Application of Artificial Intelligence Techniques to Improve Sentinel-3
The Sentinel-2 mission satellites provide multispectral images with 13
spectral bands at three different spatial resolutions (10, 20 and 60m
Ground Sample Distance - GSD). In contrast, the Sentinel-3 mission
products have 21 spectral bands at a minimum spatial resolution of 300m.
Therefore, the article’s objective is to combine the data of two
satellite missions to improve the spatial resolution of the latter. We
use a convolutional neural network (CNN), which has already been proven
to improve the resolution of Sentinel-2 bands from 20 and 60m GSD to
10m, and a generative adversarial network (GAN), both of which are
trained with data from different latitudes and terrains at lower
resolution, i.e., from 9km to 300m, to predict the step from 300m to
10m. The results of both neural networks are compared with those of the
traditional pansharpening and bicubic interpolation super-resolution
algorithms. Thus, it shows that the newly proposed methods improve the
previous ones both through quantitative analysis and visual comparison.
In particular, the outstanding performance of the GAN used is
remarkable, which manages to improve the global numerical results of
traditional algorithms by around 30%.