GPU computation of Attribute Profiles for RS Image Classification

Classification of multi and hyperspectral remote sensing images is a common task. It usually requires a previous step consisting of a technique for extracting the spatial information from the image being profiles a common approach. In particular, attribute profiles are based on the application of a morphological filter to the connected components of the image producing relevant spatial information at different levels of detail. The information is built based on attributes such as area or standard deviation. Their high computational cost makes the attribute profiles good candidates for their execution on commodity GPUs. In this paper, the first parallel implementation of attribute profiles over multispectral images in CUDA for Nvidia GPUs is proposed. The GPU proposal is based on the construction of a max-tree that is traversed from the leaves to the root by merging the connected components of the tree obtaining a considerable reduction in execution time over the CPU execution.

keywords: Remote sensing, Attribute profiles, Supervised classification, Real-time, GPU.