VEGETATION CHANGE DETECTION IN LANDSAT TM TIME SERIES USING SINGULAR SPECTRUM ANALYSIS AND REGULAR FOREST INVENTORY DATA

References: 14th International Multidisciplinary Scientific GeoConference SGEM 2014, www.sgem.org, SGEM2014 Conference Proceedings, ISBN 978-619-7105-12-4 / ISSN 1314-2704, June 19-25, 2014, Book 2, Vol. 3, 397-404 pp

ABSTRACT
The lack of the fine temporal coverage of multispectral satellite data, the insufficient historical field data and impacts of unique conditions during satellite image acquisition are significant challenges for land cover change detection and characterization. The aim of this study is to evaluate the potential of Singular Spectrum Analysis (SSA) for detecting trends in vegetation change with small number of temporal data points available. Efficiency of SSA was tested for vegetation change trend classification in three groups: stands with no changes, stands with changes associated with forest growing and stands with abrupt changes. SSA trends were robust to variations of employed image features (NDVI, Tasseled cap Greenness and subpixel vegetation fractions) not related with actual land cover and characteristics of the SSA trends improved classification of forest stands (overall accuracy (OA) 94.67 %) comparing with the use of the same descriptors for unprocessed time vectors (OA 85 %). Date of the changes was estimated using unprocessed time vectors and for the 70% of the 100 test stands change data was estimated correctly as closest year of actual changes according to RFI.

Keywords: singular spectrum analysis, vegetation change detection, Landsat TM