Retrieval Of Microwave Land Surface Emissivity In Taklimakan Desert Area

Land surface emissivity is an inherent property of the ground surface and is defined as the ratio of the radiation emission of an object at temperature T and wavelength ? to the radiation emission of black body at the same temperature and wavelength.Microwave land surface emissivity is an important parameter for assimilating microwave radiation data from various satellites,accurate values will increase the utilization of spaceborne microwave radiometer data.The microwave land surface emissivity is easily affected by the characteristics of the surface itself and the observation conditions of the sensors,which brings a certain degree of difficulty to large-scale retrieval.Therefore,in order to improve the simulation accuracy of land surface emissivity and brightness temperature,different microwave land surface emissivity retrieval models are established here.First,this article is based on Taylor's formula of the multivariate function and optimal control principle,select the central area of the Taklimakan Desert as the retrieval area,and the surface temperature,surface humidity,0.07 m soil moisture below the surface and 0.28 m soil moisture below the surface as the influencing factors.the linear retrieval models between the microwave land surface emissivity and two(four)influencing factors are constructed by using the data of observed brightness temperature from FY-3C Microwave Radiation Imager and simulated brightness temperature from Community Radiative Transfer Model(CRTM)on January 7,13,17,18,and 24,2014.Taking the observed brightness temperature in the retrieval area as a reference,the average deviations of the simulate brightness temperature from the two linear retrieval models have been reduced to 1.1324 K and 0.9919 K,respectively,it is only 56.36% and 49.37% of the average deviations of the simulate brightness temperature from the original land surface emissivity.The independence test in the whole Taklimakan Desert shows that the average deviations of 18 th in retrieval date and Jan 29 th which is not used in retrieval date were reduced from 2.6157 K,4.9997 K to 1.8789 K,1.6235 K,respectively.It shows that the two linear retrieval models improved the retrieval accuracy of the land surface emissivity in the desert area in January,and the improvement of the 4-factors linear retrieval model is higher.Secondly,because microwave land surface emissivity is affected by many factors,and the linear retrieval model may be difficult to accurately describe the functional relationship between the land surface emissivity and its influencing factors.The neural network method can fit arbitrarily complex non-linear relationships,which makes up for this shortcoming of the linear retrieval model.However,traditional neural network algorithms have problems such as poor generalization ability,easy to fall into local optimum,and slow convergence speed.Therefore,this article takes the input of surface temperature,surface humidity,0.07 m soil moisture below the surface and 0.28 m soil moisture below the surface as input,and the true value of the surface emissivity is output.A microwave land surface emissivity retrieval model based on Bayesian regularization of LM-BP neural network was proposed to improve the retrieval accuracy of land surface emissivity.In order to compare the retrieval accuracy of the 4-factors linear retrieval model,the model selected the same retrieval date(January 7,13,17,18,and 24,2014),but the Bayesian regularized of LM-BP neural network model requires a large amount of data to learn,therefore,the whole Taklimakan Desert was selected for the retrieval area.The test shows that the average deviation of the Bayesian regularized of LM-BP neural network model on the training date(18th)is reduced from 2.6157 K to 0.5579 K of the original model,which is far less than the improvement degree of the 4-factor linear retrieval model on the training date.On the 29 th without training,the average deviation was reduced from 4.9997 K to 1.6906 K.at this time,the degree of improvement of the two models is similar.Finally,based on the Bayesian regularized of LM-BP neural network model,training dates were added,and two types of surface emissivity retrieval models were established in the winter of 2014-2018 through a combination model of different input variables.Resulting from time independence test,the M1 model,which takes surface temperature,surface humidity,0.07 m soil moisture below the surface and 0.28 m soil moisture below the surface as input,has higher retrieval accuracy and less training time;the M2 model with the observed brightness temperature from FY-3C Microwave Radiation Imager 10.65GHz(H & V),18.7GHz(H & V),36.5GHz(H & V),and 89GHz(H & V)as input,has better stability.And the retrieval models obtained from the two inputs have good retrieval accuracy.It can be seen that the 2-factors and 4-factors linear retrieval models established in this paper and the Bayesian regularized of LM-BP neural network model have certain simulation accuracy.Among them,the Bayesian regularized of LM-BP neural network model has the best improvement and it can be used in the retrieval of microwave land surface emissivity in winter.