Information technology of satellite imagery pre-processing using convolutional neural network

Abstract

With urbanization and the growing interest in geographic information systems, preci-sion mapping, and environmental monitoring, there is a need to improve methods of fusing high-resolution digital satellite images to obtain more detailed and informative data. Satellite images with high spatial resolution hold significant promise for addressing a variety of is-sues. However their utility is frequently constrained by limited contrast dynamics and re-stricted spectral data. Pansharpening methods are effective means of increasing the spatial resolution of multichannel images using panchromatic ones. Although there are a significant number of proposed pansharpening algorithms, none can be universal for all types of sensors and demonstrate different results. The paper reviews the fusion methods of high spatial reso-lution satellite images and proposes an information technology for preprocessing digital sat-ellite images using a convolutional neural network. It made it possible to effectively extract texture details and spatial dependencies for local and non-local information. The paper com-pares the results of the proposed technology with existing pan-sharpening methods, such as Gram-Shmidt, HSV, Wav+HSV, ATWT, AWLP, BDSD, Brovey, and HPF. The results of test-ing conducted on eight-channel images acquired by the WorldView-2 satellite confirm that the proposed technology (ERGAS=0.32, RMSE=0.28b SSIM=0.71, PSNR=37.60) allows for achieving high spectral and spatial quality of multichannel images and outperforms existing methods.