Requested Form of Acknowledgment.

Please cite the following publication whenever these data are used:

DeFries, R. S. and J. R. G. Townshend, 1994, NDVI-derived land cover classification at a global scale. International Journal of Remote Sensing, 15:3567-3586.

Objective/Purpose.

The data set describes the geographic distributions of eleven major cover types based on interannual variations in NDVI.

The data set was developed to explore the conceptual and methodological issues that arise when using the Normalized Difference Vegetation Index (NDVI) as a basis for global classification of vegetative land cover. The purpose of the study is to use satellite data to improve currently available information on global land cover for applications to global change research.

Discussion.

Phenological differences among vegetation types, reflected in temporal variations in NDVI derived from satellite data, have been used to classify land cover at continental scales. This study explored methodologies for extending this concept to a global scale. A coarse resolution (one by one degree) data set of monthly NDVI values for 1987 (Los, et al. 1994, Sellers, et al. 1994, 1995b) was used as the basis for a supervised classification of eleven cover types (see section on data description for a list of the cover types) that broadly represent the major biomes of the world. Because of missing values at high latitudes, the Pathfinder AVHRR data set for 1987 (James and Kalluri, 1994) for summer monthly NDVI and red reflectance values were used to distinguish the following cover types: tundra, high latitude deciduous forest and woodland, coniferous evergreen forest and woodland.

The eleven cover types were selected primarily to conform with the cover types required as input to climate models. Training sets for each of the eleven cover types were identified as the areas where three existing ground-based data sets of global land cover (Matthews 1983, Olson, et al. 1983, Wilson and Henderson-Sellers 1985) agree that the land cover is present (DeFries and Townshend 1994b).

The global land cover data set is the result of a maximum likelihood classification of eleven cover types. The data set has not been systematically validated. Cursory validation indicates that the user should be aware of the following problems: 1) the distinction between "cultivated" and "grassland" cover types may be inaccurate because the NDVI temporal profiles of these two cover types are not significantly distinct, and 2) the "tundra" cover type may be inaccurate because of missing data at high latitudes.

Processing Description.

The global land cover data set was based on AVHRR maximum monthly composites for 1987 of NDVI values at approximately 8 km resolution, averaged to one by one degree resolution (Los, et al 1994). A Fourier transform was applied to smooth the temporal profiles and remove aberrant low values (Sellers, et al. 1994). At high northern latitudes, the data set was based on the AVHRR Pathfinder data set for 1987 (James and Kalluri, 1994), resampled to a spatial resolution of one by one degree and composited to obtain maximum monthly NDVI values and corresponding red reflectance values for summer months.

Spatial Coverage.

The coverage is global. Data in file are ordered from North to South and from West to East beginning at 180 degrees West and 90 degrees North. Point (1,1) represents the grid cell centered at 89.5 N and 179.5 W.

For additional information on acquisition and processing of the data sets that were used to derive global land cover, see Los, et al. (1994), Sellers, et al. (1994, 1995b), and James and Kalluri (1994).

Spatial Resolution.

The data are given in an equal-angle latitude / longitude grid that has a spatial resolution of 1-by-1 degree latitude / longitude.

Land Cover Data Description.

     Value  Land Cover Class
     -----  -------------------------------------------
     0      water
     1      broadleaf evergreen forest 
     2      coniferous evergreen forest and woodland
     3      high latitude deciduous forest and woodland
     4      tundra
     5      mixed coniferous forest and woodland
     6      wooded grassland
     7      grassland
     8      bare ground
     9      shrubs and bare ground
     10     cultivated crops
     11     broadleaf deciduous forest and woodland
     12     data unavailable

Data Format.

The database is provided in both binary and ASCII format.

The dimensions of the database are 360 pixels by 180 lines.

Click here for link to download location.

Derivation Techniques/Algorithms.

Maximum likelihood classification based on 12 monthly NDVI values was used to obtain the global land cover data set. In outline, the maximum likelihood procedure classifies each pixel to the land cover type that it most resembles in terms of its remotely sensed properties. The remotely sensed properties are used to define a multi-dimensional space within which pixels of each cover type can be located. The mean vector and variance-covariance matrix for each cover type are estimated using its worldwide population of pixels from the training set. Then, using the maximum likelihood rule (Swain and Davis 1978), the multidimensional space is partitioned into sub-spaces each uniquely associated with one land cover type. The whole of the global land mass is then classified according to the remotely sensed properties of each pixel. Thus, if a pixel falls within the sub-space associated with cover type c_i, it is labeled c_i. If the pixel falls within the sub-space associated with cover type c_j, it is labeled as that cover type, c_j.

Processing Steps and Data Sets.

To account for phasing of seasons, maximum likelihood classification was based on monthly NDVI values sequenced from the peak value at each pixel (see DeFries and Townshend (1994a) for more detail).

Training sets for each of the eleven cover types were identified as the areas where three existing ground-based data sets of global land cover (Matthews 1983, Olson, et al. 1983, Wilson and Henderson-Sellers 1985) agree that the land cover is present. Although there is considerable disagreement among these data sets (DeFries and Townshend 1994b), the locations where the three data sets agree were selected as those with the greatest confidence that the cover type actually exists on the ground. The following steps were taken to ensure that each training set was as spectrally distinct as possible or to further subdivide the training set so that each would be spectrally distinct:

1) each training set was split into Northern and Southern Hemispheres to account for phasing of seasons in the two hemispheres.

2) the feature space occupied by each training set was visually examined. Pixels that were obvious outliers were removed, and clusters were examined to determine if they were falling in different geographic areas. Where this was the case, the training set was subdivided. The most obvious example where subdivision was required was cultivated crops whose spectral signatures vary considerably among continents.

3) Bhattacharrya Distances--a measure of the separability of the training sets--and overlaps in the feature space were examined to determine if some cover types should be combined. This was the case, for example, for Southern Hemisphere broadleaf deciduous forest located mainly in Africa and Southern Hemisphere wooded grassland.

Special Corrections/Adjustments.

The global land cover data set was modified from the original maximum likelihood classification result as follows to eliminate stray pixels that were obviously incorrectly classified: pixels falling within training areas that were not correctly classified were changed to the cover type indicated by the training area; pixels surrounded on all sides by a different cover type were changed to that cover type; pixels classified as broadleaf evergreen in mid-latitudes were changed to the wooded grassland cover type; pixels classified as coniferous evergreen within the tropics were changed to the broadleaf evergreen cover type; pixels classified as mixed deciduous and evergreen forest and woodland within the tropics were changed to the wooded grassland cover type. In total, these changes altered approximately 10 percent of the total land surface.

Sources of Error.

Wintertime NDVI values were missing for large areas in high latitudes in the primary data set used for this study (Los, et al., 1994) . For these areas, results from a maximum likelihood classification using AVHRR Pathfinder data (James and Kalluri, 1994) for summertime monthly NDVI and red reflectance values were used.

Journal Articles and Study Reports.

DeFries, R. S. and J. R. G. Townshend, 1994a, NDVI-derived land cover classification at global scales. International Journal of Remote Sensing, 15:3567-3586. Special Issue on Global Data Sets.

DeFries, R. S. and J. R. G. Townshend, 1994b. Global land cover: comparison of ground-based data sets to classifications with AVHRR data. In Environmental Remote Sensing from Regional to Global Scales, edited by G. Foody and P. Curran, Environmental Remote Sensing from Regional to Global Scales. (U.K.: John Wiley and Sons).

James, M. E. and S. N. V. Kalluri, 1994. The Pathfinder AVHRR land data set: An improved coarse resolution data set for terrestrial monitoring. International Journal of Remote Sensing, Special Issue on Global Data Sets. 15(17):3347-3363.

Kuchler, A.W., 1983, World map of natural vegetation. Goode's World Atlas, 16th ed., Rand McNally, 16-17.

Leemans, R., and W. P. Cramer, 1991, The IIASA database for mean monthly values of temperature, precipitation and cloudiness on a global terrestrial grid, technical report, International Institute for Applied Systems Analysis, Laxenburg, Austria.

Los, S.O., C.O. Justice, C.J. Tucker, 1994. A global 1 by 1 degree NDVI data set for climate studies derived from the GIMMS continental NDVI data. International Journal of Remote Sensing, 15(17):3493- 3518.

Matthews, E., 1983. Global vegetation and land use: new high resolution data bases for climate studies. Journal of Climate and Applied Meteorology, 22: 474-487.

Olson, J. S., Watts, J. and L. Allison, 1983. Carbon in live vegetation of major world ecosystems. W-7405-ENG-26, U.S. Department of Energy, Oak Ridge National Laboratory.

Sellers, P.J., S.O. Los, C.J. Tucker, C.O. Justice, D.A. Dazlich, G.J. Collatz, and D.A. Randall, 1994. A global 1*1 degree NDVI data set for climate studies. Part 2: The generation of global fields of terrestrial biophysical parameters from the NDVI. International Journal of Remote Sensing, 15(17):3519-3545.

Sellers, P.J., D.A. Randall, C.J. Collatz, J.A. Berry, C.B. Field, D.A. Dazlich, C. Zhang, and C.D. Collelo, 1995a. A revised land surface parameterization (SiB2) for atmospheric GCMs. Part 1: Model formulation. submitted to Journal of Climate.

Sellers, P.J., S.O. Los, C.J. Tucker, C.O. Justice, D.A. Dazlich, G.J. Collatz, and D.A. Randall, 1995b. A revised land surface parameterization (SiB2) for atmospheric GCMs. Part 2: The generation of global fields of terrestrial biophysical parameters from satellite data. submitted to Journal of Climate.

Swain, P. H. and S. M. Davis, (ed.), 1978. Remote Sensing: The Quantitative Approach. (New York: McGraw-Hill Book Company).

Wilson, M. F. and A. Henderson-Sellers, 1985. A global archive of land cover and soils data for use in general circulation models. Journal of Climatology, 5: 119-143.

CD-ROM Distribution of This Dataset.

A derivation of this dataset was also distributed as the land cover / vegetation type grid cell data for the first ISLSCP compact disk set for climate modelers. The classes included in the ISLSCP data set are somewhat different because some classes are divided into C3 and C4 sub-classes. The ISLSCP CD-ROM set is entitled:

Global Data Sets for Land-Atmosphere Models; ISLSCP Initiative 1: 1987 - 1988; Volumes 1 - 5.

This dataset, published in March 1995, contains five compact disks of commonly gridded global data. It is available at no cost from the GSFC DAAC. The data distribution is part of NASA's Mission to Planet Earth. For further information, contact:

Glossary of Acronyms.

     AVHRR      Advanced Very High Resolution Radiometer
     CD-ROM     Compact Disk (optical), Read Only Memory
     DAAC       Distributed Active Archive Center
     GCM        General Circulation Model of the atmosphere
     GSFC       Goddard Space Flight Center
     ISLSCP     International Satellite Land Surface Climatology Project
     NASA       National Aeronautics and Space Administration
     NDVI       Normalized Difference Vegetation Index

Acknowledgments.

This project was supported by NASA under contract NAGW-2723. Thanks also goes to Chris Justice, Sietse Los, and Piers Sellers for providing the NDVI dataset as well as information about the Simple Biosphere Model. Hong Zhang and Alice Cialella at the Department of Geography at College Park provided technical assistance. Rob Sohlberg assembled the data for distribution on the WWW.

(this document was last updated 27-OCT-1995)