Patch-based adaptive temporal filter and residual evaluation

TL;DR

The paper addresses speckle-induced degradation in SAR imagery by extending nonlocal despeckling to multi-temporal data. It introduces a patch-based adaptive temporal filter (PATF) that uses a generalized likelihood ratio test to quantify patch similarity across time and a temporal weighted average to denoise while suppressing changes, complemented by a Monte Carlo-derived thresholding scheme. A residual evaluation method based on patch-wise autocovariance of the ratio between noisy and denoised images provides an automatic, ground-truth-free quality score. Experiments on simulated and real Sentinel-1 data show PATF improves denoising and detail preservation, with residual analysis offering a global assessment tool, though methods like RABASAR may outperform PATF in certain real-data scenarios. The approach is practical for long time-series and can be readily deployed on platforms such as Google Earth Engine.

Abstract

In coherent imaging systems, speckle is a signal-dependent noise that visually strongly degrades images' appearance. A huge amount of SAR data has been acquired from different sensors with different wavelengths, resolutions, incidences and polarizations. We extend the nonlocal filtering strategy to the temporal domain and propose a patch-based adaptive temporal filter (PATF) to take advantage of well-registered multi-temporal SAR images. A patch-based generalised likelihood ratio test is processed to suppress the changed object effects on the multitemporal denoising results. Then, the similarities are transformed into corresponding weights with an exponential function. The denoised value is calculated with a temporal weighted average. Spatial adaptive denoising methods can improve the patch-based weighted temporal average image when the time series is limited. The spatial adaptive denoising step is optional when the time series is large enough. Without reference image, we propose using a patch-based auto-covariance residual evaluation method to examine the ratio image between the noisy and denoised images and look for possible remaining structural contents. It can process automatically and does not rely on a supervised selection of homogeneous regions. It also provides a global score for the whole image. Numerous results demonstrate the effectiveness of the proposed time series denoising method and the usefulness of the residual evaluation method.

Figures (6)

• Figure 1: Different time series and patch-based weight comparison. The self-similarity value of the reference point is set equal to the maximum similarity value with others. $h$ is set equal to 2 during the transformation.
• Figure 2: Realistic simulated SAR image one and reference image. The background image is obtained from an average of 69 Sentinel-1 images. The change types (the corresponding time series are shown in Fig.\ref{['fig:SyntheticCC']}) are: red: step change, green: impulse change, blue: cycle change and cyan: complex change. This simulated time series will be used for the change classification.
• Figure 3: Different time series changes introduced in figure \ref{['fig:SyntheticData1']}.
• Figure 4: Denoising performances comparison based on 64 simulated Sentinel-1 images. Left: denoising results, middle: ratio with noisy data, right: ratio with noise-free image. The max ratio values between the denoised data and the noise-free image are used.
• Figure 5: Different denoising results comparison based on 339 Sentinel-1 GRD images acquired over CentraleSupélec area: denoising results (left), its ratio with noisy data (middle) and residual evaluation results (right). Buildings are appearing in the red rectangle area. Google Earth Engine is used to prepare the time series data.