Research On The Generation Method Of Spatially Seamless Passive Microwave Land Surface Temperature

Land Surface Temperature(LST)is a key factor in the energy interaction at the ground-air interface and is a direct characterization of surface thermal radiation and an input parameter for various ground-air models.Traditional thermal infrared(TIR)remote sensing is the mainstream way to estimate LST,but it can only obtain the LST under clear sky conditions due to the physical mechanism.Passive microwave(MW)remote sensing has a unique advantage in monitoring surface temperature changes at large regional scales,filling the missing surface temperature time series due to cloud cover,and generating medium resolution(e.g.,1 km)all-weather LST in integrating with TIR remote sensing.However,the use of MW remote sensing for the estimation of LST is still facing the problems of low accuracy and the existence of gaps in MW brightness temperatures(BTs)images,etc.To address these problems,the main research works of this paper were as follows:(1)A high-precision MW remote sensing LST retrieval model was constructed based on neural network.The performance of the neural network and two deep learning networks(deep belief network-DBN and convolutional neural network-CNN)in the estimation of the MW LST was compared with the MODIS LST as the label value and a series of surface and atmospheric parameters as the input parameters.The examination results showed that CNN was better than NN and DBN by 0.1-0.4 K.The test results for multiple combinations of input parameters showed that adding some easily available parameters(e.g.,air temperature at 2 m)on the basis of BTs can significantly improve the accuracy of the LST estimate.The validation results showed that the RMSD of CNN LST was 2.2-3.6 K during the daytime and 1.4-2.2 K during the nighttime.The crosscomparison results showed that the LST estimate was closer to the MODIS LST than the Glob Temperature AMSR-E LST,and maintained similar intra-annual trends when compared with the air temperature data.The comparison with other commonly used methods(look-up table algorithm and the single channel algorithm)showed that the CNN model can significantly improve the estimation accuracies(above 1 K)of MW LST on barren land pixels.(2)A MW BT gap filling method was proposed based on forward simulation.From the perspective of forwarding simulation,based on the spatial seamlessness of the reanalysis data,the deep neural network was used to achieve implicit correction of the reanalysis data and thus to fill the gaps of the MW BTs images.In order to characterize the possible frozen soil and snow cover in the MW pixel,the polarization ratio(PR)and frequency ratio(FR)were used as input parameters on the basis of surface parameters,subsurface parameters,and atmospheric parameters to enhance the forward simulation capability of the deep neural network for MW emission and atmospheric radiation transmission processes.The test results based on the missing value of simulated BT showed that the RMSD of each channel in the Qinghai-Tibet Plateau study area was 2.5-3.7 K during the daytime and 2.5-4.1 K during the nighttime,and the RMSD of each channel in the southern China study area was 1.6-2.4 K during the daytime and 1.4-2.1 K during the nighttime.The sensitivity analysis showed that the obtained PR and FR values based on temporal linear interpolation had little effect on the accuracy of the BTs prediction in the gaps.(3)The LST estimated from the BT predictions for the Qinghai-Tibet Plateau study area and Southern China study area were compared with the CNN LST.The results showed that the RMSD of the LST due to the deviation of BT predictions was less than2.4 K during the daytime and less than 1.4 K during the nighttime in the Qinghai-Tibet Plateau study area.And the RMSD of the LST due to the deviation of BT predictions was less than 1.6 K during the daytime and less than 1.1 K during the nighttime in the Southern China study area.The validation results at the three in-situ sites(i.e.,CBS,TYU,and DXI)showed that the root-mean-square error(RMSE)of LST due to BT prediction deviation was less than 1 K.Therefore,although there is some deviation in the BT prediction value compared with the original BT value,it has limited impact on the accuracy of the final LST estimate.This study further confirms that the BT prediction model constructed by the deep neural network has high accuracy and the generated complete MW BT sequence can be applied to the estimation of the LST and other related remote sensing parameters.