| Remote sensing image spatio-temporal fusion is an effective method to solve the mutual constraints of remote sensing data on spatial resolution and time resolution.It can generate remote sensing images with both spatial and temporal resolution without changing the existing observation conditions,and broadly enhance the application scenarios of remote sensing images.So as to achieve dynamic detection of the surface at a higher spatial and temporal resolution.This research takes the central part of Shanxi Province as the research area,selects the remote sensing image Landsat8 with low time and high spatial resolution and remote sensing image MODIS with high time and low spatial resolution as the research data set,and uses multi-stream data input,attention mechanism module,Differential images based on time-dependent and time-constrained rules are introduced into the improvement of the classic model SRCNN in the field of super-resolution reconstruction,and an improved SRCNN remote sensing image spatiotemporal fusion M-SRCNN model based on multi-stream data input is constructed.Improved attention mechanism SRCNN remote sensing image spatiotemporal fusion MBASRCNN model,improved SRCNN remote sensing image spatiotemporal fusion Diff-MBASRCNN model based on difference images,and adopts objective evaluation(PSNR,SSIM)and subjective evaluation to simulate the effect of the model Evaluation.The main conclusions of this study are as follows:(1)The high-resolution and low-resolution images of the non-predicted time phase are added as a priori information to the input of the SRCNN model,and the M-SRCNN model is established.The M-SRCNN model enables the network to extract fine features at high resolution image resolution,while ensuring that features at different resolutions correspond to each other.Use the M-SRCNN model to perform spatio-temporal fusion of Landsat8 and MODIS images,and use objective evaluation(PSNR,SSIM)and subjective evaluation to evaluate the simulation effect of the model.Compared with the SRCNN model,the PSNR and SSIM of the M-SRCNN model are improved by 4.0667 and 0.0840,respectively,and the optimization effect of the RED band is the best,which is improved by 5.8152 and 0.1398 respectively.Multi-stream data input can significantly eliminate the distortion of the SRCNN model results,and has a significant improvement effect on the problems of blurred image edges and poor detail information.The spatio-temporal fusion effect of the M-SRCNN model is significantly improved.(2)Aiming at the problem that the M-SRCNN model is relatively weak in the reconstruction of low-frequency detail information such as houses in the fields,the CBAM attention mechanism module is introduced,and the MBA-SRCNN model is established.Compared with the M-SRCNN model,the MBA-SRCNN model improves the two evaluation indicators of PSNR and SSIM by 1.3847 and 0.0247 respectively.The MBA-SRCNN model has a more obvious optimization effect in the reconstruction of low-frequency detail information,and the extraction accuracy of roads and houses has been significantly improved.(3)The M-SRCNN model and the MBA-SRCNN model can significantly improve the distortion and poor detail performance of the SRCNN model,but the edge expression is still weak,and the experiment will produce misty noise.The difference images based on timedependent and time-constrained rules are introduced into the MBA-SRCNN model,and the Diff-MBA-SRCNN model is established,which can express the characteristics of the areas where the types of features change and the differences in the details of the images.Compared with the MBA-SRCNN model,the PSNR and SSIM of the Diff-MBA-SRCNN model are improved by 1.4120 and 0.0189 respectively.The Diff-MBA-SRCNN model can effectively improve the edge expression ability of low-frequency detail information,and weaken the influence of noise such as haze and irrelevant results on the results of image spatio-temporal fusion.The Diff-MBA-SRCNN model can be used as an effective method for the spatiotemporal fusion of Landsat8 and MODIS remote sensing images.The Diff-MBA-SRCNN model constructed in this study can more accurately generate remote sensing images with high time and high spatial resolution,increase the amount of information carried by the images,and further broaden the application scenarios of remote sensing images for agriculture,military and urban planning,etc.Precise measures in the field provide theoretical reference.The research results of this study can provide an effective new idea for the spatio-temporal fusion of remote sensing images,and have certain promotion significance for the research in the field of spatio-temporal fusion of remote sensing images. |
