Enhancing forest cover analysis through super-resolution of Sentinel-2 multispectral images

Machine learning (ML) algorithms, combined with satellite observations, offer significant advantages in environmental studies, particularly in vegetation cover analysis. The varying spectral resolution and number of spectral bands of remote sensing imagery allow for different tasks to be addressed with different levels of detail and accuracy. A current limitation in advanced Geographic Information System (GIS) development is the availability and accessibility of data. High-resolution data with a wide spectral range are often expensive, while open-access data typically force researchers to choose between high spatial and temporal resolution or large number of spectral bands. In this study, we investigate this issue through a case study of forest type classification. We employed and trained a single-image super-resolution model based on the Residual Channel Attention Network (RCAN) to upscale Sentinel-2 multispectral images from 10 to 5 meters. We then compared image segmentation results from the original Sentinel-2 data, the upscaled data, and WorldView-3 images. In addition to experiments with spatial resolution, we explored the effect of number of spectral bands on segmentation quality. The results confirm our hypothesis that artificially upscaled data provide more information than low-resolution data, both for narrow and wider spectral ranges, with the increase in spatial resolution proving more significant than the increase in number of spectral bands.