Keywords: Time series NDVI, NOAA/AVHRR, Vegetation classification

Abstract
Vegetation classification based on a time series NDVI data from NOAA/AVHRR is investigated with the so-called “Green Census” data to evaluate the feasibility of temporal analysis of satellite data for vegetation classification. A time series of NDVI data has been popularly used for vegetation type classification since it reflects the phenological features of vegetation. It is, however, pointed out that it is difficult to validate the classified result because of the lack of actual vegetation data. Also the sensitivity of the NDVI data to vegetation classification has not been fully investigated. In this paper, vegetation classification based on a time series NDVI patterns from NOAA/AVHRR is compared with the actual vegetation data (“ Green Census” data), and the feasibility of vegetation classification from time series NDVI is investigated.

Introduction
Global environment issues such as desertification, acid deposition or deforestation have been critical all over the world, and the necessity for monitoring land cover changes is pointed out. Among a wide variety of land cover parameters, vegetation distribution is especially important for assessing the impact of human activities on the ecosystem. Therefore monitoring of vegetation cover conditions and their changes is essential for the management of environment in both of the regional and global scale. It is, however, difficult to carry out land cover survey over extensive areas only with the ground survey. Remote sensing from satellite is expected to be one of the efficient methods for observing vegetation cover conditions.

There have been many methods to analyze satellite data for land cover classification. They are categorized in two approaches including a spectral analysis approach and a temporal analysis approach. In the first approach, usually, multi-spectral data with high spatial resolution such as LANDSAT TM is used for classification based on the spectral characteristics of the images. In the second approach, on the other hand, time series data is used for classification based on the temporal characteristics of the images. Satellite data used in the latter approach has usually low spatial resolution, however, it can observe wider areas more frequently, and it enables us to utilize temporal characteristics of the target categories for classification. One of the typical examples of this approach is a time series of NOAA/AVHRR data. In this study, first, the feasibility of applying Fourier analysis of a time series NDVI data from NOAA/AVHRR ^{(1), (2)} to vegetation classification was evaluated.

Two common problems, however, are pointed out in the conventional vegetation classification in methods. First, classification results can not be fully validated because of the lack of good ground truth data. Secondly, the suitable classification system for vegetation categories from remote sensing has not been established yet. Here, the effectiveness of the classification based on the phenology of vegetation is validated with so called “Green Census” data that is actual vegetation classification data set with the scale of 1/50,000 covering whole areas of Japan archipelago. Also appropriateness of vegetation classification system from remote sensing is investigated.

Time series NDVI data and “Green Census” data
In this study a time series NDVI data derived from NOAA/AVHRR was used to detect phenological features of vegetation and the classified results were compared with precise actual vegetation data set.

Monthly NDVI Data
A time series NDVI images was produced from monthly composite NOAA/AVHRR data set from 1996 and 1997 by National Institute for Environmental Studies. The images cover most of the East Asian region with a spatial resolution of 1.1km. Geometric correction was performed with the ground control points (GCPs).

"Green Census" Data
The "Green Census" data has been produced by Japan Environment Agency around once five years since 1973, and it includes actual vegetation maps with a scale of 1/50,000, and their digital format data (vector data). It covers whole Japan archipelago prefecture by prefecture, and the number of vegetation categories in the actual vegetation map is as large as 766. Here, “Green Census” data is registered with a time series NDVI data and compared together to evaluate the accuracy and the sensitivity of vegetation classification.

Vegetation Classification with Fourier Transform Analysis

Fourier Spectrum of NDVI Time Series Data
Monthly NDVI time series data V(t) at a certain NDVI image coordinate of (x, y) is a function of time t ,which is in discrete unit of month and has a range from 1 to 24 (2 years). With Fourier Transform. V(t) is written in a form of discrete Fourier series such as

where n-th imaginary Fourier coefficient f(n) is given by

The power density spectrum P(n) of V(t) is given by

Here, Fourier power density spectra was calculated for all the pixels in the NDVI images. Figure 1 shows typical Fourier power density spectra calculated for five different land cover categories. Each Fourier coefficient is used to remove the noise of the high frequency, and to extract periodicity at the same time.


Figure 1 Examples of Fourier power spectra of 24 months NDVI data.