融合多源观测资料的微波辐射计温度廓线订正试验

(4): 494-501 Experiments on Microwave Radiometer Temperature Profile Correction by Integrating Multi-Source Observational Data

关键词:  微波辐射计  FY-4A卫星  AMDAR  BP神经网络  遗传算法  廓线订正

Key words:  microwave radiometer  FY-4A satellite  AMDAR  BP neural network  genetic algorithm  profile correction

摘要:

为增加大气探空站点，提高微波辐射计大气温度探测精度，利用 FY-4A 气象卫星温度产品和 BP 神经网络、遗传算法，分别实施杭州站、南京站微波辐射计的温度订正仿真试验，并获得准确的连续性大气温度垂直廓线；结合探空资料和民航 AMDAR 气温资料，评估模型订正效果。研究结果表明:(1) 微波辐射计温度产品存在一定误差，两站均在高度 2 km 处平均偏差最大，同站有雨时的偏差均大于无雨时的偏差；(2) 经过 BP神经网络模拟订正后的微波辐射计测温精度较原温度产品提升幅度较大；杭州站 MAE、MSE、RMSE 的降低幅度分别为 45%~55%、65%~78%、41%~53%，南京站的降低幅度分别为 58%~66%、83%~88%、55%~59%；(3) 经过遗传算法优化初始权值和阈值后的神经网络订正模型模拟效果有进一步的提升，其中有雨模型提升效果明显，RMSE降低幅度 11%~15%。微波辐射计的上述订正方法，可以推广到各地微波辐射计站点应用，具有实际使用价值。

Abstract:

In the present study, a correction method was developed to improve the accuracy of ground- based microwave radiometers in measuring temperatures. Using temperature products from the FY-4A meteorological satellite, a back propagation (BP) neural network, and a genetic algorithm, we conducted temperature correction simulation experiments to correct the temperature profiles measured by two MP- 3000 ground-based microwave radiometers located at the meteorological stations in Hangzhou and Nanjing, respectively, and obtained accurate and continuous vertical profiles of atmospheric temperature. The corrected temperature profiles were then compared with temperature data from radiosonde measurements and the Aircraft Meteorological Data Relay (AMDAR) data from the Civil Aviation Administration of China. The results show that: (1) Microwave radiometer temperature products exhibited inherent inaccuracies, with larger discrepancies during rainy conditions and the greatest average deviation observed at the altitude of 2 km for both stations. (2) The temperature measured by the microwave radiometer, after being corrected through BP neural network simulation, was a significant enhancement compared to the original temperature. At Hangzhou station, the reductions in mean absolute error, mean squared error, and root mean square error (RMSE) were observed in the ranges of 45%~55%, 65%~78%, and 41%~53%, respectively, while at Nanjing station, these metrics decreased by 58%~66%, 83%~88%, and 55%~59% respectively. (3) The simulation model of the neural network, after its initial weights and thresholds were optimized using a genetic algorithm, demonstrated further improvements. There was a significant enhancement in the rain model, with RMSE reductions of 11%-15%. The proposed correction method for microwave radiometers seems to be suitable for broader applications across microwave radiometer stations.