A STUDY OF TIME SERIES NOISE REDUCTION TECHNIQUES IN THE CONTEXT OF LAND COVER CHANGE DETECTION

The purpose of this study is to introduce concepts relevant to performance of (i) change detection algorithms within (ii) various regional contexts with differing noise characteristics according to (iii) differing strategies of noise reduction. The relevant interrelations of these three elements are presented, and focused analysis is presented from the perspective of varying (i) and (iii) for a comparative analysis across (ii). Six smoothing methods has been studied in this work: Savitzky-Golay (SG) method [7], The Savitzky-Golay method iterated to upper envelope (SG-Itr) [3], Harmonic Analysis of Time Series (HANTS) [6], Double Logistic function fitting method (DL) [1], Data Assimilation method(DA) [5]and a naive outlier identification and imputation scheme (SO). In this work, we enumerate three general data characteristics, especially relevant in the MODIS EVI data, which a given noise reduction technique may take advantage of: neighborhood coherence, quality annotation and background model. For a noise reduction technique we identify the following two questions to be of relevance: • Which observations in the time series should be imputed? • How are these observations to be imputed?