Comparative Study On Prediction Models Of Field Soil Moisture

Soil moisture is field soil water and its corresponding crop water state.Beccuse soil moisture has significant influence on crop growth,people have been working hard to get real-time value of soil moisture,and then to guide crop farming,designate irrigation system,and develop intelligent agriculture.In this paper,taking Shunyi District,Yanqing and Daxing District as examples,5 different prediction models are used to test the prediction performance of the models under 3 forecasting periods,and to find the best soil moisture prediction model for the study areas.The specific contents include:First,an overview of the research area was introduced in this paper.The soil texture in Beijing is mainly sandy loam and light loam.Sandy loam is mainly distributed in Northeast Beijing,and light loam is mainly distributed in Southwest Beijing.The precipitation in recent 10 years showed a slowly increasing trend,and the its interannual fluctuation was very large.The annual distribution of rainfall was uneven,and showed a decreasing trend from northeast to southwest.The soil moisture in the last 5 years has been growing slowly,with the highest moisture in August and the lowest at the beginning of the year.The soil water in Shunyi,Miyun and Pinggu three areas was larger than that in Huairou.Secondly,soil temperature,air pressure,humidity,wind speed,ground temperature,rainfall and initial soil moisture value were chosen as the influence factors of soil moisture prediction.The correlation analysis shows that there was a significant correlation between those factors and soil moisture.Then,models were trained and calibrated by training and test data.The research adopted five models to predict soil moisture,which were linear regression model,BP neural network model,PCA-RBF neural network model,GEP model and deep learning model.Through the analysis and learning of the basic data,the prediction performance of the models under different parameters are compared.The best parameters which made models to get a certain ability to predict were selected to complete the calibration of models.Finally,Feasibility of the prediction of the five models was tested and the results showed that the minimum correlation coefficient between the predicted results and the measured data was 0.593,which indicated that the five prediction models were feasible.Correlation coefficient(R~2),average relative error(MRE),mean absolute error(MAE)and root mean square error(RMSE)were used to evaluate performance of those five models.The results showed that the prediction performance of the depth learning model was the best,with strong correlation between the model prediction data and the measured and small error,followed by GEP model,PCA-RBF neural network model and BP neural network model.Prediction performance of linear regression model was worst.In this study,prediction performance and characteristics of different models were compared and analyzed,which provided a theoretical basis for prediction of soil moisture.