Exploiting Multi-Temporal Information for Improved Speckle Reduction of Sentinel-1 SAR Images by Deep Learning

Abstract

Deep learning approaches show unprecedented results for speckle reduction in SAR amplitude images. The wide availability of multi-temporal stacks of SAR images can improve even further the quality of denoising. In this paper, we propose a flexible yet efficient way to integrate temporal information into a deep neural network for speckle suppression. Archives provide access to long time-series of SAR images, from which multi-temporal averages can be computed with virtually no remaining speckle fluctuations. The proposed method combines this multi-temporal average and the image at a given date in the form of a ratio image and uses a stateof-the-art neural network to remove the speckle in this ratio image. This simple strategy is shown to offer a noticeable improvement compared to filtering the original image without knowledge of the multi-temporal average.

Multi-temporal speckle reduction with self-supervised deep neural networks

“…If numerous images are available, training a network to process all images becomes heavy, in particular regarding memory issues. Other approaches like ratio-based filtering [20] or a different strategy to combine images from the stack may then be preferable.…”

“…Yet, in order to perform a quantitative assessment of multi-temporal filtering, we first consider a simulated speckle framework in which both speckle-free and speckle-corrupted images are available. We build high-quality speckle-free stacks by multi-temporal filtering with RABASAR-SAR2SAR [20]. We then generate corrupted versions with simulated speckle corresponding to an ideal SAR transfer function, i.e., speckle with no spatial correlation in the simulated images.…”

“…We repeat the previous experiment with simulated speckle, this time introducing temporal correlations with a coherence matrix Γ k identical for all pixels k, and following a simple temporal decorrelation model The speckle is simulated based on the method described in [20], and details of the test set are given in III-A1. where τ is a characteristic decorrelation time.…”

“…The final despeckled images are obtained after multiplication by the super-image. To despeckle the ratio image, a deep neural network such as SAR2SAR can be used [20]. A drawback of ratio-based processing is that the lowest-contrasted structures present either in the speckled image or in the super-image might be improperly restored, leading to the suppression of these details or the apparition of a "ghost" structure leaking from the super-image.…”

“…Figure 4. Squared bias and variance averaged over 100 Multi-temporal MERLIN estimations of the reflectivities of a Sentinel-1 stack of Limagne (France).The speckle is simulated based on the method described in[20], and details of the test set are given in III-A1.…”