Research On Prediction Method Of Grassland Yield Based On Deep Learning

With the rapid development of society,due to the lack of scientific and reasonable development planning,the ecological environment in pastoral areas has been deteriorating,and then the development of animal husbandry faces great challenges.Effective monitoring of grassland productivity and improvement of grassland ecosystem restoration ability are important measures to realize a virtuous cycle of ecological environment and sustainable development of animal husbandry.They are also important to accurating measurement of pasture yield.Under this research background,this thesis uses deep learning model to accurately predict normalized vegetation index(NDVI),with CASA model to estimate grassland net primary productivity(NPP),and inverting grassland yield with grassland NPP.Finally,the prediction of grassland yield can be realized.The main research contents are as follows:Firstly,a NDVI prediction model based on Bi LSTM-attention mechanism is proposed,which can effectively use the input multi-variable time series data.It also solves the limitations of NDVI data prediction,poor accuracy and weak applicability of the model.At the same time,According to in view of the existing NDVI model focusing on the influence of temperature and precipitation of prediction,we often ignore the influence of other meteorological factors.This thesis analyzes the correlation between NDVI and various meteorological factors,and finds that evaporation and sunshine have strong correlation with NDVI.By comparison,it is found that adding evaporation and sunshine can reduce the root mean square error of the model by about 20%,and improve the prediction accuracy of the model.The attention mechanism can not only process the long time series data and multidimensional feature data but also update the weight of feature values.It also makes the model focus more on the feature data with high weight,and reduces the calculation burden brought by multidimensional input.The experiments were compared with RMSE,MAE andR~2,and LSTM,GRU,Bi LSTM and LSTM-attention models.The results show that Bi LSTM-attention model improves 13%,10.3%and 33.8%,respectively,compared with LSTM-attention model.The deep learning prediction model can provide strong technical support for crop yield estimation,grass yield estimation and so on.Secondly,in the prediction of grassland yield,using CASA model to estimate grassland NPP,and the principle and structure of CASA model were introduced.In this method,NDVI data and CASA model are used to estimate the grassland NPP,and then the grassland yield is inverted according to the NPP data.Finally,the grassland yield could be predicted.The experimental results showed that the prediction accuracy of grassland yield reached 77.13%and has higher prediction accuracy.It also proved the stability and practicability of the model.This study is based on forecasting model combining the deep learning to build artificial neural network with remote sensing.It can accurately estimate the grass yield,effectively improve scientific breeding scheme and provides the basis for the macro management of desert grassland.At the same time,it is conducive to regional resources development and utilization,which can provide a scientific reference for the social and economic sustainable development decision-making.