The Study On Machine Learning Based Methods For High Resolution Soil Moisture Retrieval

Soil moisture(SM)affects many environmental,climatic,ecological and economical processes.Higher spatiotemporal resolution in SM observation benefits more precise agriculture and drainage basin managements,timelier monitoring and warning on landslides and debris flows.Satellite based SM remote sensing is economical,convenient and macroscopic,but cannot achieve high timeliness when preserving spatial details.To fulfill spatial and temporal high resolution simultaneously,multi-source and heterogeneous observation data must be fused.Before conducting SM observation,sensors must be properly chosen to fit tasks'requirements.However,large amount and capability variety of SM remote sensors hamper making such choices.After multispectral SM observation,temporal gaps resulted from cloudy weather and satellite revisit pattern must be filled to fulfill high resolution requirements.Remote sensing data can be fused with in-situ SM data to make this happen.However,vast difference between spatial scales of both means leads to related but unequal SM data.Without in-situ observation,to recover multispectral SM observation contaminated by cloudy and foggy weather,SAR remote sensing data could be fused to take advantage of microwave penetrating clouds.However,different SM observation principles must be coordinated and unified.Consequently,this dissertation casts research on three problems:1)SM remote sensors' capability assessment and sensors clustering;2)soil moisture observation data fusing model between pixel scale and point scale;3)SAR and multispectral imager SM observation data fusing model.A clustering algorithm and two fusing model are proposed based on machine learning(ML)algorithms together with artificial neural networks(ANNs),in order to achieve three targets:1)to evaluate and cluster SM remote sensors considering potential SM observing capabilities;2)to improve optical SM remote sensing recovery quality by inventing better remote sensing and in-situ SM data fusing model;3)to fill temporal gaps between multispectral SM observations by fusing multispectral and SAR imager data.Aiming on these three problems,basic ML algorithm types and their capabilities,together with features of popular ANNs and their advantages have been briefed and introduced in Chapter 2.Conclusions are drawn that the aforementioned three problems can be regarded as a clustering problem and two regression problems,and can therefore be solved with ANNrs trained by ML algorithms.SM satellite sensors are of great amount,various clusters,and heterogeneous features.State-of-the-art sensor capability assessing methods are inadequate to uniformly and quantitatively assess remote sensors' potential observing capability in nonspecific SM observation tasks,neither can they further cluster SM remote sensors with respect to this capability.In Chapter 3,capability requirements on spaceborme remote sensing imagers by SM observation tasks have been analyzed,sensors' capability nominal parameters with respect to those requirements have been selected,and simulation experiments from which to extract sensors' capability simulation parameters have been designed.Thereafter,a sensor SM potential observing capability assessing method based on principal component analysis(PCA)has been designed to quantitativelysynthesize these two kinds of parameters.A sensors clustering method has finally been proposed,using self-organizing neural network trained by unsupervised ML upon the above quantitative assessment.In comparison with present sensor capability assessment methods,this method can uniformly quantify nonspecific SM potential observing capability of various spaceborne remote sensing imagers without artificial influences.It can also cluster remote sensors into groups corresponding to different types of SM observing tasks with respect to the above capabilities,thus can benefit selecting proper sensors for SM observation tasks in sensor planning phase.Fusing spaceborne remotely sensed and in-situ SM data can recover SM observation when nultispectral remote sensing is unavailable.State-of-the-art method use linear model to represent relationship between in-situ observed SM and corresponding pixel value on remote sensing SM images.However,their enormous scale discrepancy in combination with comprehensive environmental influences results in incompleteness to linearly model this relationship,and can lead to accumulated errors in SM recovery results.In Chapter 4,an improved data fusing model has been proposed,utilizing feedforward neural network in replacement with linear models to project in-situ SM reading toward remote sensing SM observation.Moreover,a data fusing method has been designed based on supervised ML to train the above network into fusing model,with the help of observation archive in the to-be-observed area.Compared to the present method,modified fusing model has improved SM recovery accuracy and precision.Retrieving SM with S AR data can meet the shortage of multispectral imagers in cloudy weather.Present SAR SM retrieval models are hard to apply in nonspecific area in lack of surface roughness and soil type information.To conquer the challenges by various environmental influences on SAR SM retrieval,the mechanisms,physical models and influencing factors of SAR SM retrieval have been elaborated in Chapter 5.A SM retrieval model from polarized SAR backscattering coefficients based on cascade forward neural network has been designed,assimilating five environmental as well as temporal variables for input.Moreover,a data fusion method has been proposed to train this network into SM retrieval model,fusing SAR backscattering coefficients,environmental and temporal variables as well as multispectral remotely sensed SM as training data.The retrieval model therefore synthesizes other retrieval-affecting environmental factors' influences,and can thus inverse SM from SAR backscattering coefficients under specific assumptions.In comparison with present SAR SM retrieval models,the proposed method has got rid of dependence on surface roughness input,and has wider applicability than empirical and semi-empirical models.To verify the usability and dependency of the three aforementioned innovations,a comprehensive experiment was designed,conducted and examined on a specificly chosen area,and has been detailed in Chapter 6.This experiment has taken advantage of the three aforementioned innovations in sequence.In this experiment,multispectral and SAR imagers have been chosen based on sensor clustering result.Thereafter,observation gaps of optical SM observation on SAR observation epoch have been recovered.Then,optical,SAR,environmental and temporal data and variables have been fused to build SAR SM retrieval models for 6 clusters of area.Finally,SM values have been retrieved using these models.Results of this experiment were examined,clarifying that the three innovated algorithms in this dissertation are capable,reproduceable,accurate and precise.In the future,the aforementioned three aspects have potential to be further researched.Sensor clustering method based on observing capability could be concatenated with observation resource database,automatic sensor coverage simulation environment could be developed.Thereafter,sensor capability parameters could be automatically acquired,and efficiency of such method could be enhanced.The method fusing satellite and in-situ SM observation could utilize generative model,in order to project in-situ observation toward more locations on the SM remote sensing image,so as to improve recovery accuracy.The fusing method between SAR and multispectral SM observation could utilize transfer learning mechanism to build fusing models,which could be applied to various locations and sensors,so as to improve model portability.The major innovations of this dissertation form 3 aspects:1)A satellite imaging sensor SM potential observing capability quantifying and sensor clustering method based on self-organizing neural network has been proposed,upon analyzing SM observation requirements on spaceborne remote sensors.This method can evaluate remote sensors potential capability and cluster sensors with this capability before and aiming at nonspecific SM observation tasks.2)A fusing model between spaceborne remotely sensed and in-situ SM observation data based on feedforward neural network has been proposed,upon analyzing the reasons to SM observation difference between vast scales.This model can fuse such two types of SM data at higher precision,and has improved the precision and accuracy of a state-of-the-art method.3)A multispectral SM observation reconstruction model has been proposed based on cascade forward neural network,upon elaborating SM retrieval physical models from soil microwave backscattering coefficients.By fusing SAR backscattering coefficients and environmental as well as temporal variables,such network can compute regional SM,thus can meet the shortage of multispectral SM observation in cloudy weather.Nevertheless,some aspects in this dissertation could still be improved,including degree of automation in clustering remote sensors,strict environmental assumptions on fusing models and portability of retrieving models.And these are among the future research focus.