Ordóñez Iglesias, ÁlvaroBlanco Heras, DoraArgüello Pedreira, FranciscoDemir, Begüm2021-09-302021-09-302020The Journal of Supercomputing (2020)76:9478–9492. DOI 10.1007/s11227-020-03214-00920-8542http://hdl.handle.net/10347/26958This is a post-peer-review, pre-copyedit version of an article published in The Journal of Supercomputing. The final authenticated version is available online at: https://doi.org/10.1007/s11227-020-03214-0Image registration is a common task in remote sensing, consisting in aligning different images of the same scene. It is a computationally expensive process, especially if high precision is required, the resolution is high, or consist of a large number of bands, as is the case of the hyperspectral images. HSIKAZEisaregistration method specially adapted for hyperspectral images that is based on feature detection and takes profit of the spatial and the spectral information available in those images. In this paper, an implementation of the HSI–KAZE registration algorithm on GPUs using CUDA is proposed. It detects keypoints based on non–linear diffusion filtering and is suitable for on–board processing of high resolution hyperspectral images. The algorithm includes a band selection method based on the entropy, construction of a scale-space through of non-linear filtering, keypoint detection with position refinement, and keypoint descriptors with spatial and spectral parts. Several techniques have been applied to obtain optimum performance on the GPUeng© Springer Science+Business Media, LLC, part of Springer Nature 2020Hyperspectral dataImage registrationKAZE featuresRemote sensingCUDAGPUjournal article10.1007/s11227-020-03214-01573-0484open access