MONITORING GLOBAL CHANGE: COMPARISON OF FOREST COVER ESTIMATES USING REMOTE SENSING AND INVENTORY APPROACHES DAVID E TURNER, GREG KOERPER, H E R M A N N GUCINSKI and CHARLES PETERSON Mantech Environmental Technology, Inc., USEPA Environmental Research Laboratory, 200 S. W. 35th Street, Corvallis OR 97333, U. S. A.

and ROBERT K. DIXON U.S. Environmental Protection Agency, Environmental Research Laboratory, 200 S. W. 35th Street, Corvallis OR 97333, U. S. A. (Received: June 1992)

Abstract. Satellite-based remote sensing offers great potential for frequent assessment of forest cover over broad spatial scales, however, calibration and validation using ground-based surveys are needed. In this study, forest cover estimates for the United States from a recently developed land surface cover map generated from satellite remote sensing data were compared to state-level inventory data from the U.S. National Resources Planning Act Timber Database. The land cover map was produced at the U.S. Geological Survey EROS Data Center and is based on imagery from the AVHRR sensor (spatial resolution ~ 1.1 km). Vegetation type was classified using the temporal signal in the Normalized Difference Vegetation Index derived from AVHRR data. Comparisons revealed close agreement in the estimate of forest cover for extensively forested states with large polygons of relatively similar vegetation such as Oregon. Larger forest cover differences were observed in other states with some regional patterns in the level of agreement apparent. Comparisons in inventory- and remote sensing-based estimates of current forested area with potential vegetation maps indicated the magnitude of past land use change and the potential for future changes. The remote sensing approach appears to hold promise for conducting surveys of forest cover where inventory data are limited or where rates of vegetation change, due to human or climatic factors, are rapid.

1. Introduction

Emerging issues related to global change, such as quantification of the global carbon cycle (Post et al., 1990) and analysis of biospheric feedbacks to climate change (Lashof, 1989), require data describing land cover characteristics over large areas. Satellite-based remote sensing has emerged as a powerful tool for generation of land cover maps (Tucker et al., 1986; Iverson et al., 1989a). Early efforts relied primarily on the Multispectral Scanner aboard the NOAA Landsat Satellites, with classifications based on spectral differences between land cover types. Though proven capable of differentiating specific land use or vegetation types, the spatial extent of these analyses has been restricted to some degree by cost, computational constraints and problems related to cloud cover. An approach to developing extensive land cover maps which has recently been Environmental Monitoring and Assessment 26: 295-305, 1993. (~) 1993 Kluwer Academic Publishers. Printed in the Netherlands.

296

DAVID P. TURNER ET AL.

applied at the national scale is the use of the temporal signal in the normalized difference vegetation index (NDVI) or 'greenness index' (Loveland et al., 1991). Possible mechanisms for investigating the accuracy and reliability of land cover maps produced in this manner include examination of remote sensing data at higher spatial resolution (Iverson etal., 1989b), additional comparisons with existing maps or inventories, and validation through field surveys. In this paper we compare estimates of forest cover in the United States from a remote sensing-based land cover map with estimates from the U.S. Forest Service Resource Planning Act Timber Database. The objective is to identify geographic areas and land cover classes where agreement is good and areas where additional work is needed.

2. Background and Approach The areal extent of forest cover based on remote sensing was derived from a classification of satellite data developed at the U.S. Geological Survey EROS Data Center (Loveland et al., 1991). In that analysis, the seasonal shifts in the NDVI provided the basis for classification. The NDVI uses reflectance values for the visible ( V I S ) and near infrared ( N I R ) bands. NDVI = N I R

- VIS/NIR

+ VIS

Radiation in the visible band tends to be absorbed by foliage, specifically by chlorophyll, whereas the near infrared tends to be reflected (Curran, 1983). The ratio thus maximizes the signal associated with leafy material and in field-based studies NDVI has been correlated with vegetative characteristics such as foliar biomass and photosynthesis (Box et al., 1989). NDVI also minimizes the effects of solar angle, satellite view angle, atmospheric conditions and topography which makes it desirable for time series analysis. The satellite sensor used in the EROS Data Center (EDC) analysis was the Advanced Very High Resolution Radiometer (AVHRR) aboard the NOAA polar orbiting satellites (Goward et al., 1991). The spatial resolution of the AVHRR sensor is approximately 1.1 km and data from daily coverage may be composited over time to form periodic nearly cloud-free images. The EDC analysis was based on eight monthly maximum value NDVI composites from the 1990 growing season covering the 48 conterminous states. An unsupervised clustering algorithm identified 70 unique classes based on characteristics such as seasonal duration of greenness, time of greenness onset and magnitude of greenness peak. Postclassification stratification using elevation, frost-free period, and ecoregions led to the final classification with 167 classes. Descriptive names relating to vegetation type were associated with each class based on ancillary data including existing vegetation maps. The EDC map is not considered a finished product at this time as the vegetation classification scheme is expected to be revised in response to further analyses, and year-to-year updates may be generated to evaluate the effect of inter-annual climatic variability. The present study is in

COMPARISON OF FOREST COVER ESTIMATES

297

part an effort to identify geographic areas and land cover classes which require additional analysis. The designation of each of the 167 land cover types as forest or nonforest was based on the descriptors provided by EDC. In cases where a class was identified as a mixture of woodland and some other vegetation type, the class was counted as forest. Area comparisons were made on a state-by-state basis, and results were aggregated across states for determination of regional patterns. The inventory-based estimate of current forest area which was used in this analysis to evaluate the EDC map is described by Waddell et al. (1989). Under the Forest and Rangeland Renewable Resources Planning Act (RPA) of 1974, the U.S. Department of Agriculture was directed to assess all forest and rangeland resources on a ten-year cycle. The data of the 1989 publication represents the forest inventory as of 1987. Twenty-two forest types are considered in the RPA analysis. The RPA estimates of forest area are based on a combination of photoplots and ground survey plots taken from statistically based inventories, and have sampling error associated with them (e.g., Gedney et al., 1986). The RPA inventory data provides the most reliable estimate of current U.S. forest cover available. However, there remains considerable potential for error because of rapid rates of land use change, particularly in the southeastern U.S. (Williams, 1988), and because of the limitations in the sampling design. In order to gain an indication of the potential for conversion between forest and nonforest categories, additional comparisons were made with two other digitized maps of forest distribution in the U.S. One estimate of forest area was obtained from the map by K~Jchler (1964), printed by the U.S. Geological Survey in 1985, which characterizes potential natural vegetation. The 116 vegetation categories were aggregated into forest or nonforest classes to conform with the categorizations employed in the RPA inventory and the EDC aggregation. A second estimate of forest area was taken from the map produced by the Society of American Foresters (Eyre, 1980). The map was based on U.S. Forest Service inventory data for 1967 and the types of forest are identical to those in the RPA inventory. These maps were digitized using commercially available GIS software and resampled at a 1 km spatial resolution. The minimum spatial heterogeneity indicated on the original maps was on the order of 10 km. Thus, these maps may overestimate forest cover in geographic areas of high spatial heterogeneity such as the central Rocky Mountains. One additional analysis was made to identify possible misclassifications in the EDC map. Using the Kt~chler map as a reference, the area of grid cells classified by their NDVI temporal signature as forest, but not in potentially forested areas according to the Ktichler map, was determined on a state-by-state basis, with results reported at the regional level. All area determinations were made using GRASS, a public domain geographic information system developed and supported by the U.S. Army Corps of Engineers Construction Engineering Research Laboratory, Champaign, IL.

298

DAVID E TURNER ET AL.

Fig. 1. Area classified as forest based on the EDC map.

3. Results The EDC map (Figure 1) reduced to a binary classification (forest v s . nonforest) gives an indication of considerable patchiness in some regions, particularly in the southeast. State-by-state comparisons reveal close agreement in forest cover estimates between the EDC map and the RPA database for some states, but substantial differences for others (Table I). The forest area difference was less than 25% of the RPA estimate in 22 of the 48 states, covering 50% of total RPA forest area. The agreement in forest area estimates was closest in the west coast states where the regional total for forested area from the EDC map was only 5% less than that from the RPA database (Table II). In the Great Basin states (Arizona, New Mexico, Nevada, Utah), EDC estimates were generally lower than RPA (Table I), perhaps because of problems differentiating shrubland from chaparral and pinyon-juniper communities, both of which are considered forest in the RPA inventory. The EDC map appeared to underestimate forested area in states such as Georgia, South Carolina and North Carolina but to overestimate forest area in Missouri, Kentucky and Tennessee (Table I). There is no obvious explanation for these differences which indicates the need for additional study. However, these states did exhibit significant areas of mixed-type classes (i.e. cropland/woodland) which may in part account for these patterns. Total area classified as forest by EDC, but outside of potential forest area based on the Ktichler map, was greatest in the Rocky Mountain and west coast states

299

COMPARISON OF FOREST COVER ESTIMATES

TABLE I Comparisons of remotely-sensed and inventory-based forest cover estimates for the conterminous 48 states of the U.S. State

EDC 1

Alabama

RPA 2

EDC/RPA (Ha × 1000)

Ktichler 3

SAF 4

8 770

8 799

1.00

13 301

13 366

3 350 6 581 12 501 10 010 917 243 4 939 6 194 10 228 2 667 3 002 2 765 300 7 584

7 851 6 880 15 950 8 642 735 161 6 772 9 683 8 837 1 728 1 798 633 550 4 964

0.43 0.96 0.78 1.I6 1.25 1.51 0.73 0.64 1.16 1.54 1.67 4,37 0.55 1.53

9 916 13 681 20 352 9 537 1 282 516 11 529 15 099 10 212 14 471 9 409 8 686 4 875 10 427

10 330 12 597 18 898 12 037 1 124 334 12 234 14 985 9 930 7 624 5 695 4 932 3 077 10 265

6 530 5 276 1 561

5 623 7 174 1 066

1.16 0.74 1,46

9 429 8 14l 2 483

8 774 8 132 1 541

Massachussetts Michigan Minnesota Missisippi Missouri Montana Nebraska Nevada New Hampshire New Jersey New Mexico New York North Carolina North Dakota Ohio

1 213 8 181 9 793 8 370 9 538 10 928 357 446 1 735 647 4 707 8 992 5 802 58 4 603

1 254 7 380 6 717 6 761 5 072 8 874 292 3 616 2 034 804 7 503 7 604 7 651 186 2 960

0.97 1,11 1.46 1,24 1.88 1.23 1,22 0.12 0.85 0.81 0.63 1.18 0.76 0.31 1,55

2 044 14 577 14 302 12 320 18 105 13 114 1 967 5 841 2 375 1 876 12 383 12 315 12 652 2 643 10 561

I 744 13 377 14 295 11 371 I6 406 12 357 l 962 6 857 2 392 1 013 1 l 206 12 259 12 667 1 591 7 188

Oklahoma Oregon Pennsylvania Rhode Island South Carolina South Dakota Tennessee Texas

2 076 11 990 8 764 118 2 312 876 7 671 6 844

2 950 11 364 6 884 i62 4 964 684 5 370 5 531

0.70 1.06 1.27 0.73 0.47 1.28 1.43 1.24

9 031 14 560 11 738 266 7 889 2 006 10 830 21 11313

6 350 12 644 10 839 260 7 987 1 669 10 608 18 883

4 212 2 122

6 575 1 814

0.64 1.17

9 017 2 424

10 790 2 232

7 798 9 898 6 071 8 627 4 650 252 817

6 467 8 852 4 837 6 2115 4 036 243 252

1,21 1.12 1.26 1.39 1.15 1.04

10 154 10 498 6 277 13 645 6 164 446 023

10 026 10 224 5 911 9 766 7 144 407 895

Arizona Arkansas California Colorado Connecticut Delaware Florida Georgia Idaho Illinois Indiana Iowa Kansas Kentucky Louisiana Maine Maryland

Utah Vermont Virginia Washington West Virginia Wisconsin Wyoming United States

1 Loveland et al., 1991. 2 Waddell etal., 1989. 3 Kiichler, 1964. 4 Eyre, 1980.

300

DAVID E TURNER ET AL.

TABLE II Regional comparisons of forest cover estimates for the conterminous 48 states of the U.S. Region EDC 1 RPA2 EDC/RPA Ktichler3 SAF4 (Ha x 1000) Rocky Mountains 48 530 55 935 0.87 76 184 80 652 North Central 49 176 32 491 1.51 103 757 79 284 Great Plains 1 591 1 713 0.93 11 490 8 300 Northeast 37 660 34 529 1.09 51 736 47 783 West Coast 34 389 36 166 0.95 45 410 41 766 South Central 54 426 46 878 1.16 100 120 92 213 Southeast 27 046 35 538 0.76 57 325 57 898 United States 252 817 243 250 1.04 446 023 407 895 1 Loveland et al., 1991. 2 Waddell et al., 1989. 3 Kachler, 1964. 4 Eyre, 1980.

TABLE III Area classified as forest on the EROS Data Center map, but as non-forest on the Ktichler potential vegetation map.

Region

Rocky Mountains North Central Great Plains Northeast West Coast South Central Southeast

EDC forest area not in Kiichler forest area 12 241 1 746 726 281 2 910 1 264 573

EDC forest area not in Ktichler/total EDC forest area (Ha x 1000) 0.25 0.04 0.46 0.01 0.08 0.02 0.02

( T a b l e III). T h i s p a t t e r n w a s p r i m a r i l y a f u n c t i o n o f p o t e n t i a l f o r e s t d i s t r i b u t i o n s i n c e the e a s t e r n a n d s o u t h e r n states are n e a r l y 1 0 0 % p o t e n t i a l l y f o r e s t e d ( F i g u r e 2). C o m p a r i s o n o f p o t e n t i a l f o r e s t a r e a f r o m the K t i c h l e r a n d S A F m a p s to the e s t i m a t e d f o r e s t a r e a f r o m the R P A I n v e n t o r y a n d E D C m a p s u g g e s t s t h a t c u r r e n t f o r e s t a r e a is a p p r o x i m a t e l y 5 5 % o f p o t e n t i a l f o r e s t a r e a ( T a b l e I). T h e s m a l l e s t d e c l i n e s in f o r e s t a r e a are o b s e r v e d in the w e s t c o a s t states w h e r e the p o t e n t i a l f o r c o n v e r s i o n f r o m f o r e s t to a g r i c u l t u r a l l a n d is l i m i t e d b y c l i m a t i c a n d e d a p h i c

COMPARISON OF FOREST COVER ESTIMATES

301

Fig. 2. Area classified as forest on the potential vegetation map of Kfichler.

factors. Large reductions in forest area are apparent in the midwest and southeast. When comparing the state, regional and national levels, differences between the EDC and RPA forest cover estimates appear almost randomly distributed and tend to cancel out with increasing aggregation (Tables I and II). At the national level there was only a 4% difference in total area classed as forest using the remote sensing and inventory approaches.

4. Discussion The contrast between potential and actual forest area in the U.S. is indicative of the magnitude of land use change which has occurred in this country. Similar changes in forest cover are now occurring globally in tropical as well as temperate latitudes and at an increasing rate (Myers, 1991). The global deforestation in 1990 was approximately 20 million ha (WRI, 1990). The magnitude of these changes is believed to influence the global carbon cycle (Houghton et al., 1983), with recent estimates suggesting a net annual flux to the atmosphere on the order of 0.4-2.5 Pg of carbon from tropical deforestation (Houghton, 1991). Changes in land cover of this rate and extent are nearly impossible to track using ground-based inventory approaches. Accurate and timely monitoring by satellite remote sensing may be an effective alternative. Currently available satellite imagery include products derived from the SPOT

302

DAVID P. TURNER ET AL.

sensor (10 m resolution), the Thematic Mapper (25 m), the Multispectral Scanner (80 m) and the AVHRR (~ 1100 m). The spectral classes vary among these sensors, which significantly influences their potential for classifying vegetation. Images from the SPOT, Thematic Mapper and Multispectral Scanner sensors are retrieved and stored on an as-needed basis whereas the AVHRR images are more systematically treated. Each sensor type has inherent strengths and weaknesses, however, the daily coverage and the growing global network for capturing and storing imagery makes the AVHRR sensor the primary choice for many types of land surface monitoring. The development of the EDC map considered in this study is evidence of the potential for using satellite remote sensing in production of land cover maps at the national scale. The observed differences between forest cover estimates from the EDC map and the RPA Inventory do suggest some additional work is needed, although even the RPA Inventory is not a completely accurate and up-to-date accounting of forest cover. Nor is it comprehensively georeferenced which could provide opportunities for closer comparisons. Nevertheless, the patterns in the state-specific differences in forest cover estimates can provide directions for future research. The approach appears to be most accurate where climatic gradients generate relatively narrow ecotones between vegetation types (e.g. Tucker et al., 1985), where the physiognomic character of adjacent forest and nonforest vegetation types is distinctly different, and where polygons of homogeneous vegetation are large (e.g. Oregon). Problems in classification appear to occur where the landscape is characterized by spatial heterogeneity at scales of less than 1 km and where the transition zone from forested to nonforested vegetation types tends to be gradual, i.e., with extensive areas of savannah or tall shrubby vegetation. Areas of heavily-logged forest may present instances where a 1 km spatial resolution causes problems. A recent analysis of the amount of old-growth forest remaining in the Pacific Northwest used Thematic Mapper analysis (Tepley and Green, 1991) because forest management practices on public lands in this region mandate small (less than 100 acres) clear cuts, several of which might occur within an AVHRR grid cell (~ 250 acres). The NDVI signal from the shrubby vegetation characteristic of early successional stands may have a greater seasonal amplitude than that of adjacent conifer forests. The 500 m spatial resolution of the MODIS sensor anticipated to be aboard the Earth Observation Satellites, with launches planned for the late 1990s, may represent an optimum spatial resolution for many land surface analyses (Townshend and Justice, 1988; Townshend et al., 1991). Where the transition from forest vegetation to nonforest vegetation is gradual, the temporal signal in the NDVI may not be distinct. Improvements in differentiation may come with the inclusion of additional spectral information, such as reflectance from other regions of the infrared band, which reveals details about specific vegetation properties such as foliar lignin and nitrogen concentration (Wessman et al., 1988; McLellan et al., 1991). These characteristics may also

COMPARISON OF FOREST COVER ESTIMATES

303

vary temporally and aid in distinguishing vegetation types. Several near term approaches could contribute to refinement of the current EDC land cover map for the purposes of assessing forest cover. An initial step would be a closer examination of cases where the EDC map indicates forests but the potential vegetation maps indicate nonforest vegetation. Some of these differences may involve sites where trees are maintained only through the influence of human activity. Higher spatial resolution remote sensing using spectral signatures could provide a means to readily check specific sites. Higher resolution analysis would also aid in evaluating the problem of spatial heterogeneity within grid cells. A second approach may be to use additional information in the NDVI signal. In this analysis, a simple binary classification has been used to discriminate forest from non-forest. It may be more accurate to assign fractional forest values to classes involving heterogeneous landscape patterns based on the maximum NDVI value. One other issue requiring further analysis is the naming of land cover classes. As noted, some of the differences reported in the present study may express a lack of correspondence between what is considered forest in the EDC and RPA classification systems. The thresholds chosen to differentiate vegetation types are generally subjective and cannot be expected to conform between independent data sources. EDC has made correlational analyses between their land cover classes and other databases such as the USGS Land Use/Land Cover database, the Major Land Resource Areas, and the U.S. Environmental Protection Agency's Ecoregions. Additional investigations of class by class correlations with vegetation maps, such as Ktichler's (1964), will improve the quality and utility of the EDC map.

5. Conclusions Forest areas of the U.S. and the world are subject to rapid change due to environmental and anthropogenic factors. Satellite remote sensing offers significant potential as a tool for monitoring these changes and a variety of regional to global studies have already been made. Analyses using the temporal signal in the NDVI appear to be capable of land cover characterization at the 1 km spatial resolution at the national scale; however, accuracy depends on the level of fine-grained spatial heterogeneity in the vegetation, the steepness of the broad scale climatic gradients, and the degree of physiognomic differentiation between adjacent forest and nonforest areas. Continued refinements in the class separations and identifications are needed. In the longer term, the higher spatial and spectral resolution of anticipated satellite sensors is likely to greatly enhance the reliability of this approach.

Acknowledgements This work is a component of the U.S. EPA Global Climate Research Program, Global Mitigation and Adaptation Program, R.K. Dixon, Program Leader, at the EPA Environmental Research Laboratory, Corvallis, OR and done under contract

304

DAVIDP. TURNERET AL.

b y M a n T e c h E n v i r o n m e n t a l T e c h n o l o g y , Inc. W e thank D. C o f f e y for her helpful c o m m e n t s and criticism.

Note

The information in this document has been funded wholly by the U.S. Environmental Protection Agency. It has been subjected to the Agency's peer and administrative review, and it has been approved for publication as an EPA document. Mention of trade names or commercial products does not constitute endorsement of recommendation for use. References Box, E.O., Holben, B.N., and Kalb, V.: 1989, 'Accuracy of the AVHRR Vegetation Index as a Predictor of Biomass, Primary Productivity and Net CO2 Flux', Vegetatio 80, 71-89. Curran, EJ.: 1983, 'Multispectral Remote Sensing for the Estimation of Green Leaf Area Index', Phil. Trans. R. Soc. Lond. A 309, 257-270. Eyre, EH. (Ed.): 1980, Forest Cover Types of the United States and Canada, Soc. Amer. Foresters, Washington, DC., 148 pp. Gedney, D.R., Bassett, P.M., and Mei, M.A.: 1986, Timber Resource Statisticsfor Non-Federal Forest Land in Southwest Oregon, USDA Forest Service Resource Bulletin PNW-138, Portland, Or, 26 pp. Goward, S.M., Markham, B., Dye, D.G., Dulaney, W., and Yang, J.: 1991, 'Normalized Difference Vegetation Index Measurements from the Advanced Very High Resolution Radiometer', Remote Sens. Environ. 35, 257-277. Houghton, R.A., Hobbie, J.E., Melillo, J.M., Moore, B., Peterson, B.J., Shaver, G.R., and Woodwell, G.M." 1983, 'Changes in the Carbon Content of Terrestrial Biota and Soils between 1860 and 1980: Net Releases of CO2 to the Atmosphere', EcoL Monog. 53, 235-263. Houghton, R.A.: 1991, 'Tropical Deforestation and Atmospheric Carbon Dioxide', Clim. Change 19, 99-118. Iverson, L.R., Cook, E.A., and Graham, R.L.: 1989a, 'A Technique for Extrapolating and Validating Forest Cover across Large Regions: Calibrating AVHRR Data with TM Data', Int. J. Remote Sensing 10, 1805-1812. Iverson, L.R., Graham, R.L., and Cook, E.A.: 1989b, 'Applications of Satellite Remote Sensing to Forested Ecosystems', Lands. EcoL 3, 131-143. Ktichler, A.W.: 1964, Potential Natural Vegetation of the Conterminous United States, Amer. Geogr. Soc., Spec. Publ. 36, 110 pp. Lashof, D.A.: 1989, 'The Dynamic Greenhouse: Feedback Processes That May Influence Future Concentrations of Atmospheric Trace Gases and Climate Change', Clim. Change 14, 213-242. Loveland, T.R., Merchant, J.W., Ohlen, D.O., and Brown, J.E: 1991, 'Development of a LandCover Characteristics Database for the Conterminous U.S.', Photogram. Eng. Rem. Sens. 57, 1453-1463. McLellan, T.M., Martin, M.E., Aber, J.D., Melillo, J.M., Nadelhoffer, K.J., and Dewey, B.: 1991, 'Comparison of Wet Chemistry and Near Infrared Reflectance Measurements of Carbon-Fraction Chemistry and Nitrogen Concentration of Forest Foliage', Can. J. For. Res. 21, 1689-1693. Myers, N.: 1991, 'Tropical Forests: Present Status and Future Outlook', CIim. Change 19, 3-32. Post, W.M., Peng, T.-H., Emanuel, W.R., King, A.W., Dale, V.H., and DeAngelis, D.L.: 1990, 'The Global Carbon Cycle', Am. Scientist 78, 310-326. Tepley, J., and Green, K.: 1991, 'Old Growth Forest - How Much Remains?', Geog. Info. Syst. 1, 31-32. Townshend, J.R.G., and Justice, C.O.: 1988, 'Selecting the Spatial Resolution of Satellite Sensors Required for Global Monitoring of Land Transformations', Int. J. Remote Sensing 9, 187-236.

COMPARISONOF FOREST COVERESTIMATES

305

Townshend, J., Justice, C., Li, W., Gurney, C., and McManus, J.: 1991, 'Global Land Cover Classification by Remote Sensing: Present Capabilities and Future Possibilities', Remote Sens. Environ. 35, 243-255. Tucker, C.J., Townshend, J.R.G., and Goff, T.E.: 1985, 'African Land-Cover Classification Using Satellite Data', Science 227, 369-375. Tucker, C.J., Townshend, J.R.G., Goff, T.E., and Holben, B.N.: 1986, 'Continental and Global Scale Remote Sensing of Land Cover', In: Trabalka, J.R. and Reichle, D.E. (Eds.) The Changing Carbon Cycle: A Global Analysis, Springer-Verlag, New York, pp. 221-241. Waddell, K.L., Oswald, D.D., Powell, D.S.: 1989, Forest Statistics of the United States, USDA Forest Service Res. Bull., PNW-RB-168, 106 pp. Wessman, C.A., Aber, J.D., Peterson, D.L., and Melillo, J.M.: 1988, 'Remote Sensing of Canopy Chemistry and Nitrogen Cycling in Temperate Forest Ecosystems', Nature 335, 154-156. Williams, M.: 1988, 'The Death and Rebirth of the American Forest: Clearing and Reversion in the United States, 1900-1980', In Richards, J.F. and Tucker, R.P. (Eds.) World Deforestation in the Twentieth Century, Duke University Press, Durham, NC. World Resources Institute (WRI): 1990, Worm Resources 1990-91, Oxford University Press, Oxford, 383 pp.

Monitoring global change: Comparison of forest cover estimates using remote sensing and inventory approaches.

Satellite-based remote sensing offers great potential for frequent assessment of forest cover over broad spatial scales, however, calibration and vali...
656KB Sizes 0 Downloads 0 Views