Super-Resolution of Multispectral Multiresolution Images From a Single Sensor

Charis Lanaras, Jose Bioucas-Dias, Emmanuel Baltsavias, Konrad Schindler; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 20-28

Abstract


Some remote sensing sensors, acquire multispectral images of different spatial resolutions in variable spectral ranges (e.g. Sentinel-2, MODIS). The aim of this research is to infer all the spectral bands, of multiresolution sensors, in the highest available resolution of the sensor. We formulate this problem as a minimisation of a convex objective function with an adaptive (edge-reserving) regulariser. The data-fitting term accounts for individual blur and downsampling per band, while the regulariser "learns" the discontinuities from the higher resolution bands and transfers them to other bands. We also observed that the data can be represented in a lower-dimensional subspace, reducing the dimensionality of the problem and significantly improving its conditioning. In a series of experiments with simulated data, we obtain results that outperform state-of-the-art, while showing competitive qualitative results on real Sentinel-2 data.

Related Material


[pdf]
[bibtex]
@InProceedings{Lanaras_2017_CVPR_Workshops,
author = {Lanaras, Charis and Bioucas-Dias, Jose and Baltsavias, Emmanuel and Schindler, Konrad},
title = {Super-Resolution of Multispectral Multiresolution Images From a Single Sensor},
booktitle = {Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops},
month = {July},
year = {2017}
}