Speckle Removal and Change Preservation by Distance-Driven Anisotropic Diffusion of Synthetic Aperture Radar Temporal Stacks

Satellite imagery is often collected to perform remote sensing, or uncover changes in our Earth’s surface without having to traverse every corner. One application of using radar images is to observe changes in road topography without having to physically examine all of the roads. Unfortunately, radar images are often corrupted by speckle noise, which is a direct result of interference from scatterers at the imaging aperture. There are many speckle-reducing algorithms that work on either the temporal range of the image or the spatial aspect of the image. Traditionally, with synthetic aperture radar images, the mean of time series is utilized to produce a single despeckled image that discards temporal information. We propose a method that smoothes spatially and uses, but also preserves, temporal information. Radar images are often collected of the same region over time. To provide smoothing of such imagery without effacing temporal changes in the scene, we put forth an anisotropic diffusion technique, using a PDE approach. This approach smoothes uniform areas and preserves and enhances edges, such as roads or other features. In order to use this anisotropic diffusion technique, we must first select a homogeneous region in the image to calculate statistics of the speckle noise. This is often done manually, and can be time consuming and inaccurate with a large stack of radar images. We propose a method that uses the temporal information to automatically select a homogeneous region prior to smoothing. Our proposed smoothing method is a statistical approach designed to reduce speckle noise in each image throughout the time series. Results demonstrate the efficacy of the approach on real and synthetic data, showing lower mean squared error than leading methods. The new filter incorporates temporal information essential to detection of potential change events for transportation infrastructure. Change preservation is shown with our approach because temporal information is preserved, while signal-corrupting noise is reduced.