Prediction Of Soil Moisture Based On Multi-source Data And Machine Learning Algorithm

Drought is one of the most important natural disaster in agricultural production,especially in the arid and semi-arid areas of Northwest China.By analyzing and predicting the rule of soil moisture's changes,finding practical prediction methods of soil moisture and improving the level of soil moisture's prediction are not only helpful to making full use of climatic resources,predicting the drought degree of crops,guiding agricultural irrigation,but also good for drought prevention,and providing the scientific basis of loss assessing,and so on.The main winter wheat planting areas in Baoji of Shaanxi Province were taken as research areas to study the characteristics of soil moisture's distribution and its prediction methods.Based on the measured data of soil moisture,meteorological data and soil bulk density,diving depth,plough layer thickness and the proportion of soil gravel,to set up the database of soil moisture and predictive factors.Random forest algorithm,least squares support vector machine algorithm and radial basis function neural network algorithm in machine learning were used to predict soil moisture.Combine with the predictive factors and machine learning algorithm,the corresponding prediction model were established to simulate and predict the soil moisture in the study area,and the accuracy of different prediction models were evaluated by quantitative methods.According to the prediction results of the optimal prediction model,the spatial and temporal distribution characteristics of soil moisture and soil drought in the study area were analyzed.Finally,the importance of predictive factors was quantitatively analyzed and ranked in order to analyze the influence of each prediction factor on soil moisture and its changes.The results of this study are expected to provide more scientific theoretical methods and technical support for soil moisture prediction,crop production management and drought prevention.The results are as follows:(1)Applying machine learning algorithm to soil moisture prediction,exploring the applicability of machine learning prediction model in the field of soil moisture research,and providing a new exploration scheme for soil moisture's spatial prediction research;The results show that the prediction accuracy of the three predictive models has reached more than 88%,which proves that machine learning algorithm has a good application in soil moisture modeling and prediction.And among the three models,the prediction accuracy of RBF neural network prediction model is higher than that of LS-SVM prediction model and random forest prediction model,which is most in line with the actual situation of soil moisture in winter wheat planting area of Baoji.Among the three soil moisture prediction models,the RBF neural network prediction model and LS-SVM prediction model are slightly better than 20~40 cm soil layer in 0~20 cm soil layer,while the RF prediction model is slightly better than 0-20 cm soil layer in 20~40 cm soil layer.(2)The main characteristics of spatial and temporal distribution of soil moisture in winter wheat planting area of Baoji from March to May in the year of 2014 to 2018 are that the annual variation of soil moisture is small and the monthly variation is large.In the annual scale,the amplitude of soil moisture's variation in the past five years is small,among them,the soil moisture content in the year of 2015 and 2017 is slightly higher than that in 2016 and 2018;the relative soil water content in the 20~40 cm soil layer is slightly higher than that in the 0~20 cm soil layer,while the amplitude of soil moisture's variation in 0~20 cm soil layer is smaller than that in 20~40 cm soil layer.In the time series of monthly scale,the change range of soil moisture is larger from March to May,and the relative soil moisture in March and April is lower than that in May.The amplitude of soil moisture's variation in 20~40 cm soil layer with time is generally less than that in 0~20 cm soil layer,and the soil moisture content in 20~40 cm soil layer is slightly higher than that in 0~20 cm soil layer as a whole.Analysis from the scale of spatial distribution,the annual variation of soil moisture in March is small,and the high value of soil moisture mainly distributes in Guanshan low-middle mountain area in the northwest corner of the study area and Qinling low-middle mountain area in the south of the study area,and the low value mainly distributes in Weibei loess plateau area and Weihe Valley terrace along the Qianhe River and Weihe River basin from northwest to southeast area,Weinan loess plateau area and Weibei alluvial fan area in Weinan;the annual variation of soil moisture in April is large,and the spatial distribution characteristics of soil moisture are similar in the year of 2014 and 2015.The high value mainly distributes in Qianshan area and Qianhe River basin in Linyou County,while the low value mainly distributes in the northern foot of Qinling Mountains in the southwest of the study area,and the spatial distribution of soil moisture is similar from the year of 2016 to 2018.The high values are mainly distributed in the northern foot of Qinling Mountains and the low-middle mountain areas of Guanshan;the annual variation of soil moisture is small in May,and the high values are mainly distributed in the hilly areas of the northern foot of Qinling Mountains in the southwest direction,the Qianshan Mountains in the northeast direction and the Guanshan Mountains in the Northwest direction.(3))From March to May of 2014 to 2018,the drought of Winter Wheat in the winter wheat planting area of Baoji occurred in a larger range and lasted for a longer time,but the drought degree of winter wheat was generally lighter and the drought level was generally mild.And winter wheat drought had little difference between years and big difference between months.In the annual scale,the drought in the year of 2015 and 2017 is relatively light,and in the rest of the year is relatively heavy.In the monthly scale,droughts occur in different degrees and ranges in all three months.In March and April,the drought range is larger and the drought degree is lighter,while in May,the drought range is smaller but the drought degree is heavier.And there is no significant difference between the whole month of drought.In terms of spatial distribution,winter wheat drought occurs frequently and severely in Qianhe Valley terrace area in Longxian County,Weihe Valley terrace area in Fengxiang County,Qishan County and Fufeng County,and loess plateau area in Weibei County.(4)In the winter wheat planting area of Baoji from March to May in the year of 2014 to 2018,compared with topographic factors,soil attributes and geographical location factors,meteorological factors have a greater impact on soil moisture,especially the precipitation,temperature and sunshine in meteorological factors;with the change of time,elevation,wind speed and precipitation have a greater impact on soil moisture in 0~20 cm soil layer.The influence decreases month by month,while the influence of topographic humidity index factor increases month by month;in the 20~40 cm soil layer,with the change of time,the influence of latitude and precipitation on soil moisture decreases month by month,while the influence of altitude,vapor pressure,relative humidity and maximum wind speed increases month by month.Generally speaking,the influence of each predictor on soil moisture's varies not only with the change of soil depth,but also with the change of time.Therefore,it is particularly important to select the predictor of soil moisture according to time and place for the accuracy of soil moisture prediction.