Research On Wheat Spike Detection And Yield Prediction By Satellite Remote Sensing Based On Machine Learning

The parameters of wheat growth and wheat spikes information are important basis for yield composition.Accurate and high-speed monitoring of wheat growth and wheat spikes is of great significance to high-yield cultivation.With the rapid development of smart agriculture,deep learning technology and remote sensing technology are increasingly applied to the field of agricultural production.In order to re alize the remote sensing monitoring of wheat growth and the prediction of wheat yield,this study uses multi-source image data in the later stage of wheat growth to achieve rapid detection and counting of wheat spikes,which provides technical support for wheat yield prediction.Based on the main growth indicators of wheat in different growth periods,the yield sensitivity combination index was constructed,and the multi-spectral satellite data was used to quantitatively monitor the wheat yield by using ensemble learning algorithms.The main findings are as follows:(1)Using transfer learning and active learning strategies can greatly reduce the workload of wheat spike data labeling.The precision rate,recall rate,mean average precision,average detection time,and average counting accuracy of the YOLOv5 model on the validation set is 97.53%,96.48%,98.14%,0.241s,and 98.93%,respectively.Comparing the SSD network model,Faster R-CNN network model and Centernet network model,YOLOv5 performs best in each model,especially the detection time is fast,and it has the potential to quickly count wheat ears.It can provide technical support for the process of wheat remote sensing yield estimation.(2)By screening the optimal remote sensing variables for different wheat growth periods(joining stage,booting stage and flowering stage),constructed a quantitative remote sensing monitoring model for key wheat growth indicators(leaf area index,LAI,leaf nitrogen content,LNC,above ground biomass,AGB and chlorophyll relative content,SPAD value),and made a thematic map of the spatial distribution of main wheat growth indicators in central Jiangsu in 2020.The difference,product,and normalization were used to combine wheat structural parameters and nutritional parameters,respectively,to screen out yield-sensitive comprehensive index,and constructed a multiple remote sensing indexes linear regression quantitative monitoring model.The results showed that the ratio combination of LNC and LAI had the highest correlation(r=0.715),the RMSE(Root Mean Squard Error)and R2 on the verification set were 0.114 and 0.723,respectively.And make the 2020 thematic map of remote sensing inversion of GCI(growth comprehensive index)in central Jiangsu,which provides further basis for remote sensing of wheat yield.(3)The multi-model fusion strategy based on the Blending method can improve the accuracy of wheat yield prediction.The results show that the prediction accuracy of the Blending multi-model fusion algorithm is better than the single RF(Random Forest)model and XGBoost model,and the EVS(Explained Variance Score)was 0.794,the MAE(Mean Absolute Error)was 245.238,the RMSE was 295.274 and the R2 is 0.806.Compared with the RF model for wheat yield prediction,the accuracy was increased by 10.867%,compared with XGBoost prediction model,the accuracy was increased by 4.811%.The method of ensemble learning provides a more precise prediction of wheat yield,which can provide a new technical benchmark for precision agriculture.