Soil Temperature And Moisture Prediction Based On Multi-Structure Attention Mechanism

Accurate prediction of soil temperature and moisture helps to predict and understand changes in ecosystems and plays an important role in Earth system science.However,the complex interactions between soil and atmosphere in both spatial and temporal dimensions make accurate prediction of soil temperature and moisture a huge challenge.Although deep learning methods are widely used in the field of prediction of land surface variables(e.g.soil temperature and humidity),the method’s over-parameterisation and uninterpretable training principles make its prediction performance somewhat limited.This paper thus targets optimising the prediction performance of deep learning by fusing three different attention mechanisms,while introducing and visualising an attention mechanism to bring interpretability to the model,and our main work is as follows.An Attention-aware Long Short-Term Memory Model(ILSTM＿Soil)is proposed,which incorporates a multi-feature attention mechanism,a predictor attention mechanism and a temporal attention mechanism.The Long Short-Term Memory(LSTM)network is used to generate multi-feature vectors of all predictors,and then each of the three attention mechanisms is introduced at appropriate locations and visualized to bring interpretability to the model.In this study,meteorological data from ten FLUXNET sites spread across the globe were randomly selected for testing.Inputs to the model include sky-scale historical soil moisture data,historical soil temperature data,long-wave radiation(LR),short-wave radiation(SR),air temperature(AT),Atmospheric pressure,AP),wind speed(WS),precipitation(P)and temporal variables(month information [M] and day information [D]).The model is labelled with future real soil moisture and soil temperature at a soil depth of 5 cm and a time scale of ’days’.To validate the performance of the proposed model,the evaluation metrics include the coefficient of fit(R2),root mean square error(RMSE),mean absolute error(MAE)and bias(bias).The experimental results show that the proposed ILSTM＿Soil model outperforms the Random Forest(RF),Support Vector Regression(SVR),Support Vector Regression(SVR),and the Doppler model in most cases when the prediction time is chosen to be 1 day and 7 days,respectively,by conducting prediction experiments on 10 randomly selected FLUXNET sites.Regression(SVR),Elastic-Net(ENET),LSTM(Attention-Long Short-Term Memory(A-LSTM)and LSTM with attention mechanism performed better than RF,Support Vector Regression(SVR),Elastic-Net(ENET),LSTM with attention mechanism in most cases,with errors reduced by 16.10%,22.89%,17.00%,17.56% and 14.65%,respectively.To further analyse the spatio-temporal properties of soil temperature and moisture,we explored the importance of different predictors,the importance of different time steps and the importance of different factors at different time steps by visualising the attention mechanism.The results show that historical soil moisture data,daily information,longwave radiation,precipitation and soil temperature are the five most important predictors in soil moisture prediction.In the temporal importance analysis,short-term data of temporal characteristics were highly utilised and the model focused more on recent temporal characteristics due to the memory nature of soil.In the analysis of the temporal characteristics of soil moisture,the results showed that soil moisture itself(all with weight values greater than 0.1)had a long(about 5 days)autocorrelation at most sites,and the closer to the target soil moisture,the more important the lagged soil moisture was for prediction.At the same time,at most of the occupied sites,the contribution of the predictors other than soil moisture itself to soil moisture prediction was almost the same at different time steps.