The Forecast Model Study Of Soil Water Content In Semi-arid Region Of The Loess Plateau

The bad regional climate and human activity for a long time to the destruction of thevegetation in Loess hilly semi-arid areas lead to serious imbalance of the ecologicalenvironment, in order to gradually restore vegetation, the local people did large-scaleafforestation from the sixties and seventies of the last century.Caragana is one of the mainafforestation tree species in local, it mainly depends on soil moisture, which is the commonlaw of semi-arid area plants.Artificial caragana woodland soil moisture mainly rely onrainfall, due to the rarely local average annual rainfall and the uneven distribution, how toreasonably use the limited water resources, to build vegetation and maintaining healthyecosystem and to realize the harmonious development of man and nature has been becomingone of the central task of the current vegetation restoration, so predict the grassland soilmoisture on prevention and control of soil drought and sustainable utilization of soil waterresources is of great significance.Nonlinear characteristic of the soil water content made soil moisture forecast harder, sothis paper apply neural network and time series auto-regressive model to research and forecastcaragana soil moisture in loess hilly semi-arid region, the main conclusions are as follows:1.The rainfall and planting density influence soil water contentThe intensity of the rainfall influence on soil moisture from surface to the deep is fadingafter survey and compare the soil water content, in which the5cm,20cm and40cm beenaffected serious than others. In the non-growing season the order of water storage capacity ineach density plot is contrast plot>1.0plot>0.5plot>1.5plot>2.0plot, which showed thatplanting density have contain degree influence on soil water storage capacity.2.Neural network modeling has high forecast accuracy in the soil moisture predictionUsing NARX recursive neural network model to research5cm,20cm,100cm,200cmsoil layers’soil water content in1.5plot,1.0plot and contrast plot. The change trend of soilmoisture content in three plots are as follows: soil moisture content in5cm layer decreasingfrom January to may, the trend from June to November are like parabolic and peaked inAugust, soil moisture content in20cm layer decreasing from January to may, the trend fromJune to December are like parabolic and peaked in September，100cm and200cm soil layers’ soil moisture content change less volatile, which showed that soil moisture fluctuationsincreased with the increase of soil depth.The fitting results of training set are as follows: theR2value were more than0.91and root mean square errors were less than0.91in three district,which showed that the model has good learning ability and can well fit the soil moisture datain the training set. The verification results are as follows: the root mean square error of thepredicted values and measured values in three plots ars less than0.45, which shows theprediction precision is higher. The generalizing results are as follows: R2values are more than0.92and the root mean square errors ars less than0.77in three plots, that showed the modelhas a good generalization performance.It can be seen from the results that NARX recursiveneural network has good learning ability for target layers’soil moisture content in each area,and that the results of the test set expressed the higher prediction precision and the smalleraverage relative error, which showed that the model was able to well predict soil water withvarious frequency fluctuations and with different density Caragana woodland in semi-aridLoess Hilly region.3. The time series auto-regressive model has low error in soil moisture forecastUse temporal sequence auto-regression model to study the soil moisture changing rule of1.5plot. the maximum likelihood estimate method were used to solve the parameters andAIC criterion to determine the delay order, the results of Chi-square test is bigger thanQ*,which show that three models were good. At last, through example validation all therelative errors were less than10%,which showed that temporal sequence auto-regression canwell predict soil water content of Caragana korshinskii shrub land, and can provide basis forthe regulation of the relationship between plant growth and soil water and vegetationconstruction in semi-arid Loess Hilly.