Resolution Enhanced Reconstruction Algorithm Of Spaceborne Microwave Remote Sensing Image Based On Deep Learning

Monitoring and forecasting disaster weather,such as typhoon,heavy rain and strong convection,is crucial for meteorological forecasting.The remote sensing brightness temperature image of the spaceborne microwave radiometer can invert the thermal structure and microphysical characteristics of the atmosphere,improving the monitoring and prediction of tropical cyclones.However,due to the limitation of the antenna aperture,the low spatial resolution of the satellite remote sensing microwave brightness temperature image hinders the observation of convective weather such as typhoon,rainfall,and ice cloud with small to medium scale.Therefore,it is particularly important for data post-processing algorithms to improve the resolution of microwave remote sensing brightness temperature images.The resolution of the brightness of the spaceborne microwave radiometer is affected by many factors during the imaging process,including antenna pattern,overlapping footprint,system noise,and the curvature of the earth.Many algorithms have been proposed for the microwave brightness temperature image reconstruction,such as BG,SIR,and Wiener filtering,having a large dependence on antenna pattern and the brightness temperature image prior information.Under normal circumstances,only one kind of factor of brightness temperature image degradation can be reduced to some extent.New type of microwave remote sensing brightness temperature image resolution enhancement algorithms based on deep learning is to introduce convolution neural network into microwave remote sensing brightness temperature image reconstruction application,solving the problem of insufficient resolution of microwave remote sensing brightness temperature image in low frequency channel and the resampling problem of multiple frequency observation data.The training data set is crucial for deep learning method.In this paper,two reconstruction algorithms based on deep learning,SRCNN-MW and VDSR-MW,and the methods of dataset generation are proposed.In this paper,the reanalysis data of the actual observation scene is input into the WRF model and the DOTLRT radiation transmission model to obtain the apparent bright temperature image,and the corresponding antenna brightness temperature image has been obtained by simulating the actual imaging process of the radiometer,thus constructing microwave brightness temperature reconstruction datasets.The two SRCNN-MW and VDSR-MW algorithm were implemented by supervised learning,in which the apparent brightness temperature image is regarded as the label of the antenna brightness temperature image.The SRCNN-MW algorithm directly learns the mapping relationship between two kinds of bright temperature images,however the VDSR-MW algorithm learns the mapping relationship between the residual image and the antenna brightness temperature image.The brightness temperature image experiments in this paper mainly include the stationary orbit simulation brightness temperature image reconstruction,the polar orbit satellite image reconstruction,ATMs actual observation brightness temperature reconstruction.The relationships between the SRCNN-MW network structure(network depth,width,kernel size)and its performance were explored in the static orbit microwave brightness temperature experiment,then the relationship between the VDSR-MW performance and the number of convolution layers was discussed.Experiments showed that the SRCNN-MW and VDSR-MW algorithms based on deep learning not only improve the resolution of the brightness temperature image but also the precise of the brightness temperature of the image pixels.Compared with the traditional reconstruction algorithm,the resolution improvement of SRCNN-MW and VDSR-MW are more obvious.The quantitative analysis showed that the brightness temperature value after the two algorithms reconstruction is closer to the value of the pixel corresponding to the apparent brightness temperature image.The two algorithms computation time are shorter and there are more suitable for practical applications.Compared with SRCNN-MW,VDSR-MW algorithm has provided generally better performances at computation time's expense.