Benchmarking Framework for Multitemporal SAR Despeckling

In this article, we propose a novel benchmarking framework for a quantitative assessment of the performance of despeckling algorithms for multitemporal synthetic aperture radar (SAR) imagery. A number of canonical scenes and data sets are analyzed so as to investigate both speckle reduction and feature preservation capabilities of the filters. Despeckling performance is evaluated by proper quality measures that are defined according to the scene. Due to the lack of real-world speckle-free SAR images, the proposed benchmarking tool relies on an accurate and well-assessed SAR simulator which allows for generating realistic SAR images accounting for electromagnetic (EM) and geometrical parameters of the sensed surface. Accuracy and convergence properties of filters are first measured on scenes with stationary reflectivity. Then, for a more realistic performance prediction in practical situations, the effects of temporal changes of the scene reflectivity on the despeckled images are measured on time series with time-varying reflectivity. In the latter case, performance parameters are intended to measure the capability of the filter to preserve both the perturbation and its impact on the other bands. The whole benchmarking framework is applied to a representative set of state-of-the-art multitemporal filters. Interestingly, their performance as evaluated by means of our framework is well in agreement with (qualitative) visual inspections by SAR specialists. Proposed quality metrics are measured under the hypothesis of uncorrelated bands, which defines the best case for most multitemporal filters. A numerical sensitivity analysis of the performance of filters against correlation coefficient is carried out to investigate the temporal correlation effects on the despeckled time series.