Application Of Deep Learning Methods In Spatiotemporal Reconstruction Of Soil Moisture Dat

Soil moisture is an important factor affecting the water and energy cycles between the earth’s surface and the atmosphere,and it is important to obtain the accurate spatial and temporal distribution of soil moisture for related research.At present,remote sensing is an important means to obtain soil moisture data on a large spatial scale,but satellite-based soil moisture data still has a large data gaps problem,which limits the development of related research.To address the spatial and temporal gaps in satellite-based soil moisture data,this paper firstly uses convolutional neural network deep learning reconstruction method for global soil moisture single-variable reconstruction based on two sets of satellite merged soil moisture data,CCI and CM,to obtain global long-term no gaps soil moisture data;then,based on the CCI satellite remote sensing fused soil moisture data in the Chinese region,and the multi-variable dataset（named CN05.1）obtained by interpolation based on the observed data,including precipitation,temperature,relative humidity,sunshine hours and wind speed data,the convolutional neural network deep learning method is used for multi-variable reconstruction of soil moisture data to obtain the China long-term no gaps soil moisture data;finally,the fully connected neural network is used to finely rasterize the reconstructed data of Poyang Lake basin to obtain high spatial resolution long-term no gaps soil moisture data of Poyang Lake basin.The main research conclusions of this paper are as follows:（1）In the study of global soil moisture single-variable reconstruction,the feasibility and effectiveness of the convolutional neural network for single-variable reconstruction are demonstrated by adding cross-validation experiments with missing simulations;the used deep learning single-variable reconstruction method can effectively learn the temporal and spatial information of the data for reconstruction,and the spatial characteristics,temporal characteristics and data distribution characteristics of the reconstructed data are consistent with the original data.The results of the comparison with the in-situ data show that the reconstructed data effectively retains the data quality characteristics of the original data,and the overall quality of the four evaluation indexes compared with the in-situ data shows that the CCI_recis better than the CM_rec.The results of the global soil moisture data consistency analysis show that CCI_recand CM_recshow a high positive correlation in most regions of the global,and only in southwestern China and the Sahara Desert region are the differences larger and show a slight negative correlation.（2）In the study of multi-variable reconstruction of soil moisture in China,the cross-validation reconstruction results with missing simulations show that the convolutional neural network used in this paper for multi-variable reconstruction also has high feasibility and effectiveness;the spatial and temporal characteristics of single-variable and multi-variable reconstructed data have high consistency with the original data in general,but the temporal and spatial detailed characteristics of multi-variable reconstructed data have better performance than usingle-variable reconstructed data;compared with the original data,both single-variable and multi-variable reconstructed data have some improvement in data quality and the combined results of the four evaluation indexes show that the overall quality of CCI_mvis better than that of CCI_sv;from the results of simulated continuous multi-day missing,it can be seen that the overall reconstruction effect of CCI_mvis significantly better than that of CCI_svdue to the use of other auxiliary variables and the ability to portray the details of the data in time and space is significantly stronger than that of CCI_sv.（3）In the study of fine rasterized reconstruction of soil moisture in Poyang Lake basin,the results of model training loss function and testset reconstruction show that the fully connected deep learning method used in this paper has high feasibility and effectiveness whether using the first 133 components or all components for reconstruction,and the reconstruction effect and performance of the optimal model trained with the first 133components and the optimal model using all components are basically the same.The spatial distribution pattern and the time series of daily maximum and mean values of the reconstructed data are consistent with the original data,indicating that the spatial and temporal reconstruction of the reconstructed data is effective.From the comparison results with the in-situ data,it can be seen that the reconstructed data effectively retains most of the data quality characteristics of the original data,and the time correlation coefficients of the reconstructed data has been improved.