| There are many factors that can cause the change of atmospheric water vapor content(WVC).The inversion of atmosphere WVC by remote sensing data is often ill-posed because of insufficient observational information,which yields instability of the accuracy of traditional algorithms.To overcome this problem,we developed a novel fully-coupled paradigm,which uses artificial intelligence technology to combine traditional physical methods and statistical methods to stably invert atmospheric water vapor content from thermal infrared remote sensing data.The physics-statistical coupled paradigm was firstly derived from the theoretical derivation of the physical radiative transfer equation,we determined conditions that need to be satisfied invert WVC,and then built a generalized statistical method on the basis of the physical method.The representative solutions of physical methods and statistical methods were used as training and testing data of artificial intelligence.The representative solutions of the physical methods were obtained from the physical forward model,statistical solutions were obtained from satellite remote sensing data,assimilated and ground station data which can compensate for the defect that the physical model cannot simulate mixed pixels.In this study,fully connected neural network coupled physical and statistical methods of DL were selected.In addition,the land surface radiation was an important factor affecting accurate atmospheric water vapor.To improve the accuracy and portability of model retrieval,land surface temperature(LST)and land surface emissivity(LSE)were used as prior knowledge.We divided the models into two groups,one was the model using prior knowledge,the other was the model without prior knowledge,and compared and analyzed the importance of prior knowledge in different combinations.Then,the MODIS data is used for validation analysis,and theoretical inversion and verification are carried out for the combination of bands 27,28,29,31 and 32 that meet the algorithm conditions and corresponding prior knowledge data.Finally,the VIIRS sensor similar to the MODIS band design is used to verify the portability of the algorithm.The analysis showed that the DL coupled physics and statistics method based on prior knowledge can accurately invert the atmospheric WVC,which is a significant improvement over the traditional method and solves the problem of the lack of physical interpretation and universality of DL.This is a milestone in the history of agro-meteorological remote sensing key parameter inversion.The main conclusions are as follows:(1)Through physical logical reasoning analysis,it was found that brightness temperature of 4 thermal infrared bands were usually needed as the input parameters of DL.However,if we can ensure that brightness temperature information on the satellite was mainly from the atmosphere,the inversion equation can also be constructed with three thermal infrared bands.(2)The sensitivity analysis of prior knowledge showed that the greater the error of prior knowledge,the worse the accuracy of atmospheric water vapor inversion.High precision prior knowledge can improve the inversion accuracy to some extent,make the results more stable,and reduce the number of input parameters required by the algorithm.(3)Considering the data redundancy and band quality,mean absolute error(MAE)and root mean square error(RMSE)of atmospheric water vapor inversion can reach 0.05 g/cm~2and 0.07 g/cm~2 respectively.The best accuracy of actual inversion MAE was 0.23 g/cm~2,RMSE was 0.27 g/cm~2,and the error between the actual value and the inversion result is concentrated between(-0.5 g/cm~2,0.5 g/cm~2).(4)When there were two bands in the band combination that were very sensitive to atmospheric water vapor content,that was,input parameters and output parameters has a strong correlation,high-precision inversion can be achieved without prior knowledge.In the absence of such a band,it was more appropriate to use the LST and LSE as the prior knowledge to invert the atmospheric water vapor content.(5)The new inversion method was used to perform band combination inversion in VIIRS infrared band.By comparing the corresponding band combination results of MODIS,the results were found to be similar with high precision,indicating that the deep learning coupling algorithm based on prior knowledge has portability. |
