Optimization Algorithm And Analysis Of Cloudy Temperature Based On Microwave Radiometer

As an element of meteorological science research,atmospheric temperature is the main parameter reflecting water cycle and energy conversion in the earth’s climate system.Therefore,obtaining high-precision atmospheric temperature profile is the premise of carrying out atmospheric science research.Ground-based microwave radiometer,as one of the main equipment for meteorological parameter detection,can detect and retrieve the atmospheric parameter profile with high temporal and spatial resolution all day,but its own profile retrieval products are limited by regional climate differences,and the retrieval accuracy is easily affected by cloud and water.In this thesis,aiming at cloudy weather,genetic algorithm is used to optimize the sample set of neural network model,BP neural network model is constructed to retrieve atmospheric temperature profile,and RBF(radial basis function)neural network algorithm is used to further improve the retrieval accuracy and training time of cloud temperature profile.The retrieval results are compared with cloudy sky sounding truth values and microwave radiometer LV2 products,and the feasibility and advantages of the algorithm are discussed.The main contents are as follows:(1)Based on the 8-year historical radiosonde data provided by the National High Altitude Observatory of Jinghe Meteorological Station in Xi’an,the relative humidity of water surface is calculated and converted into the face value of ice according to temperature,relative humidity and atmospheric pressure,and the threshold model of relative humidity(ice surface)is established to determine the cloud boundary,and the cloud structure and stratification at the vertical height of 0-18 km in Xi’an are statistically analyzed.Calculate the vertical decline rate of atmospheric temperature,analyze the temperature change trend inside and outside the cloud and the upper and lower layers of the cloud boundary according to the decline rate profile,and make statistical analysis on the inversion of the cloud boundary,the relationship between the temperature decline rate in the cloud and the cloud height.(2)Combining with genetic algorithm,BP neural network model is established to retrieve atmospheric temperature.According to the measured brightness temperature data of microwave radiometer LV1 for three years,combined with historical sounding data,a qualified initial sample of the network model is formed,and the genetic algorithm is used to pre-train the model to optimize the sample subset,which is used as the final high-quality sample to train BP inversion temperature.Comparing the correlation analysis of BP-ANN inversion profile and LV2 product profile with sounding,the correlation coefficient of LV2 profile is 0.9634,and that of BP-ANN is 0.9815.From the root mean square error,the error of BP inversion profile is less than 1 K below 500 m near the ground,which is 0.3-0.5 K lower than that of LV2,0.5 K lower than that of 1-3 km height and about 1 k lower than that of 4-8 km height.LV2 profile error reaches 2 K at the height of 4 km,while BP profile error begins to exceed 2 K when it reaches 7.5 km.(3)Because BP neural network uses gradient descent method to adjust the weights,it is not a complete convex optimization problem to solve the atmospheric profile inversion,and it is easy to converge to local optimum and the convergence speed is slow.RBF not only has the global optimal performance of BP algorithm,but also has good local approximation ability,and the training is fast and easy.The use of RBFNN can effectively improve the precision of the inversion of the sky temperature,and increase the training time of the network from 130～140 s to 6～7 s on average,which greatly improves the training speed of the network.According to the comparison statistics between different height layers and sounding errors,the errors of RBF algorithm are all smaller than those of BP algorithm in the height of 0-8 km,in which the average root mean square error of RBF is 0.98 K for 0-2 km,1.16 K for 2-4 km,1.33 K for 4-6 km,1.72 K for 6-8 km,3.16 K for 8-10 km,and the total height average for 0-10 km.For different cloud heights,the inversion effect of RBFNN model is better than that of BP model in low cloud weather,and the middle cloud weather is slightly improved.The errors of the two models are the same in high cloud weather,and the accuracy improvement effect is the same.