Automatic Classification of Time Series (ACTS): a new clustering method for remote sensing time series
Automatic classification methods have often been used as a first step for land cover mapping. The principle of such methods is to determine clusters of pixels with similar radiometric temporal behaviour based on their statistical properties. This allows a segmentation of the image into regions with similar radiometric properties. Most automatic classification of remote sensing data are based on the K-mean or dynamical clustering method. The latter method has two limitations. (i) It is necessary to fix the number of clusters, but this parameter is, in general, unknown. (ii) It is very slow and does not work correctly when the dimension of the problem and the number of samples become large which is typically the case for classification of remote sensing data at large scale.
To avoid these limitations we have developed a new method called ACTS (Automatic Classification of Time Series), which is based on both hierarchical and dynamical clustering principles. First of all, the method is really 'automatic' since it determines, automatically, the number of clusters. Secondly, the method is very fast and does not show a degradation of the results with large dimensions or data sets. Application to synthetic data sets shows that in most cases ACTS is able to retrieve all the clusters of the image independently of the dimensions of the problem. Comparison of classifications based on actual global 8 km NDVI (Normalized Differential Vegetation Index) composites using both ACTS and a K-mean method show very similar results but the convergence for ACTS is 20 times faster than the K-mean method using '10-day' composites.
The ability of ACTS to work with problems of large dimensions enables clustering of multi-year time series of NDVI. ACTS is here applied to the clustering of a 12-year time series of 10-day composites (1982-1993). The results show that the seasonal signal is dominant. The clusters are mainly representative of seasonal land cover regions. Moreover, the regions are more clearly delineated in comparison with the classification based on only one year of data. Such improved clustering can help avoid some confusion between biomes. Finally, ACTS is applied to 'deseasonalized' time series to investigate the interannual variability of the NDVI. The areas of higher variability are located in the tropical regions with a strong influence of El Nino events. A small positive trend in NDVI is visible in high latitudes. However, several problems linked to the quality of the data are clearly visible. For instance, the decrease of NDVI following the Pinatubo eruption in 1991 and the drift in calibration with the change from NOAA 9 to NOAA 11 in 1988 are visible in most of the regions. This unfortunately limits the possible interpretation of the signal and emphasizes the need to improve the preprocessing of AVHRR (Advanced Very High Resolution Radiometer) data for interannual variability applications.
No Reference information available - sign in for access.
No Citation information available - sign in for access.
No Supplementary Data.