Research On Related Technologies Of Soil Temperature And Soil Moisture Prediction Based On Deep Learning

Soil temperature and moisture are important variables in many fields such as geoscience.Soil temperature,as the result of the combined action of the atmosphere and the land surface hydrothermal cycle,has a significant impact on agriculture,climate,environment and so on.And it can provide a decision-making basis in many different aspects such as improving agricultural production or slowing global warming.Meanwhile,soil moisture is a key land surface variable,which links the cyclic process between land surface and atmosphere.The spatio-temporal variation of soil moisture is crucial for meteorology,climate,hydrology and so on.And it can improve the understanding of water,energy and carbon cycles,as well as the prediction of extreme climates.Therefore,accurate prediction of soil temperature and moisture is very important for many fields such as geoscience.In the past few decades,researchers usually used physical process model to predict soil temperature and moisture.However,this method has some limitations due to the uncertainty of soil properties,the inaccuracy of input data of meteorological field,the imperfect expression of physical process and the error of parameterization scheme.With the rapid development of computer technology and the continuous outdating of hardware devices,the deep learning method shows strong nonlinear learning ability.And land-atmosphere interactive patterns can be simulated by datadriven approach to achieve accurate the prediction of soil temperature and moisture.Therefore,the deep learning method has received extensive attention from researchers in geoscience and other fields.At the same time,it provides a new perspective for the prediction of the land surface variables as a supplement to the method based on physical process model.For the prediction tasks of soil temperature and moisture,how to use the deep learning method to improve the autocorrelation and uncertainty of explanatory variables(the input variables of the model)in long time series and how to use physical knowledge to guide deep learning and improve the robustness of the model performance have become the key points and difficulty for researchers.In view of the above problems,this paper uses deep learning method to research the prediction of soil temperature and moisture and proposes corresponding solutions.The main research contents of this paper are summarized as follows:1.Aiming at the scarcity of meteorological station,it is difficult to obtain relevant meteorological data.And the autocorrelation of explanatory variables in long time series of the recurrent neural networks,this paper proposed the prediction model based on multi-channel gated recurrent units for soil temperature.This method mainly solves the problem that the autocorrelation of explanatory variables in long time series gradually decreases in the process of information transmission.It causes the model to fail to accurately establish the changes in the relationship between the prediction values of soil temperature and longer-term explanatory variables.Gated recurrent units is used as the basic module of the network to extract the features of the shortterm patterns of the explanatory variables.Multiple GRU models are defined as auxiliary networks,and the timestep of each GRU model is gradually reduced to capture the features of the longer-term patterns of explanatory variable.It can avoid model error caused by information transmission for longer-term explanatory variables.The experimental results show that the method can effectively improve the accuracy of the prediction of soil temperature when the relevant meteorological data is difficult to obtain.2.Aiming at using physical knowledge to guide deep learning,a prediction model of soil temperature based on quadruplet loss function was proposed in this paper.The traditional recurrent network model takes explanatory variables as input directly and uses common loss functions such as mean-square error to train the model,which makes the prediction model unable to effectively extract the features of soil temperature.And it leads to a physical inconsistency in the prediction results.The method combines loss function with distance metrics learning between the features of soil temperature and guides the prediction model for training,which makes the prediction results conform to a certain land-atmosphere interactive pattern as much as possible.The data are divided into different intervals and differentiated with different labels by clustering.By shortening the distance of the features of soil temperature in the same interval and pushing the distance of the features of soil temperature in different intervals,it can avoid that the prediction results of traditional recurrent neural network models do not have physical consistency.The experimental results show that the method has a good prediction performance in the task of soil temperature prediction dealing with multiple explanatory variables.3.Aiming at the prediction of soil moisture based on deep learning,compared with the prediction of soil temperature,soil moisture has higher heterogeneity due to various factors such as soil properties,precipitation and vegetation.The increase of prediction scale leads to enhance the uncertainty of future information.The prediction model of soil moisture based on encoder-decoder Long Short-Term Memory is proposed in this paper to solve the problem.The encoder of this method uses Long Short-Term Memory as the basic network,which is used to encode the features within the prediction scale and enhance the feature extraction ability of the model.Then by decoding the future information and making multiple soil moisture in the prediction scale participate in the calculation of the gradient of the loss function,the error caused by the uncertainty of the future information can be improved.The explanatory variables are combined with the decoding information of the decoder to correct the error which is brought by the encoder-decoder LSTM model.The experimental results show that the method can enhance the accuracy of the prediction of soil moisture by improving the uncertainty of the future information.4.Aiming at predicting the soil moisture of SMAP(Soil Moisture Active and Passive,SMAP)product,too little total amount of data tends to cause the problem of underfitting in deep learning.This paper proposes a spatio-temporal model of soil moisture prediction integrated with transfer learning.The method uses threedimensional convolutional layer to extract the spatial features of soil moisture.Then the results are passed to the LSTM model to extract temporal features.The method can improve the prediction accuracy of traditional deep learning models.Adopting transfer learning strategy and using the ERA5-Land dataset to guide the parameters initialization for deep learning model,which can avoid the prediction model overfitting.The experimental results show that the method can effectively extract the spatio-temporal features of soil moisture.After the ERA5-Land dataset to guide the parameters initialization for deep learning model,the prediction performance of SMAP soil moisture based on deep learning model is improved.