Mining Satellite Image Time Series: Statistical Modeling and Evolution Analysis

Due to the short revisit time of high resolution satellites, huge amount of high resolution satellite images can be acquired every few days even few hours. It promotes the construction of Satellite Image Time Series (SITS), which contain valuable spatio-temporal information. Therefore, it is strongly needed to develop methods to explore such huge data to provide useful information in the context of earth observation. To address this issue, a patch based method for mining satellite image time series is proposed, consisting of statistical modeling and evolution analysis. Many statistical models has been proposed for Synthetic Aperture Radar (SAR) image modeling. Among them, G distribution has been proved efficient in modeling extremely heterogenous area especially for urban areas. In this paper, it is used to estimate the marginal distribution of SAR images by second-kind statistics. For the purpose of joint distribution modeling given the marginal distributions, optimal copula function is selected from a set of copulas by a Bayesian method and estimated using Kendall's τ. Based on the statistical model and the optimal copula, mixed information is computed between two neighboring patches along time for evolution analysis of the SITS. A v-support vector machine is applied for evolution classification. Performance of both estimation and classification are evaluated using our database produced by iterative classification.