int. j. remote sensing, 2001, vol. 22, no. 8, 1479–1493
Land-cover classi?cation methods for multi-year AVHRR data
Laboratory for Global Remote Sensing Studies, Department of Geography,
University of Maryland, College Park, Maryland 20742,USA;
(Received 28 April 1998; in ?nal form 4 October 1999)
extensively used for global land-cover classi?cation, but few studies have taken
direct and full advantageof the multi-year propertiesofAVHRR data.This study
focused on generating eVective classi?cation features from multi-year AVHRR
data to improve classi?cation accuracy.Three types offeatures were derived from
12-year monthly composite normalized diVerence vegetation index (NDVI) and
channel 4 brightness temperature from the NOAA/NASA Path?nder AVHRR
Land data for land-cover classi?cation. The ?rst is based on the shape of the
annual averageNDVI or brightness-temperaturepro?le, which was then approxi-
mated by a Fourier series.The coeYcients estimatedby the weightedleast-squares
method were used for classi?cation. The second and third features were based on
the raw periodogram of the time series and the auto-regressive modelling. A
global land-cover training database created from Landsat Thematic Mapper and
Multi-spectral Scanner imagery was used for training and validation.Both quad-
rature discriminate analysis (QDA) and linear discriminate analysis (LDA) were
explored for classi?cation, and results indicate that QDA performs much better
than LDA. The ?rst feature, based on the mean annual shape, produced much
better results than the other two features. It was also found that NDVI signals
worked better than brightness-temperaturesignals.That is probablybecause top-
of-atmosphere signals were used, and atmospheric contaminations signi?cantly
disturb the thermal signals and correlation structures of diVerent cover types.
Further validations are needed.
Land-cover maps are needed for global climate and ecosystem process models,
as well as to characterize the distribution and status of major land surface types for
environmental and ecological applications. Because of the temporal dynamics and
changes in land surface, remote sensing is the only practical means for monitoring
DiVerent global, continental and regional land-cover datasets have been derived
from remotely sensed data, particularly from Advanced Very High Resolution
Radiometer (AVHRR) data, using various methods (Tucker et al. 1985, 1991,
Malingreau 1986,Malingreau et al.1989,Loveland et al.1991,Loveland and Belward
1997,Townshend et al. 1991,Townshend 1994,DeFries et al. 1995,Cihlar et al. 1996,
Nemani and Running 1997,Gopal et al.1999). Because AVHRR has relatively coarse
spectral and spatial resolution, most studies have utilized its multi-temporal charac-
teristics to classify land cover. However, multi-year characteristics of AVHRR data
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online © 2001 Taylor & Francis Ltd
have not been fully exploited. Data from AVHRR sensors have been acquired since
1981 and are expected to continue. The long temporal history is a very useful source
to characterize land surface cover types, but it also requires advanced techniques to
extract useful information. There exist numerous time series analysis techniques in
the literature (e.g. Brockwell and Davis 1987), and several techniques have been
explored to classify land-cover types using AVHRR data, such as Fourier analysis
(Andres et al.1994), change-vector analysis(Malila 1980,Lambin and Strahler 1994),
principal component analysis (Anyamba and Eastman 1996) and others (e.g.Nemani
and Running 1997), but few studies have taken direct and full advantage of the
multi-year properties of AVHRR data.
Land-cover classi?cation accuracy depends on many diVerent factors, of which
the eVective features and the classi?cation method are critical. DiVerent land-cover
classi?cation algorithms have been explored in the literature, but the determination
ofeVectivefeatures is still farfrom mature.Becauseofserious atmospheric contamina-
tion, the original two bands in the visible and near-infrared spectrum can not be
used for eVective features. Instead, vegetation indices such as normalized diVer-
ence vegetation index (NDVI) have been widely used for global land-cover monit-
oring. Determining characteristic features from multi-temporal NDVI analysis and
other AVHRR band combinations for global land-cover classi?cation is still a
The present study focused on deriving the eVective features for classifying global
land-cover types using the monthly composite NOAA/NASA Path?nder AVHRR
Land data. Three types of features were explored. The ?rst was based on the shape
of the annual mean pro?le of the multi-year signals (NDVI or brightness temper-
ature), which was approximated by a truncated third-order Fourier series. The
coeYcients estimated by the weighted least-squares method constituted the ?rst type
of feature. The second was the subset of the raw periodogram of the multi-year
signals. The last was based on the auto-regressive (AR) model whose coeYcients
were used for classi?cation.
A training database of 13 cover types for global land-cover classi?cation has
been created (DeFries et al. 1998). These pixels were extracted from the monthly
composite dataset. One-half of these pixels were used for training and the other half
for validation. Both quadrature discriminate analysis (QDA) and linear discriminate
analysis (LDA) were used.
The AVHRR sensor has ?ve bands, one in red, one in near-infrared, one in mid-
infrared and two in thermal-infrared. Its spatial resolution varies from 1.1km at sub-
nadir to more than 10km oV-nadir. More detailed characterizations can be found
in many textbooks (e.g. Cracknell 1997).
NOAA/NASA-sponsored AVHRR Land Path?nder dataset has been created to
act as a precursor for the international Earth Observing Systems (EOS). In the
Path?nder AVHRR Land data, attempts have been made to eliminate some factors
(James and Kalluri 1994). For example, Rayleigh scattering and ozone absorption
have been corrected. All ?ve original channels have been calibrated using the post-
launch calibration algorithms. All pixels are accurately navigated and solar and scan
angles are provided. The spatial resolution of all pixels have been normalized to be
8km. The 12-year corrected, monthly composite dataset from 1982 to 1993 were
used in this study.
Land-cover classi?cation methods for multi-year AVHRR data
A training database of global land cover has been compiled based on high-spatial
resolution remotely sensed imagery and ancillary data (DeFries et al. 1998). A total
of 169 Landsat scenes, mostly from the Multispectral Scanner system (MSS) was
used to identify over 9000 pixels in the Path?nder 8km resolution data where there
is high con?dence that the labelled cover type occurs. A total of 13 cover types
(table 1)from the original database (DeFries et al.1998)were considered in this study.
3. Algorithm descriptions
The algorithms for extracting three types of features are discussed below. The
coeYcients of these models were input to classi?ers for discriminating diVerent
3.1. Weighted least-squares method
This approach consisted of two major steps. The ?rst was to generate a mean
annual pro?le. Because of the presence of noisy pixels, the median value over 12
years was used as the average. Mathematically,
The second step was to approximate the annual mean pro?le by a truncated
low-order Fourier series:
aiand biare Fourier coeYcients. The phase term is de?ned as
where n=12 represents 12 months and j ranges from 1 to n. After some initial
experiments, it was found that I=3 is a good choice; thus, there were seven
coeYcients to be estimated.
An ordinary least-squares procedure can be used to estimate these Fourier
coeYcients. However, the ?tted function may not represent the real seasonal trend
denotes the monthly averaged observations from January (j=1) to
Table 1.Land-cover types and the training pixels.
Cover typesNumber of pixels
Evergreen needleleaf forests
Evergreen broadleaf forests
Deciduous needleleaf forests
Deciduous broadleaf forests
Closed bushlands or shrubland
Mosses and lichens
Andres, L., Salas, W. A., and Skole, D., 1994, Fourier analysis of multi-temporal AVHRR
data applied to a land cover classi?cation. International Journal of Remote Sensing,
Anyamba, A., and Eastman, J. R., 1996, Interannual variability of NDVI over Africa and its
relation to El Nino/Southern Oscillation. InternationalJournal of Remote Sensing, 17,
Brockwell, P. J., and Davis, R. A., 1987, Time Series: theory and methods (New York:
Cihlar,J., Hung,L., and Xiao,Q., 1996,Land coverclassi?cationwith AVHRRmultichannel
composites in northern environments.Remote Sensing of Environment, 58, 36–51.
Cracknell, A. P., 1997, The Advanced Very High Resolution Radiometer (London: Taylor
DeFries, R., Hansen, M., and Townshend, J., 1995, Global discrimination of land cover
types from matrices derived from AVHRR Path?nder Data. Remote Sensing of
Environment, 54, 209–222.
DeFries, R., Hansen, M., Townshend, J., and Sohlberg, R., 1998, Global land cover
classi?cations at 8km spatial resolution.Part 1: Training and validation data derived
from Landsat imagery. International Journal of Remote Sensing, 19, 3141–3168.
Gopal, S., Woodcock, C. E., and Strahler, A. H., 1999,Fuzzy neural network classi?cation
ofgloballand cover from a 1 degree AVHRR data set. Remote Sensing ofEnvironment,
Holben, B. N., 1986, Characteristics of maximum-value composite images from temporal
AVHRR data. International Journal of Remote Sensing, 7, 1417–1434.
James, M. E., and Kalluri, S., 1994, The Path?nder AVHRR land data set: an improved
coarse resolution data set for terrestrial monitoring. International Journal of Remote
Sensing, 15, 3347–3364.
Lambin, E. F., and Strahler, A. H., 1994, Change-vector analysis in multitemporal space:
a tool to detect and categorize land-cover change processes using high temporal-
resolution satellite data. Remote Sensing of Environment, 48, 231–244.
Loveland, T. R., and Belward, A. S., 1997, The IGBP-DIS global 1km land cover data set,
DISCover: ?rst result. International Journal of Remote Sensing, 18, 3289–3296.
Loveland, T. R., Merchant, J. W., Ohlen, D. O., and Brown, J. F., 1991, Development of
a land-covercharacteristicsdatabasefor continuousU.S.PhotogrammetryEngineering
and Remote Sensing, 57, 1453–1463.
Malila, W. A., 1980, Change vector analysis: an approach for detecting forest changes with
Landsat.Proceedings of the 6th Annual Symposiumon Machine Processing of Remotely
Sensed Data, Purdue University, West Lafayette, Indiana, 3–6 June 1980, pp.326–335.
Malingreau, J. P., 1986, Global vegetation dynamics: satellite observations over Asia.
International Journal of Remote Sensing, 7, 1121–1146.
Malingreau, J.-P., Tucker, C. J., and Laporte, N., 1989, AVHRR for monitoring global
tropical deforestion.International Journal of Remote Sensing, 10, 855–867.
McGregor, J., and Gorman, A. J., 1994, Some considerations for using AVHRR data in
climatological studies: I. Oribital characteristics of NOAA satellites. International
Journal of Remote Sensing, 15, 537–548.
Nemani, R., and Running, S., 1997, Land cover characterization using multitemporal red,
near-IR,and thermal-IRdata from NOAA/AVHRR. Ecological Applications,7,79–90.
Privette, J. L., Fowler, C., Wick, G. A., Baldwin, D., and Emery, W. J., 1995, EVects of
orbital drift on advanced very high resolution radiometer products:normalized diVer-
ence vegetation index and sea surface temperature. Remote Sensing of Environment,
Roderick, M., Smith, R., and Lodwick, G., 1996, Calibrating long-term AVRHH-derived
NDVI imagery. Remote Sensing of Environment, 58, 1–12.
Sellers, P. J., Sietse, S. O., Tucker, C. J., Justice, C. O., Dazlich, D. A., Collatz, G. J.,
and Randall, D. A., 1996, A revised land surface parameterization (SiB2) for
Atmospheric GCMs. Part II: The generation of global ?elds of terrestrial biophysical
parameters from satellite data. Journal of Climate, 9, 706–737.
Land-cover classi?cation methods for multi-year AVHRR data
Townshend, J. G. R., 1994, Global data sets for land applications from the Advanced Very
High Resolution Radiometer:an introduction.InternationalJournalofRemote Sensing,
Townshend, J. R. G., Justice, C. O., Li, W., Gurney, C., and McManus, J., 1991, Global
land cover classi?cation by remotesensing:presentcapabilitiesand future possibilities.
Remote Sensing of Environment, 35, 243–255.
Tucker, C. J., Townshend, J. R. G., and Goff, T. E., 1985, African land-cover classi?cation
using satellite data. Science, 227, 369–375.
Tucker, C. J., Dregne, H. E., and Newcomb, W. W., 1991, Expansion and contraction of
the Sahara Desert from 1980 to 1990. Science, 253, 299–301.
Venables, W. N., and Ripley, B. D., 1994,Modern Applied Statistics with S-Plus (New York: