Reaearch Of Ground-based Microwave Radiometer Inversion Algorithm Based On Machine Learning In Semi-arid Regions

The atmospheric temperature and humidity profile is an important parameter to describe the variations of atmospheric heat and dynamic state.The traditional method is tethered balloon detection method which can be used to observed under a variety of complex weather conditions with high operation cost and low observation frequency.Ground-based microwave radiometer has automatic observation capability with high observation frequency,which has a great significance for short-term forecasting and weather modification.The studying of this paper is based on the sounding data of the national reference climatological station in Yuzhong and the microwave radiometer data of the Lanzhou University Semi-Arid Climate and Environment Observatory(SACOL).The correction algorithm of cloud sample in brightness observation is be proposed based on SVR algorithm,the applicability of different training method and the effectiveness of BP neural network,RBF neural network,multiple linear regression and SVR in the field of microwave radiometer inversion has been compared in this paper.The OBS-SVR is been selected as the optimal algorithm to inverse temperature,relative humidity and water vapor density are inversed in hourly resolution to test the applicability of OBS-SVR in semi-arid areas under various weather condition,and the inversed data is used to study the variation characteristics of atmospheric boundary layer in semi-arid regions,the main conclusions are as follows:(1)The final thresholds for judging cloud sample is 268 K infrared brightness temperature.Comparison of the cloud-affected sample and the MonoRTM simulation brightness temperature show that clouds has a great influence on 1-7 channels,which will cause an abnormal increasing in brightness temperature.After revised by SVR algorithm,the root mean square error of observation brightness temperature of 12 channels are reduced,and the SVR model better in 4th,5th,and 6th channels than others.The effect is most significant,but it causes bias increasing in9~12 channel,so this method is only applicable to the observation brightness temperature correction of 1~8 channels.(2)The Comparison of observation Tb training method(OBS)and Simulation Tb training method(SIM)with four machine learning algorithms show that,for temperature profiles inversion,simulation Tb training method work well in all four algorithms near the surface of 0~0.2 km,on all another level observation Tb training method work better;for relative humidity profiles inversion,the comparison results of RBFNN,BPNN and SVR are similar to the temperature profile's,the OBS-MLR has advantages at 0.4-3.5 km,and SIM-MLR work better at remaining level;and the comparison results of water vapor density profiles are similar to the relative humidity profiles.It can be concluded that OBS method has advantage in three nonlinear algorithms for the inversion algorithm training of three meteorological elements profiles,and only work better on linear algorithm training of the temperature profiles.(3)The comparing result of the four inversion algorithms obtained by OBS method show that,the RMSE of inversed temperature profiles of four algorithms is increasing with height and close under 4 km,above 4 km the RMSE of OBS-SVR is smaller than other three algorithms;for relative humidity profiles,the RMSE of four algorithm is increasing with height under 5 km,and decreasing with height above 5 km,the RMSE of OBS-SVR and OBS-BP are smaller than the other two algorithms at all levels,the correlation coefficient of the temperature profile obtained by OBS-SVR inversion is greater than OBS-BP at all height levels;the RMSE of water vapor density profile obtained by OBS-SVE is decreasing with height,the other three algorithms' is increasing under 1km,and decreasing above 1 km,and the RMSE of OBS-SVR is smaller than other three algorithms at all levels except near-surface level.(4)The OBS-SVR is used to inverse hourly profiles from June 2009 to 2010 to study the daily variation characteristic of the boundary layer in the semi-arid area represented Yuzhong.In Yuzhong the average annual maximum boundary layer height is 1163 m at 15:00,and the minimum is 304 m at 6:00.The seasonal variation is significant different,in spring and summer,the BLH maintains and fluctuates after reaching the maximum height,and in autumn and winter,it begins to decrease after reaches the maximum.Due to the different sunrise time of four season and the influence of solar radiation,the four season boundary layer begins to develop at different times,it is 7:00 in spring,6:00 in summer,8:00 in autumn and 9:00 in winter,and the maximum BLH is reduced from 1577 m in summer to 1012 m in spring.(5)The OBS-SVR inverting profiles to analysis the case of sunny and cloudy days,and compare the difference of boundary layer development and meteorological elements variation under sunny and cloudy weather conditions.In sunny days,the BLH changes significantly under sunny weather conditions,after sunrise,the solar radiation heats the surface and the temperature begins to rise,convection develops vigorously,water vapor is transported from near-surface levels to higher levels,after sunset,the ground is radiative cooling,and the convection is weakened and inversion layer is begin to form,the residual layer retains the characteristics of the daytimemixed boundary layer,and the water vapor distribution is uniform in the residual layer;under cloudy weather conditions,the daily variation of boundary layer height is not obvious,the transportation of material and energy is weakened,and the water vapor transportation effect is not obvious.(6)The OBS-SVR can be used to inverted the temperature,relative humidity and water vapor density profiles complex weather conditions throughout the year,and its inverted data can be used to analysis the meteorological elements characteristic under sunny and cloudy condition,and the relative humidity data can be used to analyze the humidity variation in the cloud,indicating that the SVR correction algorithm can reduce the cloud impact on observations and improve the accuracy.OBS-SVR has good capability of generalization that can be used under different weather conditions in semi-arid areas and has advantages in localization in arid regions.