1072Ecological Applications,8(4), 1998, pp. 1072–1083q1998 by the Ecological Society of AmericaINTEGRATION OF GIS DATA AND CLASSIFIED SATELLITE IMAGERYFOR REGIONAL FOREST ASSESSMENTHONGS. HE,1DAVIDJ. MLADENOFF,1VOLKERC. RADELOFF,1ANDTHOMASR. CROW21Department of Forest Ecology and Management, University of Wisconsin–Madison, 1630 Linden Drive,Madison, Wisconsin 53706-1598 USA2U.S. Forest Service North Central Forest Experiment Station, Forestry Sciences Laboratory,Rhinelander, Wisconsin 54501-0898 USAAbstract.New methods are needed to derive detailed spatial environmental data forlarge areas, with the increasing interest in landscape ecology and ecosystem managementat large scales. We describe a method that integrates several data sources for assessingforest composition across large, heterogeneous landscapes. Multitemporal Landsat ThematicMapper (TM) satellite data can yield forest classifications with spatially detailed informationdown to the dominant canopy species level in temperate deciduous and mixed forests. Westratified a large region (106ha) by ecoregions (103–104ha). Within each ecoregion, plot-level, field inventory data were aggregated to provide information on secondary and sub-canopy tree species occurrence, and tree age class distributions. We derived a probabilisticalgorithm to assign information from a point coverage (forest inventory sampling points)and a polygon coverage (ecoregion boundaries) to a raster map (satellite land cover clas-sification). The method was applied to a region in northern Wisconsin, USA. The satellitemap captures the occurrence and the patch structure of canopy dominants. The inventorydata provide important secondary information on age class and associated species notavailable with current canopy remote sensing. In this way we derived new maps of treespecies distribution and stand age reflecting differences at the ecoregion scale. These mapscan be used in assessing forest patterns across regional landscapes, and as input data inmodels to examine forest landscape change over time. As an example, we discuss thedistribution of eastern white pine (Pinus strobus) as an associated species and its potentialfor restoration in our study region. Our method partially fills a current information gap atthe landscape scale. However, its applicability is also limited to this scale.Key words: associated species; data integration; ecoregion; Forest Inventory and Analysis; forestinventory; forest landscape modeling; Geographic Information Systems; Landsat Thematic Mapper;satellite forest classification; secondary species; stand age; subcanopy.INTRODUCTIONAn ideal data source or map for the analysis of forestecosystems would contain high-resolution spatial in-formation for (a) dominant canopy tree species, (b)associated tree species, and (c) forest stand parameters(e.g., age). Such maps do not exist in digital formatfor large areas. The goal of our study was to derive adetailed forest map suitable for differentiating amongecoregions (103–104ha) within a large, heterogeneouslandscape (106ha). The data would be used in assessingregional forest patterns for making resource manage-ment decisions and as input for a large-scale, forestlandscape simulation model (Mladenoff et al. 1996, Heand Mladenoff 1999).We integrated several data sources for forest map-ping at the landscape scale. Classified satellite imageryand forest field inventories have been widely used inforest mapping, but both have shortcomings as spatialdata over large areas. Remote sensing of forest eco-systems has been the subject of many studies (e.g.,Manuscript received 27 May 1997; revised 29 November1997; accepted 5 January 1998.Franklin et al. 1986, Bauer et al. 1994, Woodcock etal. 1994). Satellite imagery is used because it coverslarge areas (e.g., 32 400 km2for one Landsat ThematicMapper [TM] scene) at high spatial resolution (e.g., 30m for Landsat TM). Current satellite imagery allowsforest classification at the dominant species level, ifmultiple image dates are used, with an accuracy rangeof 70–95% (Wolter et al. 1995). The presence of un-derstory vegetation in satellite imagery is generallytreated as a source of noise that complicates classifi-cation (Spanner et al. 1990). Remote sensing classifi-cations have not been able to provide detailed infor-mation on forest understory (Stenback and Congalton1990, Ghitter et al. 1995, Woodcock et al. 1996).Estimating forest stand parameters such as age fromsatellite imagery also faces limitations. In coniferousforests of the western United States, forest age andbasal area can be separated into several broad classes(Franklin 1986, Cohen et al. 1995). Crown closure canbe successfully mapped, but crown size classes are dif-ficult to separate (Woodcock et al. 1994). Mapping ofsuccessional stages is limited to three or four classes(Hall et al. 1991, Fiorella and Ripple 1993). In the
November 19981073INTEGRATION OF GIS AND SATELLITE DATAFIG. 1. Study area location (rectangle; size;29 000 km2)in the U.S. Midwest (MN5Minnesota, WI5Wisconsin,MI5Michigan).eastern United States, where there is less structural andage range in forest trees, deriving age and size classesfrom satellite imagery is even less successful (Mlad-enoff and Host 1994).Plot-level forest inventories are representative fieldsamples of forest stands for which a variety of param-eters are monitored, including associated species andstand age, often over large areas (Birdsey and Schreu-der 1992). The most comprehensive forest inventoryfor the eastern United States is the Forest Inventoryand Analysis data set (FIA; Hahn and Hansen 1985).Forest inventory data have been used to predict presentand future tree species ranges over the Eastern UnitedStates (Iverson and Prasad 1998). The FIA data arelimited in their spatial resolution, with 0.4-ha (1-acre)sampling plots, located on the landscape at a densityof 1–2 plots/10 km2. For forest assessments beyondsingle stands, the FIA sample points need to be aggre-gated into spatial units such as counties, states, orecoregions. Ecoregions are areas that may be definedhierarchically at a range of scales, and that are delin-eated according to their relatively homogeneous char-acteristics of soils, physiography, and climate (Bailey1988, Host et al. 1996). Ecoregions are suitable eco-logical units for aggregating FIA data because of theserelatively homogenous environmental characteristics.On their own, neither satellite-derived forest mapsnor forest inventories provide the input necessary forforest landscape modeling. The integration of satelliteforest classifications and FIA improves the applicabil-ity of both (Lachowski et al. 1992). In our study, wedescribe how these different data sources can be in-tegrated. Our approach was to derive distributions ofage classes and associated species for each dominanttree species mapped in a species-level forest classifi-cation by Wolter et al. (1995). These distributions werederived from the FIA data for each of the ecoregionsdelineated by Host et al. (1996). A probabilistic al-gorithm then assigned age and associated species foreach dominant species pixel, based on their likelihoodsof occurrence. We discuss the relevance of the resultingmaps for ecosystem management and forest landscapemodeling using eastern white pine (Pinus strobus)asan example.STUDYREGIONOur study region is situated in the western part ofthe Northern Hardwood and Conifer Forest Region ofthe northern Lake States (Fig. 1), a transitional zonebetween boreal forests to the north and temperate for-ests to the south (Pastor and Mladenoff 1992). Treespecies diversity is relatively high, with species of boththe northern and southern zones. Additionally, speciescharacteristic of the Northern Hardwoods region itself,such as eastern white pine, eastern hemlock (Tsugacanadensis), and yellow birch (Betula alleghaniensis)reach their maximum importance in this region (Mlad-enoff and Pastor 1993).The physiography and soils of the study region varyfrom nutrient-poor sands on glacial outwash plains, tosilty loams on moraines, and heavy clays of formerlake beds. Quaternary geology and mesoclimatic gra-dients are the greatest determinants of environmentalvariation in the region, leading to a wide range of dif-ferent ecosystems. Forests are dominated by hemlock,sugar maple (Acer saccharum), trembling aspen (Pop-ulus tremuloides), white pine, red pine (Pinus resino-sa), and jack pine (P. banksiana), on a gradient frommesic to xeric sites. Less important tree species foundalong the same gradient are balsam fir (Abies balsa-mea), yellow birch, northern red oak (Quercus rubra),big-toothed aspen (Populus grandidentata), and pinoak (Q. ellipsoidalis) (Curtis 1959).The landscape-scale forest variation is accompaniedby fine-scale variation due to local topography. To-pographic relief is moderate (usually,200 m); plainsand rolling hills dominate the landscape. The highgroundwater table results in lowland stands in glacialdepressions dominated by tamarack (Larix laricina),black spruce (Picea mariana), and white cedar (Thujaoccidentalis).The region was extensively altered since Europeansettlement, with large-scale, destructive logging oc-curring from the mid-1800s to the early 1900s. Re-peated severe fires followed logging, and portions ofthe landscape were in agricultural use at some point intime. Today forests are largely young second and thirdgrowth. Tree species composition, age class distribu-tion, and landscape structure are severely altered fromthe presettlement landscape (Flader 1983, Mladenoffand Pastor 1993, Mladenoff et al. 1993, Pastor andMladenoff 1993).METHODSFor our study we assumed that for each dominanttree species, its age classes and associated species can