Soil Temperature Prediction Methods Based On Machine Learning

Soil temperature is closely related to several fields and has important impacts on them,such as agriculture,hydrology and meteorology.Especially in agriculture,on the one hand,soil temperature affects germination of seeds,development of roots and physiological processes of crops.On the other hand,soil temperature affects the state of soil water,the activity of soil microorganisms and the transformation of soil organic matters,which can indirectly affect the growth of crops.Therefore,the research on soil temperature prediction is of great significance.An accurate soil temperature prediction model can effectively guide agricultural production and avoid the risks brought by extreme soil temperature in advance.Based on soil temperature data from three sites in Switzerland from the Fluxnet dataset,this paper investigates on site soil temperature prediction.Aiming at the prediction of site soil temperature sequences,this paper proposes a method named Convolutional Neural Network based on Ensemble Empirical Mode Decomposition.Representative algorithms in machine learning such as Back Propagation Neural Network,Long Short-Term Memory Neural Network,Long Short-Term Memory Neural Network Based on Ensemble Empirical Mode Decomposition,a model which has been used in study for soil surface temperature prediction,and a most basic prediction method Persistence Forecast,are selected as comparative methods.After analyzing and comparing the prediction results of soil temperature data at three stations and three depths at three kinds of time delays.Convolutional Neural Network based on Ensemble Empirical Mode Decomposition performs better on site soil temperature prediction.After the research on site soil temperature prediction,this paper further investigates on spatiotemporal soil temperature prediction in an area.Drawing on the improvement of prediction performance of machine learning models that combined with Ensemble Empirical Mode Decomposition on site soil temperature prediction research,a model called 3-Dimensional Convolutional Neural Network based on Ensemble Empirical Mode Decomposition is proposed.In order to evaluate the prediction performance of the model.Models applicable to spatiotemporal soil temperature data,including 2-Dimensional Convolutional Neural Network,3-Dimensional Convolutional Neural Network,Convolutional Long Short-Term Memory Neural Network,2-Dimensional Convolutional Neural Networks and Convolutional Long Short-Term Memory Neural Networks combined with Ensemble Empirical Mode Decomposition,and Persistence Forecast,are used as comparison models in experiments.After comparing and analyzing the results of soil temperature prediction under three kinds of time delays in the experimental area,the results show that the 3-Dimensional Convolutional Neural Network based on Ensemble Empirical Mode Decomposition has the best prediction performance.It can better obtain the correlation of spatiotemporal data and is a suitable method for spatiotemporal soil temperature prediction.