Keywords: NOAA NDVI, temporal signature, remote sensing, vegetation mapping

Abstract
Vegetation mapping at regional and global level is very important for understanding the global environment. Remote Sensing, particularly, with the use of NOAA AVHRR data, makes such large area vegetation mapping possible. However, there are still complexities involved in making an optimum vegetation classification with NOAA AVHRR data. This study explores the possibility of using a new approach for vegetation classification from dynamic analysis of NOAA NDVI data. While conventional methods consist of classifying the image into some predefined number of classes, this method is based on the use of a new index for temporal signature similarity (TSS) calculated from patterns of temporal variability of NDVI values over different months. First, a large number of clusters are formed from small test areas over different parts of the image and then these are merged into a meaningful number of classes using the index for TSS so as to produce the vegetation map from the whole image.

Introduction
Large area vegetation mapping is very important for a better understanding of the regional and global environment. The availability of NOAA AVHRR data, especially the NDVI data is very useful for such mapping. It has been demonstrated by various authors (such as Murai and Honda 1991) that because of the frequent temporal resolution, a dynamic temporal analysis of NDVI data can be very effective in the analysis of vegetation.

Existing methods for classification are mostly based on either supervised classification or clustering of pixels into groups of similar values. These are basically designed for analyzing multispectral data where we analyze data from different bands (spectrum) of the data from a given sensor. However, analysis of NDVI data for different months requires a different approach. Here we have the single layer data product derived from the same sensor but consisting of series of data for different months of the year. Therefore, the classification of data should be based on dynamic analysis of the temporal patterns of data rather than the spectral properties in different bands, and this has to be addressed in the classification process. Sugita and Yasuoka (1997) have demonstrated one possible method of such classification based on Fourier spectra analysis of monthly NDVI data. Here we present a study for another method with the use of temporal signature similarity index.

Data and Method

Data Used
The monthly NOAA NDVI data for the east Asian region for each months of 1996 and 1997 obtained from National Institute of Environmental Studies were used for this study. It is also planned to use high-resolution satellite data and existing higher resolution land cover and vegetation data from various sources so as to validate the results of the analysis.

Test areas
In order to develop confidence on the classification, three test sites each 600x600 pixels have been selected in different parts of the image. These sites were so selected that all the classes are represented in these sites. The sites have been selected as around Thailand, east China and near Mongolia. Figure 1 shows the coverage of the study area and the three test sites.


Figure 1: Coverage of the study area and the test sites
Clustering of NDVI data
The 12 monthly NDVI data were clustered into 30 groups with simple isodata clustering. These were simple clusters merely based on the statistical distribution of data values. Ten clusters were then chosen from each group of 30 clusters so that even small areas with typical spectral behavior are not left out.

Development of TSS Index
Typical feature classes were selected based on the temporal patterns of NOAA NDVI and then a statistical analysis was performed so as to characterize various statistical properties of the classes and between the classes. New indicator was proposed based on these properties and then clusters were merged into smaller number of distinct classes based on the value for temporal signature similarity.

Final classification
Classification into various vegetation classes was then made based on the new TSS index and then class names were assigned to the different patterns of NDVI dynamics based on the available knowledge.