Multispectral satellite images contain rich spectral information but often have low spatial resolution. Super-resolution of these images is important for resource development, environmental protection, disaster prevention, regional ecology, and change detection. Deep learning models, including convolutional neural networks (CNNs), have been widely applied to the super-resolution of satellite images, including Sentinel-2 images. However, their output cannot exceed the maximum spatial resolution of the sensor. Here, we propose a method that first enhances the 20 m Sentinel-2 bands to 10 m using a CNN and then fuses them with 1 m aerial images by Gram-Schmidt orthogonal transformation and its inverse, while preserving multispectral information. The method was tested on the fusion of aerial and Sentinel-2 images to visualize vegetation distribution in urban areas. The proposed network achieved a root mean square error (RMSE) of 43.5 and a correlation coefficient (CC) of 0.941, compared with 123.4 and 0.811 for bicubic interpolation and 117.8 and 0.866 for area-to-point regression kriging (ATPRK). The fused images appeared as sharp as the aerial images while preserving the shape of the original spectral curves.
Copyrights © 2026