Time series change detection using segmentation: A case study for land cover monitoring

Segmentation of a time series attempts to divide it into homogeneous subsequences, such that each of these segments are different from each other. A typical segmentation framework involves selecting a model that is used to represent the segment. In this paper, we investigate segmentation scores based on difference between models and propose two approaches for normalizing the difference based score. The first approach uses permutation testing to assign a p-value to model difference. The second approach builds on bootstrapping methodology used in statistics which estimates the null distribution of complex statistics whose standard errors are not analytically derivable by generating alternative versions of the data by a resampling strategy. More specifically, given a time series with either a single or two segments, we propose a method to estimate the distribution of model difference statistic for each segment. The proposed approach allows normalizing model difference statistic when complex models are being used in the segmentation algorithm. We study the strengths and weaknesses of the two normalizing approaches in the context of characteristics of land cover data such as seasonality and noise using synthetic and real data sets. We show that relative performance of normalization approaches can vary significantly depending on the characteristics of the data. We illustrate the utility of these approaches for detection of deforestation in Mato Grosso (Brazil).