Abstract:

Single image super-resolution （SISR） can improve the resolution of remote sensing images （RSIs）， thereby improving the application value of data. At present， the number of pixels of RSIs generally reaches hundreds of millions， and it is usually necessary to divide the image into patches when performing SISR. However， there is a lack of relevant research on how to effectively determine the patch size and whether different sizes affect the results. In this paper， taking a large-scale high-resolution RSIs as the experiment data， 3 typical SISR models are selected， 9 groups of SR experiments under different patch sizes are carried out， and the super-resolution （SR） results for the whole of the large-scaled RSI are analyzed comprehensively both qualitatively and quantitatively. The results show that： 1） Cutting of the patches results in stitching seams at the stitching place. In particular， when the patch size is small， a large number of stitching seams show a block effect and the inconsistency is more obvious. 2） With the increase of the patch size， the SR accuracy of the three models is improved， and the overall computational efficiency is also improved. When the test patches are larger than the training patches， the elapsed time and accuracy stabilize. 3） The feasibility and accuracy of the whole RSI input are closely related to the model. The ESPCN model has the best accuracy when inputting the whole RSI， the RDBPN model may cause the accuracy to decrease due to the non-square matrix of the RSI， and the HSENET model has high requirements for computing power and cannot calculate the whole RSI. In conclusion， this paper provides an experimental basis for the selection of patch size for RSI SR engineering applications.

Key words: deep learning, super-resolution, large scene remote sensing images, spatial resolution, input size, patch size