Monitoring Of Wheat Growth Indicators At Different Scales And Yield Prediction Based On Multi-source Data Fusion

Growth indicator assessments are important for monitoring wheat growth at different stages of growth and predicting wheat yields.Conventional approaches to statistical modelling,in which feature information extracted from image data from different sensors is used to parse crop growth indicators,have yielded some results in this category.However,these data-driven empirical models are not robust in assessing models and have weak generalisation capabilities due to differences in crop varieties,planting years and planting locations.In addition,canopy structure differences,sensor type and image resolution lead to large differences in the accuracy of leaf-scale and canopy-scale models for assessing crop growth parameters,and the difficulty of converting leaf-canopy upscale models.Therefore,this study aims to obtain growth indicators at the organ scale and canopy scale accurately and efficiently.The dynamic changes and correlations between multi-source image features and growth indicators are investigated based on the two scales respectively,and growth indicator estimation models and wheat yield prediction models at different scales are constructed by combining machine learning algorithms.The research results can provide a theoretical basis and technical support for the accurate and efficient management of smart agriculture.The main findings of this paper are as follows:(1)The variability of organ-scale and canopy-scale growth indicators in relation to image feature parameters was analysed.Images acquired on UAV-based platforms provide richer feature information,but also increase the complexity of crop canopy features.The correlation between different image parameters and growth indicators varies with the period of fertility.The correlation between organ image features and growth indicators is stronger and more sensitive than at the canopy scale,but the effect of different planting years on growth indicator sensitivity parameters is also greater,while the canopy scale relatively weakens the variability in the relationship between image features and growth indicators between growing seasons.(2)A model was constructed to estimate organ-scale growth indicators in a field environment.Based on two near-ground sensors(digital camera and handheld thermal camera),statistical models and machine learning algorithms were used in combination with data fusion techniques to estimate the nitrogen content of wheat organs(leaf and spike nitrogen)under multiple treatments(including variety and nitrogen application).Information such as color,texture and temperature data were extracted from the acquired near-ground images and used to assess the organ nitrogen status of wheat at late reproductive stages.The results show that organ-scale multi-source data have good potential for monitoring the nitrogen nutrient status of wheat.A machine learning model incorporating multiple features significantly improved the accuracy of the wheat organ nitrogen content estimation model.The random forest(RF)model incorporating three types of feature parameters(color,texture and temperature data)has the best prediction for nitrogen content(test set:R2=0.95,RMSE=1.89 mg/g,rRMSE=8.15%;training set:R2=0.92,RMSE=3.25 mg/g,rRMSE=(11.90%).The model prediction results indicate that the organ-scale nitrogen estimation model developed in this paper is reliable and can provide a technical reference for monitoring the nitrogen status of crops in the field.(3)A canopy-scale growth index estimation model was developed.A dataset of color,texture,vegetation indices(VIs)and the temperature was extracted from the canopy scale,and a model for estimating N content and aboveground biomass(AGB)of wheat plants was developed using different combinations of features and machine learning algorithms.The results show that the canopy-scale multi-source image feature values have the potential to estimate wheat canopy-scale growth indicators,with better results for AGB.The RF model incorporating color,texture and temperature data was the best for estimating plant nitrogen content,with an R2 of 0.68,an RMSE of 3.03 mg/g and an rRMSE of 12.33%.Variable screening using principal component analysis(PCA)algorithms,combined with machine learning models,allowed accurate estimation of AGB in late wheat fertility.the best estimation model for aboveground biomass was the RF model developed based on texture,vegetation index and temperature datasets,evaluating a model with an R2 of 0.94 and an RMSE of 743.02 kg/ha and an rRMSE of 5.85%.The canopy-scale growth indicator estimation model for wheat based on multi-source data fusion developed in this study was able to estimate the plant nitrogen status and aboveground biomass of the wheat population more accurately.(4)A model for wheat yield prediction at the organ scale and canopy scale and a combination of the two scales was developed.Combining machine learning models and variable selection algorithms can significantly improve the accuracy of yield estimation models,and model updating strategies are optimized for wheat yield prediction models.The use of feature selection algorithms to filter feature variables and develop RF models based on organ-scale and canopy-scale datasets,combined with model updating strategies,can accurately estimate wheat yields.However,machine learning models caused more variation in yield prediction than variable selection algorithms.The RF model developed based on the organ dataset produced R2 of 0.71 and 0.80,RMSE of 832.56 kg/ha and 820.12 kg/ha and rRMSE of 14.25%and 12.58%for wheat yield assessment in 2020 and 2021,respectively.The yield prediction model developed based on the canopy dataset yielded R2 of 0.77 and 0.79,RMSE of 517.50 kg/ha and 846.32 kg/ha and rRMSE of 9.22%and 12.74%for the 2020 and 2021 yield assessments,respectively.The RF model combining two scaled growth indicators assessed wheat yields with R2 of 0.81 and 0.83,RMSE of 515.36 kg/ha and 828.17 kg/ha,and rRMSE of 8.62%and 13.61%,for 2020 and 2021,respectively.The accuracy of the wheat prediction model based on the canopy scale was slightly better than the organ scale,and the highest accuracy of the wheat yield estimation model was constructed by combining the growth indices of both scales.In addition,the filling period is the best period for wheat yield estimation.The feature selection algorithm and model updating strategy were used to improve the robustness of the wheat yield estimation model across cropping years,enabling the migration and updating of the wheat yield estimation model between years.The combination of the two scales extends the application scenario of single-scale data and the accuracy of the yield estimation model.The use of the newly developed machine learning model for predicting wheat yield provides technical support for the construction of large-scale precision management of farmland and efficient breeding.