Research On The Super-resolution Of Remote Sensing Images With Whisk-brooming Space Camera

Faced with the demanding need of national defense security and economic con-struction,the new generation of space optical remote sensing is heading towards the trend of high spatio-temporal resolution and ultra-wide coverage imaging.This brand-new technology plays an essential role from the national strategic layout and high-tech industrialization of space to land inventory and disaster management.Traditional op-tical lens making is limited by the manufacturing technology and production cost.If pushbroom scanning is adopted,the obtained remote sensing imageries cannot juggle high resolution and wide swath at the same time.On the other hand,the mode of large-angle whisk-broom scanning of space camera is able to break through the limitation,as well as the contradiction between resolution and imaging swath of optical satellite.However,the "high resolution" here is only for the nadir point.In the region deviating from the nadir point,the image will still be distorted and its quality will deteriorate.Therefore,it is necessary to design a recovery algorithm for the problem mentioned above,achieving super resolution without costing the hardware.Present with the problem of distortion and resolution degradation of image ob-tained by whisk-broom scanning,this thesis explores the solution of geometric correc-tion and super resolution for the forementioned problem.In the first part,the research progress is firstly introduced of geometric correction for whisk-broom remote sens-ing image,deep learning for image super-resolution and remote sensing image super-resolution.Secondly,according to the mechanism of whiskbrooming,the resolution degeneration of ground scenery is analysed in detail.After that,the resolution inver-sion model of space camera large-angle whiskbroom image is proposed.The image geometry correction is preliminarily completed.Finally,ground simulation experiment is devised to verify the validity of the resolution inversion algorithm,according to the principle of fixing scaled ratio.For the second part,an improved super-resolution model with generative adversarial network is proposed to get rid of the blur due to the interpola-tion,further enhancing the perceptual effect of the imageries.Firstly,the basic principle of super-resolution with generative adversarial network is introduced from the perspec-tive of model architecture and loss function.Secondly,we propose our improved loss function and feature extractor.A deep learning super-resolved neural network model is then designed to fit this project.In the final experiment,a real-scene remote sensing image is utilized to create the training,validation and testing set.After our model is iter-atively trained,the reconstructed images achieve great perceptual effect,with two image quality indicators Peak Signal to Noise Ratio and Structural Similarity being 18.6049 and 0.6022 on validation set,respectively.The two numbers reach 19.6652 and 0.6045 on the testing set.In this thesis,a resolution inversion model and an improved super-resolution archi-tecture with generative adversarial network are proposed,which help realize the super-resolution of remote sensing images with whisk-brooming aerospace camera.This project facilitates the civilian-orientation of high resolution remote sensing satellite and industrialization of national aerospace high-tech.At the same time,it provides theo-retical basis for the brand-new imaging method of high resolution and wide swath with space-based payload,and helps solve the common scientific problems in the area of dynamic remote sensing.