Research On Simulation Of Soil Water Content Based On Neural Network

Changes in soil moisture affect various aspects of local residents' lives and social production.It is a major field of agricultural informatization research to timely understand soil moisture information,and to predict soil moisture and adjust irrigation.This article takes Yanchi County as the research area,selects the four-year statistical data of the weather monitoring stations from 2014 to 2017,compares the prediction effect of the data under different optimization models and selects the best model to improve,and obtains a final improved model.The main research contents are as follows:(1)The methods of XGBOOST and random forest in integrated learning are used to screen the features.After comprehensively considering the screening results of the two models,the soil temperature,water pressure,and water vapor partial pressure are selected.,Rainfall,air temperature 5 factors as input characteristics.Using two models to predict and compare the input full-factor input model respectively,the experiment proves that the effective contribution of the selected factors is 91%,which shows the accuracy of feature selection.(2)Based on the factors selected above,using integrated learning,BP neural network,support vector machine and extreme learning machine to predict,it is found that: R2 of random forest and XGBOOST in integrated learning are 0.7313 and 0.6851 respectively,the effect of random forest is better than XGBOOST;the BP neural network model uses the activation function relu and is optimized by the ADAM algorithm.The R2 is 0.8219,which is higher than 0.5049 of the SGD-ptimized BP neural network;the kernel function selection in the support vector machine is Gaussian function model.R2 is 0.5371,high The kernel function selects 0.292 of the liner model;the extreme learning machine R2 with the activation function sigmoid among the extreme learning machines is 0.7623,which is lower than 0.8246 of the particle swarm optimized model R2.In summary,it is found that the ADAM optimized BP neural network and the particle swarm optimized extreme learning machine have the best effect.Because the extreme learning machine takes too long,the ADAM optimized BP neural network is selected for improvement.(3)For the BP neural network optimized by ADAM,it is easy to fall into the problem of local optimization and inaccurate prediction.The LM algorithm and Bayesian regularization algorithm are selected to optimize the BP neural network model from the optimization of weights and the optimization of the network structure.At the same time,considering the time series problem of soil water content data,the cascade mixed model of LSTM network and BP neural network is adopted,and the time feature processing is performed on this basis.The results show that the R2 of the three are 0.8332 and 0.8621,respectively.,0.8931,the results are better than the R2 optimized by the ADAM algorithm before the improvement,and the hidden layer nodes used are lower than the ADAM algorithm.The error loss graph also shows that the decline rate of the LSTM-BP model is higher than the previous two.In summary,the optimal model is LSTM-BP.