Knowledge-Based Multispectral Remote Sensing Imagery Superresolution

Software techniques for remotely sensed imagery superresolution enhance data reliability and veracity. The most common approach for superresolution is the processing of few images of the same scene captured simultaneously with subpixel shift relatively to each other. These conditions exclude radiometric inconsistency between images, and subpixel shift allow the extracting of additional land surface details. A general superresolution approach can be adopted to multispectral remote sensing imagery registered in different spectral bands. In this case, the intrinsic radiometric inconsistency can be overpassed by translating of the input bands into some additional virtual one, joint for all inputs. Typically, such an additional band overlaps all input ones in the spectrum. Necessary knowledge for bands translation are all bands spectral responses, as well as the subpixel shifts between restored images. So, the spectral radiance for a new spectral band is estimated. Therefore, each input band image transforms into new image in the same spectral range. Obtained images are appropriate for any existing superresolution techniques, for example, using Gaussian regularization in the frequency domain. The last step of the proposed method is image improvement after superresolution using a convolutional artificial neural network.