Research On Spatio-temporal Fusion Algorithm Of Remote Sensing Images Based On Deep Learning

With the continuous improvement and development of modern remote sensing science and technology,the spatial,temporal,and spectral resolution of image data obtained by satellites has also been continuously improved,providing a large amount of data support for my country’s agriculture,waters and other fields.However,the current image data acquired by a single satellite sensor cannot take into account both high spatial resolution and high temporal resolution at the same time.Therefore,domestic and foreign scholars have proposed many remote sensing image spatiotemporal fusion methods to generate remote sensing images with high temporal resolution and high spatial resolution at the same time.Based on the theory of deep learning,this paper studies the spatio-temporal fusion of remote sensing images.The content is as follows:(1)In this paper,the current existing spatio-temporal fusion methods are classified according to their different principles,and 11 algorithms are selected from them.The two data sets of CIA and LGC are used for comparison and analysis of spatio-temporal image fusion algorithms under the same conditions.The quantitative evaluation of the fusion results shows that the combined method fusion results have less spectral loss and more spatial details,and the near-infrared band has fewer relative radiation errors than the red,green and blue bands,which has more guiding significance for practice;The visual evaluation results show that when the study area changes greatly,the reconstruction effect of the algorithm is limited,and there is still a big gap with the reference image.(2)At present,a method for spatiotemporal fusion using a convolutional neural network has been proposed,but the network is shallow and the fusion performance is limited.Aiming at the most widely used single-pair image spatio-temporal fusion,this paper uses a deep neural network to establish a novel spatio-temporal fusion method.The basic network framework consists of two 4x upsamplers cascaded to approximate the spatial difference between Landsat and MODIS.It is different from the sensor,and the reconstruction result is corrected by residual error to make the result closer to the real image.The method in this paper has been tested on different Landsat and MODIS satellite images and compared with multiple spatiotemporal fusion algorithms.Experiments show that the new method is better than the existing spatio-temporal fusion algorithm.(3)The current spatio-temporal fusion algorithm studies the data obtained from multi-source remote sensing satellites,while ignoring the research on the image data obtained by the same satellite sensor.To this end,this article takes the image data obtained by the Gaofen-1 satellite as an example.First,the wide-frame image is upsampled using the strategy of transfer learning,and then the up-sampling result is fused with the panchromatic band data with high spatial resolution.Generate multi-spectral image data with high temporal and spatial resolution.The experimental results show the feasibility of the fusion method proposed in this paper.