Visually lossless compression of multilook SAR images

Synthetic aperture radars (SARs) produce a large amount of remote sensing data useful for numerous applications. SAR image resolution improves, leading to an increased size of acquired images that must be transferred to on-land data processing centers or directly to customers. Then, SAR data compression is required. Nowadays, lossy compression is mostly applied, but it is necessary to control losses to avoid undesired (inappropriate) deterioration of useful information contained in SAR images. In this study, we consider visually lossless compression of multilook SAR images using several lossy compression techniques (including modern coders such as BPG, AVIF, and HEIF) and both conventional and visual quality metrics (including PSNR-HVS-M, HaarPSI, and MS-SSIM). Such metrics and the corresponding distortion invisibility thresholds are employed to achieve the maximum possible compression ratio. We show that, in general, the attained compression ratios are approximately 3 if calculated for 8-bit representation of the images to be compressed. Depending on the visual quality metric used, different coders might produce the largest compression ratio. The considered compression techniques are directly applied to normalized SAR without preliminary variance stabilizing transforms. The images used in the tests have the same speckle characteristics as the real-life images acquired by Sentinel-1 synthetic aperture radar operating in multilook mode, i.e., the speckle is simulated as quasi-Gaussian multiplicative spatially correlated noise. Examples of original and compressed images are presented, demonstrating their very high similarity and practical invisibility of distortions introduced by lossy compression. The obtained results are discussed, and further research directions are proposed. In particular, the use of variance-stabilizing transforms must be considered.