Band reconstruction using a modified UNet for Sentinel-2 images

Multispectral (MS) remote sensing (RS) images are of great interest for various applications, yet, quite often, a MS product exhibits one or more noisy bands, strip lines even missing bands which leads to decreased confidence in the information it contains. Meeting this challenge, current paper proposes UNet based neural network architecture reconstruct spectral band. The worst-case scenario is considered, that band, reconstruction being performed on available bands. Besides comparison with state art methods, both qualitative and quantitative analysis full-filled considering several metrics: Root-mean-square error (RMSE), Structural similarity index (SSIM), Signal (SRE), Peak signal noise ratio (PSNR) Spectral angle mapper (SAM). experiments focused Sentinel-2 (S2) open data within Copernicus programme. Various patterns urban areas, agricultural regions, regions from North Pole Kyiv, Ukraine included our dataset prove efficiency band regardless landcover diversity.