Global Soil Moisture Inversion And Data Reconstruction Based On FY-3 Satellite Microwave Brightness Temperatur

Soil moisture is an important variable in energy and water exchange between land and atmosphere.It plays an important role in many fields such as climate,hydrology and agriculture.Therefore,it is important to obtain large-scale soil moisture data.Passive microwave remote sensing can detect the all-weather and has penetration ability to clouds,fog and vegetation.So,it is the main measure of large-scale soil moisture monitoring.FY-3B,FY-3C and FY-3D satellites of China Meteorological Administration provide passive microwave data from 2011 to present,enriching the data source of global soil moisture retrieval.In this paper,we retrieve the global soil moisture based on the LPRM method and FY-3B and FY-3D passive microwave data.And we improve the soil moisture retrieval algorithm and obtain the global soil moisture based on the TCA method and FY-3C passive microwave data;Then,we merge the soil moisture data for prolong the timeseries length of single satellite data based on the SNR-Opt method and the soil moisture retrieval data;Finally,the merged data is reconstructed by deep learning method to fill in the missing value.The main conclusions of this paper are as follows:（1）The retrieval accuracy of FY-3C is more stable than that of night observation,especially in sparse vegetation area;By comparing FY-3B and FY-3C with different observation time,it can be found that the data with observation time in the daytime（FY-3B ascending and FY-3C descending）have higher accuracy in humid areas,and the satellite observation data at night（FY-3B descending and FY-3C ascending）have higher accuracy in arid areas.（2）SNR-Opt method is used to merge FY-3B/3C/3D soil moisture datasets,which prolongs the timeseries length of satellite data.Using the in-situ observation data for verification,the results show that the timeseries change characteristics of the merged data are consistent with the in-situ observation data,and its value is closer to the in-situ observation data than the retrieval data;Through the comparative analysis with SMAP dataset,it can be found that the mean square error of FY-3 merged dataset is less than SMAP dataset,and the improvement effect of merged dataset is not significant in terms of correlation.（3）Using the deep learning method to reconstruct the merged data,the missing values in the dataset are filled,and the soil moisture dataset with continuous spatial distribution is obtained.The loss function value of the data reconstruction model is greater than 0.05 at the beginning of the iteration,stable below 0.05 after 100 durations,and RMSE reaches the lowest after 1000 iterations.Using the station observation data and fusion data for verification,it can be found that the reconstructed data has high correlation with the fusion data,the maximum range of numerical distribution density is 0.2-0.35 m³m^-3,and the reconstructed data is consistent with the seasonal variation characteristics of the in-situ data.