Fast statistically homogeneous pixels (SHP) selection
The i.i.d sample is essential for statistical inference (or homogeneous pixels from the viewpoint of image processing). Although the comparison of time-wise samples from SAR stack by two-sample non-parametric hypothesis test is a very powerful tool, the heavy computational burden limits the application in practice, especially for the wide scene coverage. To solve this issue, we developed FaSHPS algorithm under the parametric statistical framework. We deduced a likelihood ratio test and confidence interval respectively based on SAR intensity stack. Unlike non-parametric family that frequently compared the empirical distributions of samples, only logical operation is needed for FaSHPS as the critical region and interval are both analytic. Therefore, FaSHPS is a very fast algorithm and can be applied for SAR Big Data. FaSHPS is generally 200 times faster than non-parametric techniques, with smaller type II error.
Estimation theory for InSAR coherence
Coherence magnitude is a basic observable for InSAR data processing. The sample coherence magnitude, however, is biased towards the larger value over low coherence areas. It is clear that the unbiased coherence estimator is non-analytic and the bias correction is difficult in the real world. To solve the issue, we deduced a non-parametric double bootstrapping estimator for non-Gaussian scenarios, and a parametric bootstrapping estimator in Gaussian scenes, based on statistically homogeneous pixels from the same population. The case above demonstrates that bootstrapping is a pragmatic method to correct the coherence bias without loss of the resolution.
Distributed scatterers interferometry over low coherence scenarios
This research highlights the impact of coherence error on distributed scatterers interferometry over low coherence scenarios. We extended our previous work into coherence matrix to reduce the adverse effects of estimator at lower coherence values and therefore improve the estimate accuracy of optimal phase series. We further optimized the spatial network by maximizing the temporal coherence using shortest path algorithms. Thanks to the smaller bias and variance of sample coherence matrix, as well as the spatial network with the better quality, we found that 2%-40% accuracy improvement can be reached for both the time-series displacement and displacement rate over moderate to low coherence scenes.
Change detection in the built environment from SAR stack and multi-spectral imagery
We presented an MRF-based accumulative change detection method (MRF-ACDM) towards the detection of built-up land change by combining individual strong points of SAR and optical imagery. The proposed methodology includes the generation of a change indicator, the Markov modelling procedure and the delineation of changes over built-up areas. The generation of the change indicator aims to provide a feature with abundant contrast between changed and stable areas, a high signal-to-noise ratio and detail preservation. To this end, we for the first time investigated the use of coefficient of variation from SAR stack. After error removal, we used Markov random field (MRF) criterion function to delineate the boundary between changed and stable classes. Rather than MRF modelling by iteration with very complex stochastic models, we used SAR temporal trajectory under a hypothesis test framework and interferometric coherence series to establish conditional density for each class. We finally constrained the study extent to built-up areas using spectral information of optical datasets based on the Bayes theory. We found that in a complex built environment that is challenging for classical change indicators and state-of-the-art techniques, MRF-ACDM can provide smaller overall error with better detail preservation.
To be published