagery for mapping vegetaa currently used method, which requires expert knowltion in its 20 National Forests. Over that time, the deedge of the area as well as extensive manual editing. In mand for vegetation maps has increased, even as sensor fact, once field observations of vegetation classes had technology and methods for deriving information from been collected for selected sites, ARTMAP took only a few remote sensing images have continued to improve. The hours to accomplish a mapping task that had previously result is ongoing pressure to refine the knowledge detaken many months. The ARTMAP network features fast rived from remote sensing, leading to new explorations online learning, so that the system can be updated increof map production methods. This article reports on findmentally when new field observations arrive, without the ings concerning the utility of one such new method, the need for retraining on the entire data set. In addition to ARTMAP neural network. The study compares ARTmaps that identify lifeform and Calveg species, ARTMAP MAP capabilities with those of a conventional method on produces confidence maps, which indicate where errors the benchmark problem of mapping vegetation in the Siare most likely to occur and which can, therefore, be erra National Forest. used to guide map editing. Elsevier Science Inc., 1999 The vegetation maps employed for management of National Forests in California identify basic lifeforms INTRODUCTION: VEGETATION MAPPING such as conifer, hardwood, water, and barren. Many of FROM REMOTE SENSING DATA these lifeforms are further subdivided by species associations, and, when appropriate, by tree size and cover. The Vegetation maps serve a wide range of functions in the present analysis considers the problem of mapping lifefmanagement of natural resources, including inventory asorms and species associations, with species labeled according to the California vegetation, or Calveg, classifica*Center for Adaptive Systems and Department of Cognitive and tion system (Matyas and Parker, 1980). The mapping Neural Systems, Boston University methodology that Region 5 of the USFS currently ap†Center for Remote Sensing and Department of Geography, Bosplies to this problem has evolved over the years into a ton University ‡Center for Earth Systems Analysis Research and Department of rather cumbersome system which uses two separate proGeography, San Diego State University cessing streams (Woodcock et al., 1994) (Fig. 1a). The Address correspondence to Gail A. Carpenter, Dept. of Cognitive first stream produces lifeform maps from Landsat Theand Neural Systems, 677 Beacon St., Boston Univ., Boston, MA 02215. matic Mapper (TM) input data and a terrain-derived imE-mail: [email protected] Received 31 December 1998; revised 14 May 1999. age of local solar zenith angle.