KP Labs Sp. z o.o. (PL)
The aim of the activity is to apply super-resolution reconstruction multispectral Sentinel-2 images, using multiple images of the same region, captured at different points in time (MISR, multiple-image super-resolution). This is achieved by adapting recent deep neural networks that were recently proposed for dealing with MISR, to the particularities of Sentinel-2 data. In particular the project focusses on three aspects: adapting the existing networks to process multispectral images, proposing techniques for preparing the training data, and selecting and pre-processing the input low-resolution data. The existing networks are applied to super-resolve the Sentinel-2 images in a band-wise manner (each band treated independently), followed by exploiting the correlation among the multiple bands. The output contains a panchromatic image, as well as an RGB/multispectral image of higher resolution than the one presented at the input.
Read about the project achievements in the following publications:
1. M. Kawulok, J. Nalepa, P. Benecki, D. Kostrzewa (2020): Deep learning for super-resolution reconstruction of Sentinel-2 images, Phi-Week 2020.
2. M. Kawulok, T. Tarasiewicz, J. Nalepa, D. Tyrna, D Kostrzewa (2021): Deep learning for multiple-image super-resolution of Sentinel-2 data, in Proc. IEEE IGARSS 2021, pp. 3885–3888.
3. J. Nalepa, K. Hrynczenko, and M. Kawulok (2021): Multiple-image super-resolution using deep learning and statistical features, in Proc. Joint IAPR International Workshops on Statistical Techniques in Pattern Recognition (SPR) and Structural and Syntactic Pattern Recognition (SSPR). Springer, 2021, pp. 261–271.
