A Deep Learning Algorithm For Retrieving Land Surface Temperature Based On Simulated Data Of Radiative Transfer Model

The retrieval of land surface temperature(LST)is a typical "ill-posed" problem,and the difficulties in determining atmospheric conditions and land surface emissivity(LSE)bring serious challenges for retrieving LST using traditional methods.Deep learning,a good tool for retrieving LST,has the ability of simulating,solving complex equations and self-learning.Deep learning methods require a large number of high-quality and representative training samples.While obtaining sufficient samples to satisfy the requirements is impossible,which makes it difficult to apply in quantitative extraction of remote sensing.To address these problems,a deep learning algorithm based on simulated data of radiative transfer model for retrieving LST is proposed in this paper.In this method,MODTRAN is used to simulate a large number of real scene data as the training and test database which is then trained by the deep belief network(DBN)to get the LST retrieval models.The inputs of DBN models are only thermal infrared radiances and satellite zenith angle without any prior knowledge.LST retrieval models for Moderate Resolution Imaging Spectroradiometer(MODIS)and Visible Infrared Imaging Radiometer Suite(VIIRS)data were built in this study.Ground Surface Radiation Budget Network(SURFRAD)data were utilized for validation.The results showed that the DBN results are of high precision.And the DBN results had good spatial consistency with corresponding LST products.The main contents of this paper are as follows:(1)Using radiative transfer model simulates real scenes to build training and test database.Deep learning method requires a large number of reliable training samples.The radiative transfer model can overcome the errors in measurements and maintain the relationships between geophysical parameters accurately and its simulation data has high accuracy.Therefore,this paper proposed the use of radiative transfer model MODTRAN to simulate the real scene data as the training and test data set.The range of each input of MODTRAN was set by considering the characteristics of emissivity and LST of objects and the global atmospheric conditions,and the thermal infrared radiances were simulated according to the spectral response functions of sensors.And a large number of satellite signals at different scenes are obtained as the training and testing database of DBN.(2)Determining the parameters of satellite remote sensing for LST retrieval and building the deep belief network models.The selection of parameters for LST inversion has a direct impact on the demand of training data and the design of deep belief network.The radiative parameters and geometric parameters of satellite remote sensing for LST retrieval are determined by analyzing the radiative transfer process of different parameters.The setting of network parameters has important influence on the training accuracy of DBN.The network structure,the pre-training learning rate and the training times were determined through experiments.The pre-training times,the learning rate and the momentum of fine tuning were set according to the experience.(3)The retrieval of land surface temperature based on MODIS and VIIRS data and the authenticity test of the results.The DBN models were applied to MODIS and VIIRS data to retrieve LST.The ground SURFRAD data was used for validation the DBN retrieval results and LST products,and indexes such as bias,root mean square error(RMSE)and correlation coefficient(R)were applied to quantitative evaluations.The results suggested that the retrieved LST by DBN algorithm proposed in this paper had good consistency with ground measurements.This work solves the problem that LST retrieval relies on prior knowledge because of"ill-posed" inversion.The DBN algorithm can realize high-precision retrieval of LST based on a single satellite image without prior knowledge support,and it can effectively reduce or avoid the errors caused by the introduction of prior knowledge.