Change Detection from Satellite Data Time Series Using Pixel Value Distributions
Saarela, Olli; Molinier, Matthieu
VTT Technical Research Centre of Finland, FINLAND

The paper proposes a novel approach for change detection from image time series. In this approach changes are detected from evaluated distances between the (possibly multivariate) distributions of pixel values. Basing change detection on these distributions facilitates, e.g., joint analysis of images having different resolutions and comparisons of smaller areas against larger images. Furthermore, clouded areas can be excluded from each image separately, allowing the data in the remaining pixels to be utilized independent on the whether the corresponding pixels have been covered by clouds in the other images in the time series. In the paper the proposed method is applied to forest cover change detection using Landsat data covering Mexico.

In the proposed method the distances between pixel value distributions are evaluated using the Kolmogorov-Smirnov test statistic. It is computed as the maximum difference between two cumulative distributions, D=max(|F(x)-G(x)|). Using cumulative distributions avoids having to specify a bin size for histograms. Other advantages of this test statistic are its relative insensitivity to individual outlier values and that its statistical significance can be evaluated without having to assume a specific shape of the distributions being compared.

Rapid changes are detected when the distance exceeds a threshold value as in the statistical Kolmogorov-Smirnov test, but accounting for the number of comparisons carried out so that desirable detection sensitivity is reached. Gradual longer-term drifts are detected by computing distance D from the most recent image to each previous one, and examining the resulting distance time series with the statistical Mann-Kendall test. This test evaluates the degree of monotonicity without assuming a particular shape of the drift (e.g., linearity with respect to time).

The application-specific parametrization of the method involves simply specifying detection thresholds and an area for evaluating the distributions. The latter of these balances spatial resolution with detection sensitivity. In this choice also the number of extracted features needs to be considered, and whether the joint distribution or the set of marginal densities are considered.

This work has been carried out in the ReCover project in FP7, funded by EU.