Efficient GPU-based local histogram analyzer for change detection in satellite SAR images

This paper presents an efficient GPU-based algorithm to perform histogram analysis of sub-images for change detection in SAR images. It is designed to use three-stage parallelisms: the SAR images are sub-divided and distributed to all the GPU devices; raster-scans for the sub-divided images are parallelized with many thread-groups inside each GPU; and 32 threads inside the group cooperatively compute the 32 statistical elements one by one by reusing the sub-histograms generated from the overlapped pixels until the dynamic range is modified. The analyzer with quad-GPU takes 1.7 s, which is 5.2 and 57.5 times faster than the conventional analyzer and that with 32-threaded dual-CPU, respectively, when applied to the problem of flood-water detection in ALOS-2 images with 37305x26811 pixels.