Research On Grain Yield Forecasting Based On Satellite Remote Sensing Images

Accurately forecasting grain yields has broad implications for economic trading,food production monitoring,and global food security.Grain production forecasting is the key to grain storage,farmland management,and national agricultural decision-making.It is an important part of national food security assessment and food policy formulation.A series of factors such as population increase,reduction of arable land and water resources,environmental degradation,and global warming have a significant impact on agricultural production and threaten food security.Therefore,accurate regional crop growth monitoring and yield forecasting are essential to guide agricultural production,ensure national food security,and maintain sustainable agricultural development.However,the variety of environmental variables across time and space presents challenges and has limited the precision of yield models.The traditional production estimation model is based on a physical model.Its disadvantage is that only representative points can be selected for estimation.The area is small and it is difficult to achieve large-scale production estimation.After research,there is a strong correlation between crop yield and its remote sensing image data,and crop yield prediction through satellite remote sensing data becomes possible.Here,this paper develops a deep learning-based prediction model,which uses a hybrid method of convolutional neural network CNN and long-term and short-term memory network LSTM for modeling.It uses the corn remote sensing data and yield from 2001 to 2016 on the American Corn Belt Data for model training,using 2017 and 2018 data to test the accuracy of the model.In this paper,by analyzing the importance of physiological characteristics during crop growth,it can be considered that extreme temperatures and the increase in annual average temperature may have a negative impact on corn yield.In addition,changes in some important variables have been identified,in particular early surface temperature,intermediate surface temperature and vegetation index.Finally,this paper compares several key remote sensing data including normalized vegetation index(NDVI),enhanced vegetation index(EVI),land surface temperature(LST)and soil moisture(Soil Moisture,SM).It is found that it is difficult to obtain a high-precision prediction model using a single index.Its prediction RMSE curve is basically above 15%.Through the combined use of vegetation index and environmental index,the prediction accuracy will be significantly improved.Finally,the combination of EVI,LST,and SM has the smallest prediction error,and its RMSE curve is between 4%and 5%.The prediction accuracy at this time has reached the level used for actual prediction.Traditional models are used to calculate and fit the input data through pure mathematical methods to achieve crop yield prediction.When these models are used,a big problem is the need to manually select features,feature selection and final model The accuracy has a great relationship,and the fitting method is mostly linear fitting,but the actual situation is mostly non-linear.In order to make up for these shortcomings,this article uses a deep learning method to build a prediction model.Through simulation and comparison,the prediction model proposed in this paper performs well,it has the lowest error in the most important production areas,and has very high prediction accuracy.In order to further improve the accuracy and versatility of the network,this paper makes a detailed analysis on the setting of hyperparameters,overfitting problems,and optimization algorithms,including the setting of the learning rate,the size of the convolution kernel,the number of layers in the network,and the number of neurons.Because CNN is affected by many factors when extracting features,it has high requirements on input data,and it is difficult to learn the time correlation between data.Therefore,the model of combining two networks can compensate for certain aspects of a single network.Insufficient,can achieve the best results in various situations.Finally,through comparative simulation with a single CNN network and LSTM network,The results show that the model designed in this paper has the highest prediction accuracy,and the prediction error RMSE curve can converge to about3%.With the optimization of the model structure,the quality of predictions is also improving.