Applying principal components analysis to image time series: effects on scene segmentation and spatial structure

Principal components analysis (PCA) has been applied to multitemporal data for over a decade, frequently in preparation for land cover classification. With the increasing availability of standardized multitemporal datasets, such as Pathfinder AVHRR Land products and the maximum biweekly AVHRR NDVI composites for the conterminous US, the authors wondered about the efficacy of using PCA to identify the dominant spatio-temporal modes within the data. How many principal components are useful? Does inclusion of higher order components improve scene segmentation or hinder it? What are the effects on the spatial structure of a segmented scene? To address these questions the authors selected a portion of the North American Great Plains from the Pathfinder AVHRR Land product for 1986. The subset region nominally covers over 2 million km² and spans significant temperature and precipitation gradients. The resulting image time series comprised 30 10-day maximum NDVI composites. These data were submitted to a series of PCAs with 3, 6, 9, or 12 principal components output as different images. Each PCA set was then submitted to an unsupervised classification, followed by maximum likelihood decision rule applied to the signature set to yield 4, 6, 8, or 10 classes. Each class was analyzed with lacunarity analysis, which quantifies spatial heterogeneity and anisotropy in binary data. There was a significant effect of the number of PCs used in the classification on the lacunarity of certain classes but not others. Visual inspection revealed that several higher order principal components were compositing artifacts. These results suggest that PCA performed on image time series can effectively filter compositing noise when data are reduced to a small number (<6) of components