Research On Remote Sensing Retrieval Method Of Soil Moisture And Air Temperature Based On Machine Learning

Soil moisture(SM)and surface air temperature(SAT)are important climatic variables,which are critical in weather and climate evolution,energy cycle,and hydrological cycle.Large-scale,high-precision,and high-quality SM and SAT data are of great significance for preventing disasters and reducing damages.To obtain timely,accurate,comprehensive,and stable data,it is necessary to carry out researches on SM retrieval and SAT retrieval based on remote sensing observations.This thesis aims to study the retrieval methods based on machine learning in terms of SM and SAT from the perspectives of data-driven and problem-driven.The performance and retrieval accuracy of each model are evaluated and analyzed.The main research contents and conclusions of this paper are as follows:(1)SM retrieval from satellite observations based on the deep convolutional neural network.The SM retrieval model based on deep convolutional neural network(DCNN)was established for solving the multi-source data fusion,using the microwave brightness temperature of Soil Moisture and Ocean Salinity(SMOS),backscattering coefficient of Advanced Scatterometer(ASCAT),Moderate Resolution Imaging Spectroradiometer(MODIS)Normalized Difference Vegetation Index(NDVI)and soil temperature.The retrieved results were compared with insitu measurements and soil moisture reanalysis products.The experimental results show that DCNN performs very well when compared with in-situ measurements at the continental scale and gives better results than backward propagation neural networks(NN).Besides,the temporal and spatial correlations of the retrieved soil moisture by NN(DCNN)and soil moisture products from the European Center for Medium-Range Weather Forecasts Reanalysis Interim(ERAInterim)were 0.558-0.570(0.5680.576)and 0.922-0.924(0.926-0.927),respectively.It is proved that DCNN can be effectively used to obtain near-real-time SM on a large scale.(2)SM retrieval from satellite observations based on multi-view and multi-task learning.The SM retrieval model based on multi-view multi-task learning(MVMTL)was designed according to the spatial and temporal distribution characteristics of SM and the prior assumptions of geoscience.First,temporal and spatial features were extracted from SMOS microwave brightness temperature,ASCAT backscattering coefficient,MODIS NDVI,and soil temperature data.Then,the spatial and temporal features extracted from each grid cell were used to estimate the corresponding SM value from the perspective of multi-view learning.Finally,the spatial correlation among different grid cells was used to carry out the joint estimation of SM value for all grid cells over the continental U.S.using multi-task learning.Results show that the temporal and spatial correlations between the retrieval results and the ERA-Interim SM products are 0.527 and 0.892,respectively.Compared with the in-situ measurements from Soil Climate Analysis Network(SCAN),the model can achieve better performance than other models,and the mean and median values of temporal correlation are improved by 3.54% and 5.33%.(3)Fined soil moisture retrieval from satellite observations based on multi-view and multitask learning.To verify the potential of Fengyun-3B/Microwave Radiation Imager(FY-3B/MWRI)microwave brightness temperature for soil moisture retrieval and the applicability of the improved MVMTL method in the Tibetan Plateau region.Fined MVMTL for SM retrieval was designed,taking the application requirements of satellite-borne microwave remote sensing for SM retrieval and a priori assumption of spatial correlation into account.First,FY-3B MWRI microwave brightness temperature,MODIS NDVI,albedo,topography,soil temperature,roughness,and other data were divided into different parts according to their spatiotemporal attributes.Then,the corresponding temporal or spatial,or spatiotemporal of the corresponding grid cells were extracted according to different spatiotemporal attributes.Finally,the model parameter MVMTL was optimized via the mean square loss function,considering the constraints of spatiotemporal inconsistent,spatial correlation constraints for multi-task and L21 regularization,and then the joint estimation of soil moisture of the whole study area was estimated.Compared with the ERA-Interim SM products against the in-situ measurements,the presented method has a smaller root mean square error(RMSE)and a higher correlation skill.Comparing with the in-situ measurements from the observation network at different resolutions,the retrieval technique of MVMTL is inferior to the FY-3B SM level 2 product but better than that of the ERA-Interim SM products.(4)Surface air temperature retrieval from satellite observations based on stepwise convolutional neural network.A novel Temperature Estimation based on a step-wise convolutional neural Networks(TENet)using combined satellite observations and in-situ measurements was designed for the sparse observation stations in local areas under complex terrain.The model was realized in two steps utilizing thermal infrared observation of geostationary satellites and topographic data,and the hourly SAT with the spatial resolution of0.01° under clear-sky across southwest China was retrieved.In the first step,the relationship between the collocated satellite observations and the measured surface air temperature from limited surface meteorological stations was modeled.In the second step,the surface air temperature derived from other satellite observations nearby stations without measured surface air temperature was estimated by applying the achieved relationship and then employed to augment the training set for further refining the model.Comparing with the in-situ measurements from meteorological stations,the mean RMSE of the model across the whole study area is 1.61?,and in the Qinghai-Tibet Plateau is less than 2?.Compared with two commonly used temperature products in different time scales and geographical characteristics,the model can achieve better or similar performance,which verifies that the proposed method can effectively improve the retrieval accuracy.