Spatiotemporal Analysis for NDVI Time Series Using Local Binary Pattern and Daubechies Wavelet Transform

Abstract

NDVI time series has shown to be very efficient for vegetation change dynamic analysis over a long period. However, noise and illumination variations present significant challenges to perform an accurate change detection. This paper aims at capturing global vegetation change dynamics within 16-days MODIS-NDVI time series by considering the inter-annual variations. To determine the appropriate scale that characterizes the long term variation, an efficient way relying on wavelet transform is used. First, the Daubechies 4 wavelet transform is employed to perform a multi-scale decomposition to extract the inter-annual variations and remove noise. Second, critical point theory is used to identify a set of points indicating potential vegetation change within time series, which, allows a time series reduction. Then, for each critical point, LBP code is computed to characterize the corresponding local patterns, which provides the ability to deal with illumination variations. Based on the extracted features, a change map is produced by computing similarity between neighboring time series, assessing dynamic vegetation change over the period of study. Experiment results using NDVI time series show clearly the potential of the proposed approach to detect change.