Remote Sensing Image Fusion based on Improved Super-Resolution Convolutional Neural Network

By fusing multi-source remote sensing images, higher spatial resolution and richer detail information can be obtained to better serve the fields of environmental monitoring, crop estimation, and urban planning. In order to effectively improve the quality of low-resolution multispectral remote sensing images, this work proposes a remote sensing image fusion method based on improved super-resolution convolutional neural network. Firstly, the characteristics of super-resolution technique and convolutional neural network are investigated, and a novel three-layer convolutional neural network, SRCNN, is introduced. Then, the multispectral image is divided into four different channels for processing, and all the four different channel images are fed into the SRCNN for the enhancement of high-frequency detail information. The predicted multispectral and panchromatic images are sparsely represented before fusion. Secondly, the weights of SRCNN are generally initialised using two methods, namely Gaussian distribution as well as encoder assignment. However, these two algorithms have uncertainties that affect the reconstruction accuracy of the images. Therefore, the PSO algorithm is used to optimise the SRCNN weights, thus improving the resolution reconstruction accuracy. Finally, multiple sets of images from different areas of Landsat satellite data are used for simulation by both subjective and objective evaluation metrics. The experimental results show that the proposed methods all better maintain the rich information of remote sensing images and achieve better fusion results. The indicators such as source entropy, correlation coefficient, average absolute error and mean square error of the fused images are improved after the introduction of PSO algorithm. © 2024, J. Network Intell. All rights reserved.