Preprocessing
Original NDVI values in each pixel have noise due to the cloud effect or the sensor noise. First, noise of low frequency was removed with the Median Filter where each NDVI value is replaced by the median value of three successive NDVI values. Noise of high frequency was removed by depositing the 3-th, 5-th to 24-th Fourier power density spectrum calculated as shown in the previous section. Moreover, DEM data were taken into consideration to get rid of the influence of the shadow by the geographical features.

The periodicity of the NDVI value is extracted as 0-th to 2-th and 4-th Fourier power density spectrum. Vegetation distribution due to the height above sea level was also considered utilizing the DEM data

Classification
In this study the number of classified categories which can be distinguished by satellite data is supposed to be beyond 30,therefore training area is set automatically from actual vegetation data by random sampling. As phenology of vegetation is the key characteristics of land cover classification, 0-th, 1-th, 2-th and 4-th power spectrum, corresponding to averaged value of two years period , 12 months period component and 6 months period component are used. Classification is done with maximum likelihood classification. Figure 2. shows an example of land cover classification with Fourier spectrum.


Figure 2 An example of vegetation classification with Fourier spectrum.

Comparison of Classified Vegetation Distribution and “Green Census” Data
There are 766 categories in "Green Census" data which is used as the precise validation data. It is impossible to categorize all the vegetation classes in the validation data directly from the satellite data, therefore 766 categories are stratified according to phenology.

Original vegetation vector data whose precision is about 100m is converted into 50m mesh data. It is integrated to the 1km mesh data to be compared with the result of classification. As the representative vegetation for 1km mesh, the following three are considered; the vegetation that occupies the maximum area, the vegetation that contributes most to the change of NDVI and the rate of vegetation as the index. Figure 3 shows an example “Green Census” data for Osaka Prefecture. Percentage of the area of the classified categories corresponding to each actual vegetation categories is calculated.


Figure.3.” Green Census” data Osaka (1km mesh)

Conclusions
Vegetation classification based on the time series NDVI data from NOAA/AVHRR is investigated with the so-called “Green Census” data, and the feasibility of time series NDVI data for vegetation classification was evaluated.

There have been many methods for vegetation classification, however, a good quality vegetation map has not been derived yet from remotely sensed data. One major reason for it is that the good quality ground truth data is not available. In this study newly produced digital map of actual vegetation is used for validation of the classified results. Also it is used to organize reasonable vegetation classification system from remotely sensed data. Although the results is still at the preliminary stage it is expected to extend the developed method and a classification system for more extensive areas covering East Asian region.

Acknowledgments
Authors would like to express our thanks to the Japan Environment Agency for providing us with the “Green Census” data, and also thanks to National Institute for Environmental Studies for providing us with the NDVI data set .

References

• R.S.DeFies, et.al.: NDVI-derived land cover classification at a global scale,Int.J.Remote Sens,Vol.15,No.17,pp3675-3586,1994.
• M. Sugita and Y. Yasuoka : Land Cover Classification of East Asia Using Fourier Spectra of Monthly NOAA AVHRR NDVI Data,IGARSS(1996):