Research On Hyperspectral Estimation Of Rice Leaf SPAD And Nitrogen Content Based On Machine Learning

Chlorophyll and nitrogen are two very important indicators to characterize crop nutrition and growth status.Among them,chlorophyll is the most important pigment in plant photosynthesis and is one of the important reaction indicators for high-yield and high-quality growth of rice.Nitrogen is one of the essential nutrient elements for crop growth and development,and is the important limiting factor affecting crop growth and yield.Therefore,real-time and accurate monitoring of chlorophyll and nitrogen is of great significance to the accurate management of crop fertilizers in the field and the improvement of crop yield and quality.Spectroscopy technology can meet the requirements of real-time,rapid,non-destructive,and accurate diagnosis of crop leaf nitrogen content and SPAD,and timely grasp of crop growth and nitrogen status,providing a basis for nutritional diagnosis and fertilization regulation.The combination of hyperspectral technology and machine learning modeling analysis has greatly improved the accuracy of crop leaf nitrogen content and SPAD inversion,in order to accurately grasp the dynamic changes of leaf nitrogen content and SPAD,crop growth monitoring,yield estimation and quality prediction provide new means.In this study,rice cultivars Zhuliangyou 819 and Zhuliangyou 39 were taken as the research objects,and rice leaf hyperspectra were collected during the five key periods of early rice growth(tillering stage,booting stage,full heading stage,filling stage and maturity stage).Data and corresponding physiological and biochemical indicators(SPAD value and leaf nitrogen content).The study first explored the hyperspectral characteristics of rice leaves and their changing laws under different growth periods and different nitrogen application gradients,and systematically analyzed the response of the SPAD value and leaf nitrogen content of rice leaves to the growth period;The load value analysis of the wave band has determined the hyperspectral sensitive band of rice SPAD and nitrogen content;based on the determination of the hyperspectral sensitive band,Savitzky-Golay filtering,multi-scattering correction,standard normal transformation,and principal component analysis are adopted.4 kinds of pretreatment methodsPerform noise reduction filtering on the original spectral reflectance,and compare the effect of the four pre-processing methods and the model without pre-processing;use the sensitive spectral band as the input,and the corresponding actual physiological and biochemical index value as the output,respectively Three different algorithms: Partial Least Squares Regression,Random Forest,and Support Vector Regression were used to construct rice leaf nitrogen content and SPAD inversion models.The coefficient of determination and root mean square error were used as evaluation indicators to evaluate the accuracy of the model,and the best was selected.A combination of optimal pretreatment methods and modeling methods;finally,the study uses intertemporal data to construct an intertemporal prediction model of rice leaf nitrogen and SPAD based on hyperspectral data,and puts forward a “intertemporal prediction of crop physiological indicators” "Sex" is helpful for early prediction in the actual production process,and it also provides new ideas for the prediction of crop phenotypic traits and internal quality.The main conclusions are as follows:(1)Combine hyperspectral technology and machine learning technology to estimate the SPAD of rice leaves.From the results of the models constructed by three different algorithms,the estimation of rice leaf SPAD based on hyperspectral data is feasible,and the specific effect of the model is expressed as PLSR model>SVR model>RF model.Among them,the SPAD inversion model based on the PLSR algorithm has the best effect.Under different preprocessing,the R2 of the model has reached 0.64 or more,and the RMSE is controlled within 1.88.From the results of different preprocessing,the accuracy of the PLSR model based on Savitzky-Golay processing is the best,with R2 of 0.89 and RMSE of 1.02.(2)Combining hyperspectral technology and machine learning technology to estimate the nitrogen content of rice leaves.The research results show that the rice leaf spectrum has a high accuracy in estimating the nitrogen content of rice leaves.The inversion models of rice leaf nitrogen content constructed based on PLSR,RF,and SVR algorithms all have good fit.Among them,it is based on Savitzky-Golay.The PLSR model has the best estimation effect,and its model R2 and RMSE are 0.88 and 0.10,respectively.(3)The leaf nitrogen content and SPAD of rice were respectively predicted across the period.The results showed that the SPAD showed the potential for intertemporal prediction,while the leaf nitrogen content was not predictable across the period.Among the inter-period prediction models of rice leaf SPAD,the inter-mature prediction model has the best effect,with R2 reaching 0.74 and RMSE of 1.77.However,the fitting effect of the inter-temporal prediction model for rice leaf nitrogen content in each period is very poor,and the leaf nitrogen content cannot be predicted inter-temporal.