Unsupervised Land Cover Change Estimation Using Region Covariance Estimates

In this letter, we demonstrate the utility of estimating a probabilistic model of the underlying seasonal and interannual variations experienced by land cover time series in a given geographical region. Time series that deviate from these trajectories due to the human-induced change appear as outliers and can be detected using their Mahalanobis distance from the mean under the joint distribution of time samples. We apply this model to a collection of pixel time series acquired by the Moderate Resolution Imaging Spectroradiometer platform over Limpopo province, South Africa, for the task of identifying human settlement expansion. For estimation of the time of change, we present a hypothesis testing approach that tests for a decrease in correlation between samples before and after the change. This was found to be highly effective, yielding a mean absolute error of 52 days.