Prediction Of Soil Moisture Based On Residual And Spatio-temporal Deep Learning Model

In many fields such as ecosystem management and precision agriculture,soil moisture prediction can play an important role in many practical applications.Because of the great uncertainty of the long-term change of soil moisture,there are great difficulties and challenges in the prediction of soil moisture,especially in the long-term prediction.Therefore,This paper develops a new coder-decoder MCR-LSTM deep learning model with residual learning based on short-term and short-term memory and ECs-LSTM deep learning model that captures the relationship information of environmental variables,as an alternative tool of data intelligence to enhance the prediction effect of soil moisture.The MCR-LSTM model mainly includes two layers:the coder-decoder convolution LSTM layer and the LSTM layer with a fully connected layer.The residual learning is used to enhance the prediction ability by considering the intermediate time series data between the input time step and the prediction time step.ECs-LSTM is improved with convolution LSTM as its core,which is used to capture the relationship between the environmental variables of soil moisture,rainfall,soil temperature and other time series data to further improve the prediction effect.This paper uses FLUX NET site data to test MCR-LSTM for soil moisture prediction at 1,3,5,7 and 10 days in advance.The results show that based on the R~2 index of LSTM-based comparison model,the average improvement of MCR-LSTM is about 7.95%(1 day),10.10%(3 days),12.68%(5 days),15.49%(7days)and 19.71%(10 days);At the same time,ECs-LSTM model has achieved good results compared with the ordinary LSTM model with single environment variable and multiple environment variables.In addition,this paper also extensively discussed the predictability of soil moisture under different conditions,that is,the impact of adjusting different super parameters,different prediction models,different climatic regions and different locations on the results in MCR-LSTM,and the importance of different covariates in ECs-LSTM on the prediction results of soil moisture,in order to further understand the performance of the model.The specific research contents are as follows:Research on improving prediction performance based on codec structure.Encoder-decoder cyclic neural network consists of two recurrent neural networks(RNNs).In this paper,two LSTMs are used as encoders and decoders in the experiments of coder-decoder structure.The first encoding network converts the input sequence into a fixed-length internal representation vector,and the second decoding network takes this vector as the input to predict the output sequence.The connection between the two networks is a vector in the middle.The coder and decoder of the model jointly train to maximize the conditional probability of the given original sequence and the target sequence.The conditional probability of the combination of input and output sequences calculated by the coder-decoder is used as an additional feature of the existing log-linear model.At the same time,the depth of the deep learning prediction model network is effectively expanded,and the model performance is improved.Research on improving prediction effect based on residual learning.Generally speaking,the deeper the neural network is,the more difficult it is to train.Because of the uncertainty and complexity of future data,long-term prediction of time series data is more difficult than short-term prediction.The depth of the deep learning model network is crucial to the performance of the model.When the number of network layers is increased,the network can extract more complex feature patterns,so when the model is deeper,better results can be achieved theoretically,But in fact,this is not the case.We propose a codec model with residual learning to refine the training of the network,solve the long-term prediction accuracy of the deep learning model and network degradation,and further improve the prediction model effect.Research on ECs-LSTM Capturing Dynamic Environment Covariate Relationships Based on Convolutional LSTM.Soil moisture prediction is expressed as a spatio-temporal series prediction problem.By considering multiple dynamic covariable inputs and the relationship between multiple environmental variables,the multidimensional input of the model can be understood as a laterally extended spatio-temporal series from the perspective of data horizontal relations.By adjusting the structure of fully connected LSTM(FC-LSTM)cells,the convolution algorithm is applied in both input-to-state transition and state-to-state transition.On this basis,we propose the ECs-LSTM deep learning network model and establish an end-to-end trainable model for soil moisture prediction.The experimental results show that,Our ECs-LSTM deep learning neural network can better capture the temporal and spatial correlation information of dynamic environmental covariates such as soil moisture,rainfall,and soil temperature.Study on the importance of covariates in soil moisture prediction.Each prediction variable has an impact on the model generated by the machine learning algorithm.The significance of the statistical effect can be measured by the importance of the variable.Calculate the importance of variables in different LSTM-based models in the form of RMSE increase,and determine the importance of variables in the form of percentage.On the other hand,the importance of variables can also be used as the relative contribution of prediction variables to soil moisture.The average absolute error loss(MAE)can represent the impact of the prediction variable in the generated model and show the extent of the impact of the environment variable if it is not considered in the prediction model.