Multispectral imaging sensors collect information about the reflectance characteristics of land-cover materials over multiple wavelengths of the electromagnetic spectrum. The sensor can be designed to have low or high spatial resolutions in different spectral bands, which significantly affects the reliability of classification accuracy results. This study investigates the effect of band upsampling, which increases low spatial resolution of a spectral band to high resolution, on the classification of Sentinel-2 multispectral data. Both traditional Bicubic Interpolation and deep learning-based DSen2 algorithms are utilized to increase the spatial resolution of spectral bands having 20m and 60m to 10m resolution. Then, random forest and support vector machines classifiers are used to classify the different band combinations. Experiments are realized on two real Sentinel-2 image datasets called SeasoNet covering different seasons such as fall, spring and summer. The quantitative results demonstrate that the highest classification accuracy values are achieved when the band combinations include the upsampled 20 m and 60 m bands, in addition to the original 10 m resolution bands. This improvement is observed consistently across all seasons when Bicubic Interpolation and DSen2 methods are utilized for upsampling. The qualitative results, on the other hand, demonstrate that upsampling with DSen2 yields visually superior images compared to the Bicubic Interpolation.
Tuğcan Dündar (Tue,) studied this question.