Research On The Prediction Method Of Soil Moisture In Wheat Field Based On Deep Learning

Soil moisture is one of the important factors affecting the yield and quality of wheat in the field.Water shortage and drought will lead to the reduction of wheat yield,while the traditional flooding irrigation method in the field has poor uniformity and serious waste of water resources.Therefore,it is of great significance to develop soil moisture prediction for the fine management of wheat fields and the saving of agricultural irrigation water.Because soil moisture is a kind of data with time series nonlinear,although the traditional prediction method has good mechanism interpretation ability,it has some shortcomings in application,such as requiring a lot of professional knowledge in the field,difficult parameter acquisition and relatively low prediction accuracy.With the development of machine learning and deep learning technologies in the field of big data,researchers begin to apply these technologies to soil moisture prediction in order to improve the prediction accuracy and timeliness.Based on the monitoring data of agricultural Internet of Things,this paper constructs a multi-factor soil moisture prediction model based on two-stage attention mechanism from two aspects of input factor characteristic weight distribution and time series characteristic weight distribution,aiming at the soil moisture prediction at field scale and combining the deep learning model and the attention mechanism,so as to improve the soil moisture prediction accuracy.The main research contents of this paper are as follows:(1)Processing and analysis of multi-source soil moisture data.There are missing values and outliers in the original data,which will have a negative impact on the stability and stability of the data,resulting in deviation or error of the prediction results.Based on the soil moisture,soil temperature and meteorological data such as solar radiation and precipitation obtained by agricultural Internet of Things equipment,this study firstly eliminated the outliers according to the Laida criterion,then used the mean replacement method to supplement the missing values,and finally analyzed and processed the data through the Pearson correlation coefficient method to obtain the correlation degree of different influencing factors and soil moisture.Moisturedata-wheat soil moisture data set was constructed with 10 kinds of soil and meteorological data being selected as characteristic data.(2)Multi-factor soil moisture prediction based on LSTM.In order to solve the problem of dependence on long time series of important information in soil moisture prediction,the multi-factor soil moisture prediction model LSTm-soil was constructed based on LSTM and adopted two-layer network structure.In order to verify the validity of the model,the method of SVR,ELM,BP,RNN and GRU was used to construct soil moisture prediction model,and the experiment was compared.The results showed that the root mean square error(RMSE),mean square error(MSE),mean absolute error(MAE),and coefficient of determination(R2)of LSTM-soil model were 0.5238,0.2744,0.4388,and 0.9434,respectively,which were better than the comparison model.Among them,three deep learning recursive neural network models(RNN,LSTM,GRU)had better prediction effect than BP,SVR and ELM models,and LSTM-soil neural network model had better prediction effect than GRU and RNN model.(3)Multi-factor soil moisture prediction based on two-stage attention mechanism.In order to capture the attention of input characteristics and time series,the problem of vertical and horizontal distribution of soil moisture distribution is solved.Based on the above research results,this study used LSTM network to build encod-decoder structure,and introduced the characteristics and temporal attention mechanism to further study the correlation between input characteristics and prediction targets,and constructed a soil moisture prediction model DA-LSTM-soil with two-stage attention mechanism.By comparing with LSTM-soil and CNN-LSTM models,the experimental results show that RMSE,MSE,MAE and R2 of DALSTM-soil model are 0.1764,0.0311,0.0466 and 0.9938,respectively,and the prediction accuracy is higher.In order to prove the generalization of DA-LSTM-soil model prediction,experiments were carried out on four groups of soil moisture data sets of different soil types in different wheat growing areas of Henan Province.The prediction effect of the model was stable,and the R2 of the model all reached more than 0.95,which verified the good generalization ability of the model.In this paper,the deep learning model is used to predict the soil moisture in the wheat field,and the DA-LSTM-soil model based on the two-stage attention mechanism is constructed,which can improve the short-term prediction accuracy of soil moisture in the wheat field scale,and then provide technical support for the realization of water-saving irrigation.