Spaceborne Microwave Radiometer Image Restoration Based On Deep Learning

The satellite-borne microwave radiometer can obtain temperature characteristic information of multi-band in the observation for the earth,and plays an important role in the meteorological parameter inversion.However,the limitation of the antenna size in the satellite payload design,the system noise and the curvature of the earth in the scanning process,etc.those degradation factor affected the quality of the data received by the microwave radiometer in different degrees.In this paper,a method based deep learning is first proposed to improve the quality of the radiometer data.This method handle the degraded effects of complex degrading factors such as antenna pattern,system noise,and curvature of the earth on the observation based on the data of the microwave radiation imager on Fengyun-3.In the imaging process of microwave radiometers,due to the influence of satellite payload design,the antenna aperture cannot be designed to be too large,which result in diffraction blur caused by low-pass filtering effects,and image quality will be affected.Meanwhile,system noise and curvature of the earth will also affect image quality.The traditional method of Wiener filtering deconvolution establishes a digital filter in the Fourier domain to eliminate the diffraction blur,and then handle the other degradation factors.This will not only make it difficult to ensure the reconstruction accuracy,but also increase the complexity of the image restoration model.This paper first proposed a general framework based on deep learning to solve the problem of image degradation caused by multiple degradation factors in the radiometer imaging process.The image restoration is defined as a regression problem in the spatial domain.The higher reconstruction accuracy can be obtained through data-driven ideas and multiple transformation of feature space.Before training on the deep learning model,a flexible method of making dataset was proposed.Using the characteristics of the multi-frequency band of the microwave radiometer,the data of the high resolution frequency band was defined as the simulated apparent temperature of the dateset.A large amount of sample can be obtained through established image degradation model.The experimental results show that the trained model achieve good results both in quantitative evaluation indicators and subjective visual effects.For the problem of image degradation caused by non-oversampling in the microwave radiometer scanning process,through the above-mentioned image restoration framework based on deep learning,a 32-layer deep residual network is used to handle super-resolution resconstruction of non-oversampling data.The experimental results show that deep residual network can obtain better reconstruction accuracy than general deep convolutional neural network.The image restoration technology can effectively improve the quality of the brightness temperature image of the microwave radiometer,and can provide high-quality brightness temperature data when the bright-temperature image data is used.This has an important influence on the accurate inversion of meteorological and physical parameters such as soil moisture,sea surface temperature,precipitation,water vapor content,ice thickness,and so on.