GISdevelopment.net ---> AARS ---> ACRS 1990 ---> Global Change of Environment

Global Land Cover Classification by NOAA AVHRR data

Takayuki Odajima, Koji Kajiwara, Ryutaro Tateishi
Remote Sensing and Image Research Center, Chiba University
1-33, Yayoi-Cho, Chiba City, Chiba, 260 Japan


Abstract
For the purpose of global scale and cover classification, cluster analysis was performed two times to monthly global vegetation index (GVI) data. First cluster analysis was applied separately to the north hemisphere and south hemisphere. N both hemispheres, 31 clusters were obtained form 40 initial clusters. Second cluster analysis was applied to both hemispheres with the data by using 62 centroids of resultants clusters of the first cluster analysis as initial clusters. 44 clusters were obtained by the second cluster analysis. As a result, land cover map of the world with 13 types was produced fro the viewpoint of seasonal vegetation dynamics.

Introduction
Recently, the problems of the earth environment such as deforestation, desertification, destruction of ozone layer … and so on , have been the object of public interest, and many papers have been reported about that. NOAA data especially GVI data is the great tools for monitoring global vegetation distribution. Many phonological studies have been achieved by using NOAA data (justice et al., 1986; Townsend et al., 1989 etc)

Authors have studies by using NOAA data about classifying and monitoring the land cover in eh Asian Region (Tateishi et. al 1990). Since the NOAA data has the potential for analyzing the larger region, the study area was extended from the Asian region to eh whole earth and classified global and covers by cluster analysis of NOAA AVHRR data in 1987.

Data
The following data were used in this study
  1. NOAA AVHRR weekly channel data (Ch1, Ch2) in 1987

  2. NOAA weekly GVI data in 1987

  3. Global DEM data (ETOPO 5)

  4. Wilson's global land cover data

  5. Philip's great world atlas

  6. Atlas zur biogeographic.
Since the area of this study is the whole earth the plate Carree projection data was chosen. The advantage of this projection is the easiness of registration with other image data. The characteristics of data 1) and 2) are well described in NOAA POLAR ORBITER DATA USERS GUIDE (SDSD, 1986) and GLOBAL VEGETATION INDEX USERS GUIDE (SD SD 1986)_____ TEXT NOT CLEAR IN HARD COPY_____ Tateishi et al 1990. The data 4), 5) and 60 were used for checking the results of the cluster analysis.

Pre- processing
  1. Production of monthly GVI image

    In order to reduce remaining clouds and eliminate noises in weekly GVI images, a monthly GVI Image was produced from four or five weekly GVI images. If GVI images have no noise, the monthly data is picked up from the lowest SNVI (highest NVI) value of weekly data in the month. But almost all of the weekly data have many noises and clouds. The reduction of those noises and clouds is as follows. If channel 1 or channel 2 data has extremely high value in a certain weekly data that week data is regarded as a noise pixel and candidate of monthly data is searched from the other week data. But this is not the complete noise reduction method, so a little abnormal patches are still seen in the produced monthly images.

  2. Extraction of land area

    Since the sea and lake area data re not needed in this study, those areas were filled with 0 value in the monthly GVI images. In the discrimination of the land area, global DEM data was used. But the data showing land area needs a modification because lake above the sea level under the sea level area not discriminated from global DEM data.
Cluster analysis
  1. First Cluster analysis

    Since land covers with different vegetation types have different seasonal vegetation dynamics, land cover can be classified using 12 monthly GVI data. As the NVI fluctuations of the north and south hemisphere have normally reverse pattern and the amount of north hemisphere data are much larger than that of south hemisphere, cluster analysis applied together to the north and south hemisphere. Applied algorithms of cluster analysis is K-means method with merger of clusters. On both hemisphere, 31 clusters were obtained from 40 initial clusters.

  2. Secondary cluster analysis

    Cluster analysis was applied again to both hemispheres with the data shift of 6 months for south hemisphere data. In this cluster analysis, 62 cancroids of resultant clusters of the first cluster analysis were used as the initial clusters. 44 clusters were obtained by the second cluster analysis.

  3. Manual merging

    44 Cluster were merged annually into 13-land cover groups with considering of the annual mean NVI value, fluctuation pattern and geographical distribution of the each cluster.
Results and Conclusion
Global land cover map was produced by the cluster analysis of the monthly NOAA GVI data (fig. 1). The graphs of monthly mean NVI value of the land cover groups are also drawn in Fig. 2 Table 1 shows brief description of the land cover type of the each group. From the visual interpretation of the map, the distribution of the each land cover group correspond well to the vegetation map by Smithusen (1976), especially in equatorial area and in the south America, Africa and Australia continent.

But the land cover groups in a high latitude area (>600) of north hemisphere do not well correspond to that vegetation map. One o that reason is the solar zenith angel effect (Sing 1988). In a high latitude are, even in summer, NOAA AVHRR data is often recorded at the solar zenith angle (SA) up to 700 and in winter the data is recorded at SZA up to 900. According to Singh (1988), as the SZA beyond 300, the NVi decreases gradually. So in the high latitude area, the NVI does not correspond to the vegetation activity.

Another reason is cloud contamination (Holben 1986). As mentioned in the earlier section, monthly GVI images are not completely cloud free images. Visual checking of weekly GVI images shows that high latitude area s widely covered with clouds almost every week especially in winter. So in winter, the NVI value of this area may not be from vegetation but cloud. In this situation, vegetation is not active its NVI values are low. While clouds have generally low values of NVI. So it is difficult to tell which is the true source of the NVI.

For further study, the characterization of cash cluster and land cover group will be achieved, and a supervised classification will be done by the same data.

Table 1 Brief description of merged 13 land cover groups
Gr. Cover type Gr. Cover type
1 Tropical rain forest 8 Mediterranean scrub
2 Savanna 9 Evergreen needleleaved forest
3 Cold deciduous forest with ever greens 10 Cold deciduous wood land
4 Cold deciduous forest with evergreens 11 Scrub, steppe and semi-desert
5 Monsoon forest 12 Tundra and Ice
6 Savanna, grassland 13 Desert
7 Grass land    



Figure 1 (a). Global land cover map with 13 land cover groups in North and South America.


Figure 1 (b). Global land cover map with 13 land cover groups in Africa, Eurasia and Oceania.


Figure 2. Annual mean NVI fluctuation of 13 land cover groups, solid lines and dotted lines indicate those in north and south hemisphere respectively.

Acknowledgements
The authors are grateful to Mr. Hiroshi Okumura of Remtoe Sensing and Image Research Center, Chiba Univ. for his software development of K-means clustering and also wish to thank Mr. Hikaru Kuki of Remote Sensing and Image Research Center, Chiba Univ. for his assistance of data processing.

References
  1. Holben, B.N 1986, "Characteristics of maximum value composite images from temporal AVHRR data", Int. J. Remote Sensing, Vol. 7, No. 11, 1417-1434.

  2. Justice, C.O. Townshend, J.R.G, Holben, B.N. and Tucker, C.J. 1985, "Analysis of the phenology of global vegetation using meteorological satellite data", Int. J. Remote Sensing, Vol6, no. 8, 1271-1318.

  3. Kajiwara,K. and Taeishi, R., 1990, "Integration of satellite data and geographic data for global land cover analysis", proceedings of ISPRS Comm., IV Sym. Tsukuba, Japan.

  4. Schmithusen, J., 1976"ATLAS ZUR BIOGEOGRPHIE", Bibliographisches Institute, Mannheim F.R.G.(In German).

  5. SDSD, 1986, "NOAA POLAR ORBITER DATA USERS' GUIDE", Satellite Data Services Division (SDSD), National Climatic Data Center, NESDIS, NOAA, U.S. Department of Commerce, Washington D.C, U.S.A.

  6. SDSD, 1986, GLOBAL VEGETATION INDEX USERS' GUIDE", Satellite Data Service Division (SDSD), National Climatic Data Center, NESDIS, NOAA, U.S Department of commerce.

  7. Singh, S.M, 1988, "Simulation of solar Zenith angle effect on global vegetation index (GVI) data", Int. J. Remote Sensing,Vol. 9, No. 2 237-248.

  8. Tateishi, R. and Kajiwara, K., 1990, "Global land cover monitoring by NOAA GVI data", proceedings OF ISPRS Comm. VII Sysm. Victoria, Canada.

  9. Willet, B.M. et. al. 1987, "Philips great world Atals", George Philip and Son Ltd., U.K.

  10. Wilson, M.F. and Henderson - Sellers, A., 1985, "a global archive of land cover and soils data for use in general circulation climatic models", J. Climatology, vol. 5, 119-143.