Construction Of Soil Temperature And Moisture Prediction Model Based On Meteorological Data And Machine Learning

Soil temperature and moisture,as significant physical properties of soil,affect the distribution of nutrients and fertility,microbial activities,and the germination,development and growth of crops in soil.Through the prediction of soil temperature and soil moisture,crop irrigation can be adjusted precisely,so as to increase crop yield and income.But as a result of the sample data is missing value,abnormal values and extreme value,and the multicollinearity effect between variables,such as improper handling,lead to soil temperature and soil moisture prediction model has low reliability,accuracy is not high.In this paper,data preprocessing methods were used,soil temperature and soil moisture prediction model were established based on the dataset of Hailun Agricultural Ecological Experimental Station and Changbai Mountain Station,which improved the prediction accuracy of the model and broadened the method of filling missing samples.The soil temperature data of Hailun Station,soil temperature and moisture data of Changbai Mountain Station were preprocessed by data cleaning and interpolation.Pearson correlation was used to analyze the influencing factors of soil temperature and soil moisture.To reduce the influence of soil temperature and soil moisture extremes on the model prediction accuracy,this paper proposed to use K-means clustering and partial least squares（PLS）to cluster and reduce the dimension of meteorological data to improve the model prediction accuracy and convergence speed,and established a prediction model based on multi-layer perceptron（MLP）neural network.Four models,multiple linear regression（MLR）,MLP neural network,K-MLP combined with K-means clustering and MLP,K-P-MLP combined with K-means clustering,PLS and MLP,were compared to predict soil temperature and moisture respectively with meteorological data as input.Four evaluation indicatos,R²,RMSE,MAE and RRMSE,were used to analyze and evaluate the performance of the prediction model established in this paper,and the optimal models for predicting soil temperature and soil moisture in the study area were determined respectively.The results of soil temperature research are as follows:（1）Based on the analysis of soil temperature data at Hailun Station,the monthly average soil temperature changes in a sinusoidal curve.With the depth of soil layer,the maximum and minimum soil temperature have an obvious time lag,and the range of soil temperature change is smaller.The correlation coefficient between soil temperature within 20 cm depth and air temperature,surface temperature is above 0.9,which is highly positive.Soil temperature is negatively correlated with air pressure,and positively correlated with month,sunshine duration and other predictors,and the correlation coefficient between soil temperature and month increases with the increase of depth.Except for month and relative humidity,the correlation between soil temperature predictive factors and soil temperature decreased with increasing depth.（2）In this paper,four soil temperature prediction models based on the machine learning method,the R² of K-MLP soil temperature prediction model is the largest,above 0.93,and the RMSE,MAE and RRMSE are the smallest.In the performance of R²,RMSE,MAE,and RRMSE,K-P-MLP model is slightly worse than K-MLP model,followed by MLP model,MLR model has the worst accuracy.Among the seven depths of K-MLP model,RMSE range is 1.481-1.603℃and MAE range is 1.129-1.206℃,where the depth with the largest error is 10 cm and the depth with the smallest error is 15 cm.The convergence rate of the two hybrid models is compared,and the K-P-MLP model is faster than the K-MLP model.The results of soil moisture research are as follows:（1）Based on the analysis of soil moisture data of Changbai Mountain Station,the soil moisture of 5 cm and 20 cm depth is accompanied by heavy precipitation in two peaks around May and July each year.The average value of soil moisture increased with the increase of soil depth,and the variation range showed a decreasing trend.There is a positive correlation between soil moisture and air temperature relative humidity,evaporation,surface temperature and soil temperature,and there is a negative correlation between soil moisture and air pressure.Among the three soil depths,the correlation between soil moisture at 20 cm and the predictors are the strongest,followed by soil moisture at 50 cm and soil moisture at 5 cm.（2）The errors of the four soil moisture prediction models based on machine learning method are k-MLP hybrid model,K-P-MLP hybrid model,MLP model and MLR model in descending order.At depths of 5 cm,20 cm and 50 cm,the R² and RMSE of K-MLP model were 0.608 and 0.070 m³m^-3,0.728 and 0.065 m³ m^-3,0.679 and 0.031 m³ m^-3,respectively.The prediction model of 20 cm soil moisture was the best.In this paper,hybrid models of MLP based on clustering and dimensionality reduction preprocessing was proposed to predict soil temperature and moisture,which improved the accuracy and reliability of prediction,expanded the regional prediction means of soil temperature and soil moisture,and provided scientific theoretical basis for regional agricultural production regulation.