Applications of Satellite Remote Sensing to Forested Ecosystems

Home , Satellite imagery

Landscape Ecology vol. 3 no. 2 pp 131-143 (1989) SPB Academic Publishing bv, The Hague Applications of satellite remote sensing to forested ecosystems

Louis R. Iverson', Robin Lambert Graham2 and Elizabeth A. Cook' 'Illinois Natural History Survey, 607 E. Peabody, Champaign, IL 61820 and Department of Forestry, University of Illinois, Champaign, IL 61820, USA; 2Environmental Sciences Division, Oak Ridge National Laboratory, P.O. Box 2006, Oak Ridge, TN 37831, USA

Keywords: satellite, remote sensing, forest ecosystems, GIS, monitoring

Abstract

Since the launch of the first civilian earth-observing satellite in 1972, satellite remote sensing has provided increasingly sophisticated information on the structure and function of forested ecosystems. Forest classification and mapping, common uses of satellite data, have improved over the years as a result of more discriminating sensors, better classification algorithms, and the use of geographic information systems to incor- porate additional spatially referenced data such as topography. Land-use change, including conversion of forests for urban or agricultural development, can now be detected and rates of change calculated by superim- posing satellite images taken at different dates. Landscape ecological questions regarding landscape pattern and the variables controlling observed patterns can be addressed using satellite imagery as can forestry and ecological questions regarding spatial variations in physiological characteristics, productivity, successional patterns, forest structure, and forest decline.

Introduction over varying scales of spatial resolution. The sophistication of applications evident in re- Since the launching of the first earth-observing cent years has been made possible by (1) the use of civilian Landsat satellite in 1972, satellite remote more spectrally and/or spatially discriminating sensing has been used for gathering synoptic infor- sensors; (2) the improvement of hardware and mation on forests. In the early years, satellite data software systems designed to process spatially- were used mostly by geographers to create maps of referenced digital data, and (3) the increased availa- forest types. These early efforts relied almost en- bility, standardization, and compatibility of other tirely on satellite-collected digital spectral data with spatially-referenced digital data sets such as digital no integration of ground-based digital information topographic variables generated from digital eleva- such as topography. More recently, ecologists have tion models. The most common sources of satellite joined the geographers in utilizing satellite technol- data relevant to forests are the U.S. Landsat ogy for a variety of forest-related applications Thematic Mapper (TM), the U.S. Landsat Mul- which will be reviewed in this paper: detecting land- tispectral Scanner (MSS), the U.S. Advanced Very scape change over time, relating landscape patterns High Resolution Radiometer (AVHRR), and the to biological or physical phenomena, evaluating French Systkme Probatoire d'observation de la physiological processes of forest canopies, and Terre (SPOT). The spectral characteristics and spa- quantifying forest cover, biomass, or productivity tial resolution of data from these sensors are por- 132

2 0.725 1.10 3 3.55 3.93 4 10.5 11.5

4 0.80 1.10

50 - 40 RELATIVE - REFLECTANCE 30 - (PERCENT) 20- - 10- \ WAVELENGTH 1 2 0.52 0.60 3 0.63 0.69 I Resolution: 30 rn “sr 4 0.74 0.91 5 1.55 1.75 6 10.4 12.5 -7 2.08 2.35 -BANDS 1 1 2 0.61 0.68 Resolution: 10-20 m 3 0.79 0.89 PAN 0.50 0.74 31 I, I I I I I I I I , , , , , , .I,. . . - I IS 1 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8 4.0’ 10.0 12.0 Spectral Wavelength, pm

Fig. 1. Spectral and spatial resolution for four common satellite sensors: AVHRR, MSS, TM and SPOT. Also shown is a spectral response curve for typical green vegetation. trayed in Fig. 1 and compared to the electromagnet- mation on forests, but are beyond the scope of this ic spectrum typically found in green vegetation. paper. Sensors on the recently launched Japanese More details on each sensor’s characteristics can be satellite and the Russian satellite are also useful in found elsewhere (e.g., Billingsley 1984; Greegor forest applications although their full potential is 1986). Several other sensors have been used in untested. forest-related applications but much less frequent- Current trends in ecological studies have dictated ly; for example, the Scanning Multichannel Micro- the integration of remotely sensed digital spectral wave Radiometer (SMMR) to monitor vegetation in data into geographic information systems (GIs). semiarid regions (Choudhury and Tucker 1987) or This merger moves satellite spectral data beyond for assessing global primary productivity (Choud- standard image processing and permits the use of hury 1988) and radar data for detecting forest remotely sensed spectral data in conjunction with change (Lee and Hoffer 1988; Stone and Woodwell such other spatially referenced digital data as eleva- 1988). Several sophisticated airborne sensors are tion, slope aspect, vegetation type, and soils. In this capable of detecting a great deal of ecological infor- way, information about a landscape can be en- 133 riched beyond what is possible by the separate sys- on the basis of ground-based data and the spectral tems (Logan and Bryant 1987). The integration of properties of the class (e.g., water has unique image processing systems and multilayered spectral reflectance characteristics so it can often be dis- data (as provided by satellite sensors) with GIS and cerned directly from its spectral signature). In su- digital geographic databases allows for the develop- pervised classification, the operator assigns specific ment of more sophisticated models of landscape- pixels (training sites) to particular landcover classes scale variables such as regional forest cover (Iver- on the basis of ground-based data. Computer al- son et af. 1989). gorithms are used to analyze the spectral properties The objective of this paper is to review ways in of those sets of pixels and to assign the remaining which satellite remote sensing can be useful in pixels to landcover classes on the basis of the delineating structural and functional characteristics statistical similarity of their spectral properties. of forests at a variety of geographical scales. We fo- Satellite data of all resolutions have been used to cus on the following uses of satellite imagery: (1) generate forest type maps, from high resolution classification and mapping of forest types; (2) de- SPOT and TM land-use maps (e.g., Hopkins et af. tection of areal change in forestland due to clearing 1988; Buchheim et af. 1985; Nelson et af. 1984) to or reforestation; (3) determination of patch disap- mid resolution MSS maps (e.g., Beaubien 1979; pearance and compositional change during succes- Dodge and Bryant 1976) to coarse resolution sion; (4) assessment of forest structure (basal area, AVHRR maps (e.g., Tucker et af. 1985; Norwine biomass, leaf area index, density, crown closure); and Greegor 1983; Townshend et af. 1987). (5) determination of damage or forest decline; (6) Comparisons of the various sensors for classifi- assessment of physiological processes and (7) as- cation and mapping accuracy have shown the su- sessment of forest cover and productivity. periority of the finer resolution TM data over the MSS data (DeGloria 1984; Williams et af. 1984; Malila 1985; Toll 1985; Hopkins et af. 1988). Toll Applications of remotely-sensed data to forests (1985) found that the improvement in classification of a scene of rapidly urbanizing Washington D.C. Mapping of forest types was due primarily to the better spectral discrimina- tion of TM data (especially TM bands 1, 5, and 7; Using satellite data to classify and map various Fig. 1) and to a lesser degree to the increase in quan- forest and/or land-use types has historically been, tization of the spectral data within a band (a raw and still is, the most frequent use of satellite data. MSS band value can range from 1 - 128; a TM band Pixels are classified according to their ground value can range from 1-256). Interestingly, Toll reflectance values as measured by the satellites. The found that the increased spatial resolution of TM desired map is created by displaying the classified reduced his ability to differentiate land-use classes pixels in their appropriate geographic context. Two of the first order such as urban, forest, agriculture, types of classification procedures may be followed and water. This reduction occurred because the to create a forest type map from a satellite image finer resolution TM data increased spectral varia- (Colwell 1983; Lillesand and Kiefer 1987). In un- bility within the pixels of first-order classes but the supervised classification, computer algorithms are spatial context of the pixel was not incorporated used to examine the spectral data of the entire scene into the classification algorithms (e.g., forested ur- and to clump pixels with like spectral properties ban areas such as yards and small parks were classi- into common classes according to the specific fied as forest rather than urban). However, Hop- clustering algorithm used. The classes are indepen- kins et af. (1988), examining forested areas of Wis- dent of any apriori assumptions as to what ground consin, USA, found the spatial detail of TM to be cover they actually represent. After the classes are advantageous in classifying second- and third-order generated, the operator assigns meaning to the forest land-use types such as upland coniferous classes (i.e., converts the classes to landcover types) forests and central hardwoods. The difference in 134 classification accuracy lies in the nature of the land- bles were integrated with TM data to increase the cover classes desired: for classification of a finer, accuracy of classifications of vegetation communi- higher order, the higher spatial resolution of TM is ties in Rocky Mountain terrain (Frank 1988). By in- beneficial (Williams and Nelson 1986); for classifi- corporating topographic variables, the shadowing cation of a coarser, lower order, TM is disadvanta- effects created by the angle of the sun can be ac- geous unless the spatial context of a pixel is incor- counted for. Topo-climatic variables can also pro- porated into the classification procedure. vide indirect information about vegetation cover The usefulness of SPOT data in classifying forest which can be incorporated directly into classifica- types has received mixed reviews. In an urban study tion algorithms. Other biogeographical variables - in Athens, Georgia, SPOT data were found to in- soil types, landforms, geology, or vegetation maps crease the accuracy of all second- and third-order - can also be helpful in classifying by providing classifications by about 15 to 20% over that of TM strata (e.g., forest-non-forest, cultural-non- (Welch 1985); further, these data were suitable for cultural, or wetland-non-wetland masks) that al- cartographic mapping at a scale of 1:24,000. How- low image classification to be focused on a particu- ever, SPOT data may be less helpful for mapping lar area or resource of interest. forest types (away from urban regions) because its Using satellite imagery to classify forest types is reduced spectral resolution (fewer bands) relative to still a subjective procedure and as much an art as a TM may obscure vegetation differences. science. Nonetheless, the technique has proven very AVHRR classifications are useful for maps of useful not only to researchers but also to agencies large areas and can be verified with higher resolu- that manage land resources. Classifications tend to tion images or map data (Schneider 1984). For ex- be more accurate in flatter terrain and when the ample, multitemporal AVHRR data were used to vegetation types are sharply contrasting, for exam- develop a vegetation map of South America in ple, coniferous versus hardwood or forested versus which 16 vegetation classes were differentiated, agriculture. The use of multitemporal scenes to several with accuracy greater than 90% (Town- capture phenological differences in vegetation shend et al. 1987). often improves accuracy. In inaccessible parts of Much research has been conducted in an effort to the world such as the tropical and boreal regions, enhance classification results. Raw spectral data satellite imagery is invaluable in mapping forest- may be pre-processed prior to classification. Vari- land because often no other current data are ous mean or median filters, in which pixels are reas- available. signed the mean or median spectral value of their surrounding pixels, may be applied to reduce intra- class variance while retaining the boundary detail Detection of forest change of classified areas (Atkinson et al. 1985; Cushnie and Atkinson 1985). Raw spectral values are also Changes in forest cover over time are important be- sometimes converted to their principal component cause of the role forests play in the global carbon values via principal components analysis of the en- cycle, in global climatic trends, and in providing tire scene. More sophisticated techniques for clas- species habitat (Woodwell et al. 1984). Although sification include stepwise discriminant analysis understanding forest change is important world- (Nelson et al. 1984) and per-field algorithms (Dean wide, it is especially important in the tropics, where and Hoffer 1982) in which the classification of a land-use transformation is occurring very rapidly pixel depends not only on its own spectral charac- and where timely ground data are scarce. teristics but also on those of adjacent pixels. The basic methodology for detecting change is Recently, classification accuracies have been im- straightforward: two or more satellite images of the proved by using a CIS to integrate digital bio- same area, preferably taken at the same phenologi- geographical data with satellite sensor data (e.g., cal period but in different years, are overlaid to CERMA 1985). For example, topographic varia- show geographically specific changes in landcover. 135

In some cases, raw satellite spectral data can be lack of timely ground data. Radar data also holds taken from the two scenes and merged to make a good potential for assessing deforestation in the multiple band combination data set, which is then tropics since radar data are not constrained by classified. Usually, however, the two images are cloud cover. For example, Stone and Woodwell classified separately prior to combining the data; (1988) found Shuttle Imaging Radar-A (SIR-A) this technique permits the use of varying data types data to have the brightest returns (the highest signal such as MSS and TM or even historic ground-based returns to the radar sensor result from smooth, de- maps. forested areas) on recently deforested regions in By comparing digitized ground-based maps of Amazonia. Costa Rican forestland from 1940, 1950, and 1961 Although most forest change studies using satel- and MSS-derived forest cover maps of 1977 and lite data have focused on deforestation in the 1983, Sader and Joyce (1988) found that forest tropics, temperate forest changes have also been cover had decreased from 67 to 17% between 1940 studied because of their importance with regard to and 1983 with the most rapid rate of clearing be- soil protection, water retention during flooding, tween 1977 and 1983. Furthermore, four of the ll wildlife habitat, timber resources, and recreation Costa Rican life zones had disappeared completely: sites. For example, loss of bottomland forest coin- the dry tropical, the moist premontane, the moist cident with upland forest regeneration has been lower montane, and the wet montane. They also documented in southern Illinois, USA, using classi- demonstrated the close relationship between road fied 1978 MSS and 1984 TM scenes (Iverson and building and deforestation by overlaying transpor- Risser 1987), as has forest degeneration in the high- tation network maps with forest cover maps. elevation forests in the Green Mountains of Ver- Deforestation in the Amazon basin of Brazil has mont (Vogelmann 1988). been quantified by using AVHRR band 3 thermal The accuracy of satellite-based studies of forest data which, unlike the visible bands, can penetrate change in the tropics are difficult to determine, the ubiquitous cloudcover of the region (Tucker et given the lack of ground-based data for verifica- al. 1984; Malingreau and Tucker 1987). Estimates tion. Nonetheless, the results are valuable because of deforestation were obtained by using band 3 to they are often the only source of timely, regionally detect both the fires associated with lines of active consistent information on deforestation. In tem- deforestation and the devegetated areas, which are perate regions where ground-based data are often warmer. The studies of Rondonia, Brazil, indicate available (e.g., national forest inventories), satellite that the deforested area increased from 4200 km’ in studies are nevertheless valuable because they can 1978 to 10,000 km’ in 1982 to 27,000 km’ in 1985 show the spatial pattern of change, which most in- to over 35,000 km’ in 1987 (Malingreau and Tucker ventories cannot, because they can look at shorter 1987; Malingreau and Tucker 1988). time intervals (two to three years versus 10 or more Deforestation rates in Rondonia, Brazil, have years for most ground-based surveys), and because also been evaluated using AVHRR data in combi- the methodologies developed to assess forest cover nation with selected cloudfree 1976, 1978, and 1981 (or change) over large regions can be validated with MSS scenes of much smaller portions of the region the independent inventory data sets (as in Iverson et (Nelson and Holben 1986; Nelson et al. 1987; al. 1989). Woodwell et al. 1987). The spatially precise MSS data revealed a doubling of deforestation rates between the 1976-1978 and 1978-1981 intervals Forest succession (Woodwell et al. 1987). The MSS data were also used to check the accuracy of AVHRR band 3 esti- The spatial and temporal patterns of forest succes- mates of cleared areas for the entire state of Rondo- sion can be studied using spatially referenced vege- nia. The estimates appeared reasonably accurate tation data from two or more dates. Transition given the constraints on the satellite data and the probabilities of forest succession pathways in 136 northern Minnesota, USA, were calculated using spectral data generated from airborne sensors such classified MSS scenes from 1973 and 1983 (Hall et as the thematic mapper simulator (TMS), which has al. 1987). Transition probabilities of the managed bands identical to TM, rather than satellite-borne areas differed from those of the wilderness areas sensors. The resolution of airborne spectral data is primarily because of the influence of logging, often finer than that of satellite data. Virtually all which altered not only the rates of transition but studies have focused on coniferous forests, which also the possible types of transition. In a second tend to be more uniform and more distinguishable type, Walker et al. (1986) successfully used Landsat from other vegetation types than are deciduous MSS data in Australian semi-arid eucalypt wood- forests. Whether the techniques used to relate satel- lands to detect stage of seccession based on struc- lite data to forest structure in coniferous forests will tural differences in 0 to 50 year old clearings. also be appropriate for non-coniferous forests is yet In another forest succession study, which utilized to be determined. both image processing and GIS technology, the sta- Canopy closure in montane, coniferous forests bility and fate of abandoned pasture patches in a of California, USA, correlated well with the spec- mosaic of mountainous forest were found to be sig- tral intensity of several TMS bands (r = 0.62 to nificantly related to original patch size and eleva- 0.69, n = 103) irrespective of forest type (Peterson tion (Graham et al. 1987). This study used 1934 et al. 1986). Total stand basal area, however, was vegetation maps depicting the abandoned pastures poorly related to the spectral data (r < 0.33). and 1984 TM imagery. The fate of the pastures Stratification by forest type improved the spectral patches was determined by comparing the 1984 relationship with basal area. The data suggested spectral signature of pixels within the historic that the relationship between total basal area and boundaries of the abandoned patches with the spec- spectral signature will be strongest in young, low tral signature of pixels just outside the patch density, even-age stands. In another study of boundaries. Californian coniferous forests, TMS bands 1,2and Satellite imagery holds considerable promise for 3 (analogous to TM bands 1, 2 and 3) were most determining the rate and spatial context of succes- strongly related to stand basal area and leaf bi- sion; however, this use is still experimental and not omass (Franklin 1986). without problems. The accuracy of transition prob- A relatively high relationship between TMS spec- abilities will depend in large part on the accuracy of tral band intensity and canopy closure (r = 0.80, n the original classifications. Furthermore, there are = 32 for band 5) was found for the pine-aspen theoretical problems in calculating transition prob- forest of Colorado, USA (Butera 1985). By apply- abilities with data that have a time interval equal to ing a regression model of this relationship to the or greater than that of the change phenomena. raw band 5 value of every pixel in the mountainous scene, Butera generated a map of forest canopy closure. The accuracy of the map was 71070, 74%, Assessment of stand structure and 54% for canopy closures of 0-25%, 25-75'70, and 75- 100% respectively. Satellite data have been used with varying degrees Spanner et al. (1984) used a classification ap- of success to quantify spatially such forest structure proach to study the ability of TMS imagery to characteristics as crown cover, tree density, tree di- differentiate crown closure and tree size classes in ameter, basal area, tree height, tree age, biomass, a fir-dominated forest in Idaho, USA. They found and leaf area index. In general, the technique is to > 60% accuracy in classifying crown closure class- collect spatially-referenced ground data on the es of > 70%, 40-69'70, and 10-39%, with less ac- forest structure variable of interest and then to de- curacy on sites of very low (< lOQ7o) crown closure. termine the statistical relationship between the Sawtimber and pole size classes were also classified ground-obtained data and the spectral data for the with 72-87% accuracy. The optimal bands in these same location. Thus far, most studies have used analyses were, in order, 4, 7, 5, and 3. 137

Again using TMS imagery, researchers have Assessment of forest damage related leaf area index (LAI) of coniferous forests to spectral band intensity (Running et al. 1986; The assessment of forest damage is an important Peterson et al. 1987). In these studies, LA1 of use of remote sensing data. Many of the changes in coniferous forest stands along a transect across the tree or foliage morphology resulting from stress can mountains of Oregon, USA, was strongly related to be detected with remote sensors (Jackson 1986). the ratio of band 4 to band 3 (r2 = .82, n = 18). Furthermore, the spectral signature of stressed trees LA1 of these stands ranged from < 1 to > 16. may indicate not only the degree of stress but also Danson (1987) correlated SPOT data with struc- the type of stress. For example, TMS imagery of tural characteristics of pine forests in England and damaged red spruce (Picea rubens) stands in Ver- found highly significant correlations of SPOT band mont shows a large reduction in the near and 3 (near infrared) to tree density, diameter at breast shortwave-infrared reflectance (bands 4 and 5 height, and tree age but not to canopy cover. Wu respectively) (Rock et al. 1986). The location of and Sader (1 987) showed that airborne Multipolari- highly damaged stands was readily apparent in the zation Synthetic Aperture Radar (SAR) also may be scene if the ratio of these bands was displayed. Field used with some success to estimate basal area, tree verification of the image revealed that the foliage of height, and total tree biomass. the highly damaged spruce stands was drier and less An airborne, pulsed laser system, called the Light dense than that of undamaged stands (Rock et al. Detection and Ranging (LIDAR) system was also 1986; Vogelmann and Rock 1986). These authors used to predict total tree volume and green weight have continued their work with the TM sensor and biomass of pine plantations in Georgia (Nelson et have been successful in assessing forest damage in al. 1988). They were able to predict overall tree Vermont and New Hampshire (Vogelmann and volume to within 2.6% and mean biomass to within Rock 1988). 2.0070,but were not successful at predicting volume Damage produced by insect defoliation has also or biomass on a site by site basis. been successfully assessed from remotely sensed im- Results from these studies are encouraging in agery. This type of damage is easily perceived by ex- that statistically significant relationships between amining scenes of an area before and after defolia- spectral data and forest structure data generally do tion. For example, areas of heavy gypsy moth exist. The results are also frustrating, however, be- defoliation in Pennsylvania, USA, were quite evi- cause the relationships are not consistent across dent in a foliage difference map created from June studies and are generally too weak to offer predic- 1976 and July 1977 MSS data (Williams and tive accuracy at a per pixel scale. As a consequence Stauffer 1979). The key to successful defoliation as- of the latter weakness, the relationships cannot be sessment is to use scenes that capture the period of used to examine the spatial patterns of structural at- heaviest defoliation (Dottavio and Williams 1983). tributes. Nonetheless, in some cases the relation- Spectral imagery is used routinely by forest ships can be used to accurately determine the mean managers to detect and measure insect defoliation, value of a structural attribute over a landscape although the data often come from airborne rather (Iverson et al. 1989). New approaches such as the than satellite sensors. Stress detection of forests is incorporation of biogeographical data, along with still in the research stage, but results thus far are additional research and probable technology de- promising. velopment will be necessary before satellite imagery - forest structure relationships are sufficiently accurate and robust to be truly useful in large scale in- Assessment of physiological parameters ventories or to detect spatial changes in stand structure. Many physiological attributes - such as photosynthesis, evapotranspiration, plant maintenance respiration, turnover of organic carbon, and 138 moisture retention - are related to the interaction water-stressed vegetation (Rock et al. 1986). Hand- of solar radiation and vegetation (Knipling 1970). held spectral sensors have been used to detect water As such, satellite sensors, which measure the light stress in buffelgrass in Texas, USA (Richardson reflectance of the earth’s surface, should potential- and Everitt 1987). To our knowledge, current satel- ly be able to indirectly measure changes in these lite data have not yet been used to satisfactorily radiation-mediated physiological processes. Re- evaluate moisture availability of forested canopies. flectance measurements should be useful in infer- Mounting evidence suggests that remotely sensed ring spatial and/or temporal variations in pho- spectral data may become as successful, if not more tosynthesis and evapotranspiration rates because successful, at estimating forest function (e.g., pho- the structural and functional properties of leaves tosynthesis or evapotranspiration) than forest determine the radiatiodinterception characteris- structure (e.g., biomass or leaf area) because of the tics of tree canopies (Sellers 1985; Tucker and dynamic nature of the reflectance-physiological in- Sellers 1986). Spectral data can provide informa- terface (Tucker and Sellers 1986; Kimes et al. 1987). tion on the amount of chlorophyll pigment (visible In the future, remote sensing may be able to detect wavelengths), the projected green leaf density (near portending ecosystem shifts by detecting changes in infrared), and the leaf water content of the canopy rates of key physiological processes that reflect (shortwave infrared). The first two can be used to basic ecosystem parameters (e.g., photosynthesis infer potential photosynthesis although actual pho- and productivity) (Waring et al. 1986). Evidence tosynthesis will be determined by solar flux, also suggests that some of these physiological para- moisture availability, and other environmental fac- meters such as photosynthesis can be estimated tors operating on the system at the time (Tucker and without knowledge of species (Aber and Fownes Sellers 1986). 1985). Detection of ecosystem parameters without Spectral reflectance data should also be useful in identification of species is necessary to integrate identifying many important biochemical features data across landscapes and eventually the globe. of forest canopies because many biochemical com- However, examples of satellite detection of phys- pounds possess unique spectral absorption proper- iological processes of forested ecosystems are rela- ties (Waring et al. 1986). Determination of leaf tively scarce at the present time. Running and starch, nitrogen, protein, and lignin content should Nemani (1988) found a high relationship between be feasible from spectral data, although probably photosynthesis, transpiration, and aboveground not with current satellite technology (Waring et al. primary productivity as ascertained by an eco- 1986). For example, Spanner et al. (1985) were able system simulation model and the annual integrated to relate canopy nitrogen content to spectral data normalized difference vegetation index (NDVI, taken with the Airborne Imaging Spectrophotome- (near infrared-red)/(near infrared + red) from the ter (AIS). Peterson et al. (1988) have extended this AVHRR sensor, Fig. 1) over seven sites in the U.S. work over several sites to find relationships be- They found the relationship to be especially rigo- tween nitrogen, lignin, and total water content with rous on sites located at high latitudes with little the AIS spectral signatures. The infrared region of seasonal water stress. Serafini (1987) used diurnal the electromagnetic spectrum has been shown to be and seasonal variations in the difference between especially rich in information about canopy bio- satellite-derived earth surface temperature (based chemical characteristics. on AVHRR data) and air temperature near the sur- Leaf water content and consequently forest stand face (as measured by ground-based, shelter-height water relations should also be able to be inferred sensors). Spatial variation of evapotranspiration from canopy spectral reflectance properties in the could account for the variation in the derived shortwave infrared bands (e.g., TM bands 5 and 7) differences. Tucker et al. (1986) and Fung et al. (Tucker 1980). In the field, the higher values of the (1987) found a high correlation over a 3 !h year peri- ratio of the percent reflectance at 1.65 pm to the od between globally averaged NDVI and globally- reflectance at 1.26 pm corresponded to highly averaged monthly atmospheric C02concentra- 139 tions. This relationship suggests that satellite data 1987; Cook et al. 1989). In general, forest produc- can be used to estimate terrestrial photosynthesis tivity was more accurately predicted with a combi- and productivity, since atmospheric CO, varies nation of TM and biogeographical variables than seasonally according to the amount of drawdown with either data type alone. The best regression occurring via photosynthesis. The intensive studies models in each of the three study regions were high- of the First ISLSCP Field Experiment (FIFE), per- ly significant (p < 0.002) but left a considerable formed during 1987 and 1988 on the Konza Prairie, amount of the spatial variance in forest productivi- Kansas, used a large number of ground, airborne, ty unexplained. Because of the extreme heteroge- and satellite sensors to assess the potential to under- neity of forests stands at the 30-m2 resolution of stand the physiological characteristics of vegetation TM and because of the many abiotic and biotic (especially with regard to the effect on climate) via variables involved, it may not be reasonable to ex- remote sensing (Sellers et al. 1988). Most research, pect a high degree of predictability on small, site- however, that relates spectral data to forest physio- specific areas (Franklin 1986; Peterson et al. 1986). logical features has been done using various air- Predictability may be improved by changing the borne sensors (Sader 1987) or portable sensors scale of reference to cover larger areas or by pooling mounted on platforms or low flying helicopters. and/or stratifying data (Cook et al. 1989; Franklin Furthermore, often the sensors have been quite 1986). different from the sensors currently employed on As a means to scale up to regional levels, Iverson satellites. More refined satellite sensors and much et al. (1988) used the TM-derived models men- research will be needed before satellite spectral data tioned previously for the Illinois and Tennessee will truly promote a better quantitative understand- sites in combination with TM and AVHRR scenes ing of the temporal and spatial pattern of physio- of the same areas to develop predictive relation- logical properties of the earth’s vegetation. ships between the much coarser but more extensive AVHRR data and forest productivity. Multiple regression was used to develop the models relating Assessment of forest productivity AVHRR spectral data to TM-derived estimates of forest productivity. The resulting models were then If satellite sensors could accurately detect forest applied to each AVHRR pixel in the region to de- productivity, they would provide obvious cost and velop regional maps of forest productivity. The va- effort advantages over traditional field survey lidity of these maps was tested by aggregating the methods. Productivity estimates based on satellite AVHRR pixel productivity into county-level pro- data have been produced with some success for ductivity estimates and then comparing these agronomic ecosystems (Olang 1983), wetlands county-level estimates with independently derived (Butera et al. 1984; Hardisky et al. 1984), and county-level forest productivity estimates from the shrublands (Strong et al. 1985). Productivity assess- U.S. Forest Service. ments of forests using satellite data are rare. Forest For the 428 counties centered on the southern 11- productivity classes in northwestern California, linois region, the correlation coefficient of the two USA, were predicted with moderate success using a productivity estimates was 0.72 P (p < 0.0001). CIS with classified MSS data, topographic data, For the 168 counties centered on the eastern Ten- and ecological zone data (Fox et al. 1985). nessee region, the coefficient was 0.52 (p < In another study, predictive models of wood 0.0001). The lower success in the eastern Tennessee mean annual increment of volume in three regions region was attributed to the more variable land- of the United States (southern Illinois, eastern Ten- scapes of counties > 100 km from the original TM- nessee, and northeast New York) were developed AVHRR-forest productivity calibration center. using CIS, TM data, and digital biogeographical Closer to the calibration center (within 100 km), the data on forest productivity and soils, slope, solar correlation coefficient was 0.86. To extend this radiation, and/or vegetation type (Cook et al. methodology of using nested TM and AVHRR 140 scenes to scale up relationships between spectral ably focused on parameter identification rather values and productivity to continental or global than on spatial relations of ecological parameters. scales will probably require stratification of the ini- Satellite data provide two main applications to tial calibration sites by ecological regions such as forest ecology: (1) the ability to monitor ecological Kiichler’s (1964) potential vegetation type or attributes in inaccessible regions and/or spatially Bailey’s (1980) ecoregions (Logan 1985). extensive regions, and (2) the capacity to detect the The above methodology for developing regional spatial ecosystem patterns and processes of forests. estimates of productivity differs from most re- Much progress has been made toward the first ap- gional-scale remote sensing studies of productivity plication although much is left to be done. The sec- which generally rely solely on AVHRR data. A ond application has just begun to be explored. more common approach is to use multiple scenes of Forests are fundamental to the healthy function- AVHRR data to capture the change in a spectral ing of the biosphere. With the current global cli- greenness index over the growing season (Goward mate warming, loss of biodiversity, environmental et al. 1985, 1987; Tucker et al. 1985; Shimoda et al. degredation, and increased need for forest products 1986; Townshend and Justice 1986). The most suc- (all problems that rely on forests as a key in the cessful index has been the normalized difference process or mitigation of the process), it is impera- vegetation index (NDVI). For example, Goward et tive that we monitor and understand the forests of al. (1987) found a high correlation between this globe. Synoptic, timely information, which can seasonal changes in NDVI values and literature esti- be provided only with satellite data, is needed to mates of biome productivity across 24 North and support local, national, and global decision-makers South American biomes. Choudhury (1988) also in the crucial planning efforts designed to preserve found Nimbus-7 37 GHz SMMR data highly cor- the habitability of this planet for generations to related with estimates of global net primary pro- come. ductivity. Generally, though, valid satellite-based estimates of productivity or other ecological parameters across a large area are difficult to obtain be- Acknowledgments cause of the problems with securing ground obser- vations over such large regions (Curran and Wil- The authors are grateful to Tom Frank, Paul Ris- liamson 1986). ser, Jerry Olson, and Ying Ke for assistance in our work reported here, and to all those, whether men- tioned in this paper or not, who have done research Conclusions in the field of remote sensing of forests. Funding by NASA Contract NAS5-28781 is gratefully ack- The use of satellite data as an aid in understanding nowledged. Thanks to Monica Turner, Dan Neely, the ecological nature of forests is a very recent and and Audrey Hodgins for helpful reviews. Publica- rapidly evolving phenomenon. Although most tion number 3299 of the Environmental Sciences forest applications are still in the experimental Division of the Oak Ridge National Laboratory. stage, research suggests that satellite data will prove extremely useful in extracting spatial information on forest ecosystem attributes. Because satellite Literature cited sensor data integrate optically over the pixels, they are not as useful as finer resolution data if informa- Aber, J.D. and Fownes, J. 1985. Modeling the controls of forest tion on site specific ecological parameters is productivity using canopy variables. In Machine Processing desired. However, satellite sensors are indispensible of Remotely Sensed Data Symposium Proceedings. pp. 157-161. West Lafayette, Indiana. if one wishes to evaluate or monitor large areas. Atkinson, P., Cushnie, J.L., Townshend, J.R.G. and Wilson, The synoptic quality of satellite data is just begin- A. 1985. Improving thematic mapper land cover classification ning to be exploited; most research has understand- using filtered data. Int. J. Rem. Sens. 6: 955-961. 141

Bailey, R. 1980. Description of the ecoregions of the United DeGloria, S.D. 1984. Spectral variability of LANDSAT-4 States. United States Department of Agriculture Forest Serv- thematic mapper and multispectral scanner data for selected ice, Ogden, Utah. crop and forest cover types. IEEE Trans. Geosci. Rem. Sens. Beaubien, J. 1979. Forest type mapping from LANDSAT digital GE 22: 303-311. data. Photogr. Eng. Rem. Sens. 45: 1135-1144. Dodge, A.G. and Bryant, E. 1976. Forest type mapping with Billingsley, F.C. 1984. Remote sensing for monitoring vegeta- satellite data. J. For. 74: 526-531. tion: an emphasis on satellites. In The Role of Terrestrial Dottavio, C.L. and Williams, D.L. 1983. Satellite technology: Vegetation in the Global Carbon Cycle. pp. 161-180. Edited an improved means for monitoring forest insect defoliation. by G.M. Woodwell. John Wiley and Sons, New York. J. For. 81: 30-34. Buchheim, M.P., Maclean, A.L. and Lillesand, T.M. 1985. Fox, L., Brockhus, J.A. and Tosta, N.D. 1985. Classification of Forest cover type mapping and spruce budworm defoliation timberland productivity in northwestern California using detection using simulated SPOT imagery. Photogr. Eng. Landsat, topographic, and ecological data. Photogr. Eng. Rem. Sens. 51: 1115-1122. Rem. Sens. 51: 1745-1752. Butera, M.K. 1985. A correlation and regression analysis of per- Frank, T.D. 1988. Mapping dominant vegetation ecosystems in cent canopy closure vs. TMS spectral response for selected the Colorado Rocky Mountain Front Range and Landsat forested sites in the San Juan National Forest, Colorado. In TM. Photogr. Eng. Rem. Sens. 54: 1727-1734. Proceedings of IGARSS '85 Symposium. pp. 287-302. Ann Franklin, J. 1986. Thematic mapper analysis of coniferous Arbor, Michigan. forest structure and composition. Int. J. Rem. Sens. 7: Butera, M.K., Browder, J.A. and Frick, A.L. 1984. A prelimi- 1287-1301. nary report on the assessment of wetland productive capacity Fung, I.Y., Tucker, C.J. and Prentice, K.C. 1987. Advanced from a remote-sensing-based model - a NASA/NMFS joint very high resolution radiometer vegetation index to study research project. IEEE Transactions on Geoscience and Re- atmosphere-biosphere exchange of CO,. J. Geophys. Res. mote Sensing GE-22: 502-5 11. 92: 2999-3015. Center for Earth Resource Management. 1985. Integration of Gholz, H.L. 1982. Environmental limits on aboveground net Remotely Sensed Data in Geographic Information Systems primary production, leaf area, and biomass in vegetation for Processing of Global Resource Information. Proceedings zones of the Pacific Northwest. Ecology 63: 469-481. of 1985 CERMA Conference, Washington, D.C. Coward, S.N., Dye, D., Kerber, A. and Kalb, V. 1987. Compar- Choudhury, B.J. and Tucker, C.J. 1987. Monitoring global ison of North and South American biomes from AVHRR ob- vegetation using Nimbus-737 GHZ data - some empirical re- servations. Geocarto Int. 1: 27-39. lations. Int. J. Rem. Sens. 8: 1085-1090. Goward, S.N., Tucker, C. J. and Dye, D.G. 1985. North Ameri- Choudhury, B.J. 1988. Relating Nimbus-7 37 Ghz data to global can vegetation patterns observed with the NOAA-7 advanced land-surface evaporation, primary productivity, and the at- very high resolution radiometer. Vegetatio 64: 3-14. mospheric CO, concentration. Int. J. Rem. Sens. 9: Graham, R.L., Iverson, L.R. and Cook, E.A. 1987. Evaluating 169- 176. spatial patterns of forest productivity in a disturbed, moun- Colwell, R.N. (ed.) 1983. Manual of Remote Sensing. Am. SOC. tainous landscape using Landsat data and a GIS. Bull. Ecol. Photogr. Rem. Sens. Falls Church, Virginia. SOC.Am. 68: 3-4. Cook, E.A., Iverson, L.R. and Graham, R.L. 1987. The rela- Greegor, D.H. 1986. Ecology from space. BioScience 36: tionship of forest productivity to Landsat Thematic Mapper 429-432. data and supplemental terrain information. In Proceedings of Hall, F.G., Botkin, D.B., Strebel, D.E., Woods, K.D. and Pecora XI Symposium. pp, 43-52. Am. SOC.Photogr. Rem. Goetz, S.J. 1987. Ten year change in forest succession and Sens. Falls Church, Virginia. composition measured by remote sensing. NASA-CR- Cook, E.A., Iverson, L.R. andGraham, R.L. 1989. Estimating 180948, NAS 1.26: 180948. Washington, D.C. forest productivity with thematic mapper and biogeographic Hardisky, M.A., Daiber, F.C., Roman, C.T. and Klemas, V. data. Rem. Sens. Env. 27: 131-141. 1984. Remote sensing of biomass and annual net aerial Curran, P.J. and Willamson, H.D. 1986. Sample size for primary productivity of a salt marsh. Rem. Sens. Env. 16: ground and remotely sensed data. Rem. Sens. Env. 20: 31- 91-106. 41. Hopkins, P.F., MacLean, A.L. and Lillesand, T.M. 1988. As- Cushnie, J.L. and Atkinson, P. 1985. Effect of spatial filtering sessment of thematic mapper imagery for forestry applica- on scene noise and boundary detail in thematic mapper im- tions under Lake States conditions. Photogr. Eng. Rem. agery. Photogr. Eng. Rem. Sens. 51: 1483-1493. Sens. 54: 61-68. Danson, F.M. 1987. Preliminary evaluation of the relationships Iverson, L.R., Cook, E.A., Graham, R., Olson, J.S., Frank, between SPOT-1 HRV data and forest stand parameters. Int. T.D. and Ke, Y. 1988. Interpreting forest biome productivity J. Rem. Sens. 8: 1571-1575. and cover utilizing nested scales of image resolution and bio- Dean, M.E. and Hoffer, R.M. 1982. Evaluation of thematic geographical analysis. Final Report to NASA, Champaign, 11- mapper data and computer-aided analysis techniques for linois. mapping forest cover. Ind. Ac. Sci. 92: 203-210. Iverson, L.R., Cook, E.A. and Graham, R.L. 1989. A tech- 142

nique for extrapolating and validating forest cover across Norwine, J. and Greegor, D.H. 1983. Vegetation classification large regions: calibrating AVHRR data with TM data. Int. J. based on advanced very high resolution radiometer (AVHRR) Rem. Sens. 10 (in press). satellite imagery. Rem. Sens. Env. 13: 69-87. Iverson, L.R. and Risser, P.G. 1987. Analyzing long-term Olang, M.O. 1983. Vegetation classification using satellite im- changes in vegetation with geographic information system agery and area sampling frame to locate sampling stands. In and remotely sensed data. Adv. Space Res. 7: (11)183-194. Machine Processing of Remotely Sensed Data Symposium Jackson, R.D. 1986. Remote sensing of biotic and abiotic plant Proceedings. pp. 102-109. West Lafayette, Indiana. stress. Ann. Rev. Phytopath. 24: 265-287. Peterson, D.L., Aber, J.D., Matson, P.A., Card, D.H., Swan- Kimes, D.S., Sellers, P.J. and Diner, D.J. 1987. Extraction of burg, N., Wessman, C. and Spanner, M. 1988. Remote sens- spectral hemispherical reflectance (albedo) of surface from ing of forest canopy and leaf biochemical contents. Rem. nadir and directional reflectance data. Int. J. Rem. Sens. 8: Sens. Env. 24: 85-108. 1727- 1746. Peterson, D.L., Westman, W.E., Stephensen, N.J., Ambrosia, Knipling, E.B. 1970. Physical and physiological bases for the V.G., Brass, J.A. and Spanner, M.A. 1986. Analysis of reflectance of visible and near-infrared radiation from vegeta- forest/structure using thematic mapper simulator data. IEEE tion. Rem. Sens. Env. l: 155-159. Transaction on Geoscience and Remote Sensing GE-24: Kiichler, A.W. 1964. Potential natural vegetation of the conter- 113-120. minous United States. Am. Geogr. SOC.Spec. Publ. No. 36. Peterson, D.L., Spanner, M.A., Running, S.W. and Teuber, New York. K.R. 1987. Relationship of thematic mapper simulator data Lee, K.S. and Hoffer, R.M. 1988. Assessment of forest cover to leaf area index of temperate coniferous forests. Rem. Sens. changes using multidate spaceborne imaging radar. In Tech- Env. 22: 323-341. nical Papers, 1988. ACSM-ASPRS Annual Convention, Vol. Rock, B.N., Vogelmann, J.E., Williams, D.L., Vogelmann, 4. Image Processing/Remote Sensing. pp. 198-21 1. Am. A.F. and Hoshizaki, T. 1986. Remote detection of forest SOC.Photogr. Rem. Sens. Falls Church, Virginia. damage. Bioscience 36: 439-445. Lillesand, T.M. and Kiefer, R.W. 1987. Remote Sensing and Richardson, A.J. and Everitt, J.H. 1987. Monitoring water Image Interpretation (second edition). John Wiley and Sons, stress in buffelgrass using hand-held radiometers. Int. J. New York. Rem. Sens. 8: 1797-1806. Logan, T.L. 1985. An ecoregion-continuum approach to global Running, S.W. and Nemani, R.R. 1988. Relating seasonal pat- vegetative biomass estimation. In Proceedings, Tenth Pecora terns of the AVHRR vegetation index to simulated pho- Conference, Ft. Collins, Colorado. tosynthesis and transpiration of forests in different climates. Logan, T.L. and Bryant, N.A. 1987. Spatial data software in- Rem. Sens. Env. 24: 347-367. tegration: merging CAD/CAM/Mapping with GIS and im- Running, S.W., Peterson, D.L., Spanner, M.A. and Teuber, age processing. Photogr. Eng. Rem. Sens. 53: 1391-1395. K.B. 1986. Remote sensing of coniferous forest leaf area. Malila, W.A. 1985. Comparison of the information contents of Ecology 67: 213-276. Landsat TM and MSS data. Photogr. Eng. Rem. Sens. 51: Sader, S.A. 1987. Airborne remote sensing of forest biomes. 1449- 1457. Geocarto Int. 1: 9-17. Malingreau, J.P. and Tucker, C.J. 1987. The contribution of Sader, S.A. and Joyce, A.T. 1988. Deforestation rates and AVHRR data for measuring and understanding global trends in Costa Rica, 1940 to 1983. Biotropica 20: 11-19. processes: large-scale deforestation in the Amazon Basin. In Schneider, S.R. 1984. Renewable resource studies using the Proceedings of IGARSS’87 Symposium. pp. 443-448. Ann NOAA polar-orbiting satellites. In 35th International As- Arbor, Michigan. tronautical Federation Congress. pp. 15-22. Lausanne, Swit- Malingreau, J.P.and Tucker, C.J. 1988. Large-scale deforesta- zerland. tion in the southeastern Amazon basin of Brazil. Ambio 17: Sellers, P.J. 1985. Canopy reflectance, photosynthesis and 49-55. transpiration. Int. J. Rem. Sens. 6: 1335-1392. Nelson, R.F., Latty, R.S. and Mott, G. 1984. Classifying north- Sellers, P.J., Hall, F.G., Asrar, G., Stebel, D.E. and Murphy, ern forests using thematic mapper simulator data. Photogr. R.E. 1988. The first ISLSCP field experiment (FIFE). Bull. Eng. Rem. Sens. 50: 607-617. Am. Meteor. SOC.69: 21-27. Nelson, R., Horning, N. and Stone, T.A. 1987. Determining the Serafini, Y.V. 1987. Estimation of the evapotranspiration using rate of forest conversion in Mato Grosso, Brazil, using surface and satellite data. Int. J. Rem. Sens. 8: 1547-1562. LANDSAT MSS and AVHRR data. Int. J. Rem. Sens. 8: Shimoda, H., Fukue, K., Hosomura, T. and Sakota, T. 1986. 1767-1784. Global vegetation monitoring using NOAA vegetation index Nelson, R. and Holben, B. 1986. Identifying deforestation in data. In International Symposium on Space Technological Brazil using multiresolution satellite data. Int. J. Rem. Sens. and Science Proceedings. pp. 1625- 1630. AGNE Publishing, 7: 429-448. Inc., Tokyo, Japan. Nelson, R., Krabill, W. and Tonelli, J. 1988. Estimating forest Spanner, M.A., Brass, J.A. and Peterson, D.L. 1984. Feature biomass and volume using airborne laser data. Rem. Sens. selection and the information content of thematic mapper Env. 24: 247-267. simulator data for forest structural assessment. IEEE Trans- 143

actions on Geoscience and Remote Sensing GE-22: 482-489. Vogelmann, J.E. and Rock, B.N. 1988. Assessing forest damage Spanner, M.A., Peterson, D.L., Acevedo, W. and Matson, P. in high-elevation coniferous forests in Vermont and New 1985. High-resolution spectrometry of leaf and canopy Hampshire using thematic mapper data. Rem. Sens. Env. 24: chemistry for biogeochemical cycling. In Proceedings of the 227 -246. Airborne Imaging Spectrometer Data Analysis Workshop. Waring, R.H., Aber, J.D. Melillo, J.M. and Moore, B. 111. pp. 92-99. Edited by G. Vane and A.F.H. Goetz. Jet Propul- 1986. Precursors of change in terrestrial ecosystems. Bio- sion Laboratory Publication 85-41, Pasadena, California. science 86: 433-438. Stone, T.A. and Woodwell, G.M. 1988. Shuttle imaging radar Walker, J., Jupp, D.L.B., Penridge, L.K. and Tian, G. 1986. A analysis of land use in Amazonia. Int. J. Rem. Sens. 9: Interpretation of vegetation structure in Landsat MSS im- 95-105. agery: a case study in disturbed semi-arid eucalypt wood- Strong, L.L., Dana, R.W. and Carpenter, L.H. 1985. Estimat- lands. Part 1. Field data analysis. J. Env. Manage. 23: 19-33. ing phytomass of sagebrush habitat types from microdensito- Welch, R. 1985. Cartographic potential of SPOT image data. meter data. Photogr. Eng. Rem. Sens. 51: 467-474. Photogr. Eng. Rem. Sens. 51: 1085-1091. Toll, D.L. 1985. Effect of Landsat Thematic Mapper sensor Williams, D.L. and Nelson, R.F. 1986. Use of remotely sensed parameters on land cover classification. Rem. Sens. Env. 17: data for assessing forest stand conditions in the Eastern Unit- 129- 140. ed States. IEEE Trans. GeoSci. Rem. Sens. GE-24: 130-138. Townshend, J.R.G. and Justice, C.O. 1986. Analysis of the dy- Williams, D.L., Irons, J.R., Markham, B.L., Nelson, R.F., namics of African vegetation using the normalized difference Toll, D.L., Latty, R.S. and Stauffer, M.L. 1984. A statistical vegetation index. Int. J. Rem. Sens. 11: 1435-1445. evaluation of the advantages of Landsat thematic mapper Townshend, J.R.G., Justice, C.O. and Kalb, V. 1987. Charac- data in comparison to multispectral scanner data. IEEE terization and classification of South American land cover Trans. Geosci. Rem. Sens. GE-22: 294-302. types using satellite data. Int. J. Rem. Sens. 8: 1189-1207. Williams, D.L. and Stauffer, M.L. 1979. Forest insect defolia- Tucker, C.J. 1980. Remote sensing of leaf water content in the tion in Pennsylvania. In Monitoring Forest Canopy Altera- rear infrared. Rem. Sens. Env. 10: 23-32. tion Around the World with Digital Analysis of LANDSAT Tucker, C.J., Fung, I.Y., Keeling, C.D. and Gammon, R.H. Imagery. pp. 23-28. Edited by D.L. Williams and L.D. 1986. Relationship between atmospheric CO, variations and Miller. NASA, Washington, D.C. a satellite-derived vegetation index. Nature 319: 195-199. Woodwell, G.M., Hobbie, J.E., Houghton, R.A., Melillo, Tucker, C.J. and Sellers, P.J. 1986. Satellite remote sensing of J.M., Moore, B., Park, A.B., Peterson, B.J. and Shaver, primary production. Int. J. Rem. Sens. 7: 1395-1416. G.R. 1984. Measurement of changes in the vegetation of the Tucker, C.J., Holben, B. and Goff, T.E. 1984. Intensive forest earth by satellite imagery. In The Role of Terrestrial Vegeta- clearing in Rondonia, Brazil, as detected by satellite remote tion in the Global Carbon Cycle. pp. 221-240. Edited by sensing. Rem. Sens. Env. 15: 225-261. G.M. Woodwell. John Wiley and Sons, New York. Tucker, C.J., Townshend, T.R.G. and Goff, T. 1985. African Woodwell, G.M., Houghton, R.A. Stone, T.A., Nelson, R.F. land-cover classification using satellite data. Science 227: and Kovalick, W. 1987. Deforestation in the tropics: new 369-375. measurements in the Amazon basin using Landsat and Vogelmann, J.E. 1988. Detection of forest change in the Green NOAA AVHRR imagery. J. Geophys. Res. 92(D2): Mountains of Vermont using multispectral scanner data. Int. 21 57-2163. J. Rem. Sens. 9: 1187-1200. Wu, S.T. and Sader, S.A. 1987. Multipolarization SAR data for Vogelmann, J.E. and Rock, B.N. 1986. Assessing forest decline surface feature delineation and forest vegetation characteriza- in coniferous forests of Vermont Using NS-001 Thematic tion. IEEE Trans. Geosci. Rem. Sens. GE-25: 67-76. Mapper Simulator data. Int. J. Rem. Sens. 7: 1303-1321.