Research On Near-surface Air Temperature Retrieval Method Based On Artificial Intelligence And Remote Sensing Data

Near-surface air temperature usually refers to the air temperature at a height of 1.5-2 m above the ground,and is a physical quantity that characterizes the degree of hot and cold air.It is a very important parameter in weather forecasting and agricultural disaster detection,as well as an important object of study in regional-global climate change,ecological change,urban heat island effect,crop growth and phenology,and extreme temperatures.Near-surface air temperature is influenced by a variety of factors in the three system units of land-air-sea,and is a key input parameter for various geophysical models,climate models,and numerical assimilation models of land surface.It is widely used in the fields of ecology,climatology,environmental science,carbon cycle and epidemiology.In recent years,near-surface temperature anomalies have led to frequent natural disasters such as droughts,low-temperature freezes,and high-temperature disasters,which also lead to increasing agricultural,livestock,and socioeconomic losses year by year.Therefore,it is of great importance to obtain the spatial and temporal distribution characteristics of near-surface temperature quickly and accurately for a more comprehensive understanding of the Earth-air-ocean energy exchange process,the spatial and temporal patterns of global changes and the change patterns.The inversion of near-surface air temperature from remote sensing data is usually pathological due to insufficient observational information.There are many factors that cause the variation of near-surface air temperature,which leads to the unstable accuracy of the conventional algorithm.Overcoming this problem,this study develops a new fully-coupled framework for stable direct inversion of instantaneous near-surface air temperature from thermal infrared remotely sensed data.The framework integrates physical methods,statistical methods,and deep learning,and is called the PS-DL method.The framework first establishes a complete set of solvable equations by deriving the physical radiative transfer equations to determine the existence of correlations between the deep learning inputs and outputs.The best combination of bands is selected to build the set of inverse equations for near-surface temperature,and a training and testing database is built using multiple sources of data,including physical model simulations,remote sensing data,and assimilation products.Finally,the developed fully-coupled framework is optimized and computed using deep learning to effectively solve the pathological problem of near-surface air temperature inversion,and the accuracy of near-surface air temperature inversion is greatly improved after using the surface temperature(LST)and surface emissivity(LSE)as a priori knowledge.Finally,the accuracy of the inversion algorithm and the performance of the analysis algorithm are evaluated from various aspects by validating them on MODIS data and Himawari-8/AHI data,respectively,and comparing them with the ground observation data.The main conclusions were obtained as follows:(1)For the MODIS data,the inversion maximum mean absolute error(MAE)and root mean square error(RMSE)of near-surface air temperature are 0.78 K and 0.89 K,respectively.In the cross-validation for the Chinese Meteorological Forcing Dataset(CMFD),the MAE and RMSE are 1.00 K and 1.29 K,respectively.In the ground validation of the actual inversion results for the best MODIS band combination MAE and RMSE are 1.21 K and 1.33 K,respectively.The proposed method effectively overcomes the limitations of the conventional method,because the inversion accuracy is improved by adding atmospheric water vapor and more waveband information,and the applicability(portability)of the algorithm is enhanced by using surface temperature and surface emissivity as a priori knowledge.has been enhanced.(2)For Himawari-8/AHI data,it was applied and validated in the Beijing-Tianjin-Hebei(JJJ)region and the Yangtze River Delta(YRD)region.The results showed that the MAE ranged from 1.36 K-2.8 K and RMSE ranged from 1.78 K-3.48 K in the JJJ region,and the MAE ranged from 1.04 K-1.70 K and RMSE ranged from 1.34 K-2.21 K in the YRD region.The inversion accuracy characteristics of different seasons and different times of the day are shown,with the highest accuracy in summer,followed by spring and autumn,and slightly worse in winter,and the inversion accuracy is generally better at night than during the day.Three validation methods are used: cross-validation on different data sets,out-of-region validation,and out-of-time validation to evaluate them.The actual inversion result of the best inversion combination is 1.36 K for MAE and 1.62 K for RMSE,which exhibit temporal stability and expansion.This study shows that the model can be called a universal inversion paradigm for geophysical parameter inversion and is a landmark due to its accuracy and physical interpretability of deep learning.