Inversion Of Nitrogen And Chlorophyll Content In Crop Leaves Based On Hyperspectral And Machine Learning

Rapid,accurate and non-destructive monitoring of vegetation nutrient content is of great significance to the dynamic management of crop fertiliza-tion and growth monitoring.Chlorophyll is an important bioactive component of crop growth and development,and is closely related to light energy utiliza-tion,atmospheric carbon dioxide uptake,and photosynthesis.The level of chlorophyll content directly affects the photosynthetic efficiency and develop-ment of plants.Nitrogen is the main nutrient that promotes leaf growth and is also a key nutrient element involved in the formation of chlorophyll.Most of the existing studies use multi-spectral remote sensing data to construct various vegetation indices or a combination of 3-5 bands to estimate leaf nitrogen and chlorophyll content.Due to the limitation of the number of bands,it is difficult to improve the inversion accuracy.Hyperspectral remote sensing can obtain vegetation reflection information of more than 2,000 spectral bands,which can better characterize crop nutrient content and growth information,and is an important means to detect vegetation nitrogen and chlorophyll content.Some scholars use hyperspectral curve features to obtain crop nutrient content in-formation,such as the position of the red edge,etc.However,these models only use a small number of band information and do not give full play to the multi-band advantages of hyperspectral data.Therefore,this paper introduces a machine learning algorithm to screen the hyperspectral multi-band informa-tion,establishes an inversion model,and predicts the nitrogen and chlorophyll content of crop leaves with high accuracy.This study firstly discusses the effect of different dimensionality reduction methods for hyperspectral data,the purpose is to minimize the influence of multicollinearity among hyperspectral data,retain all the effective hyperspec-tral features as much as possible,and reduce data redundancy;secondly,the optimal dimensionality reduction algorithm is coupled to the machine learn-ing model,and the hyperspectral inversion model of winter wheat in each growth period is established,and the influence of multicollinearity on different machine learning models is also fully studied;finally,The optimal machine learning model is applied to the Stacking ensemble learning model,and the point scale inversion is extended to the regional scale.This paper has the following research contents and conclusions:（1）Using the elastic net algorithm to reduce the dimension can effectively retain the effective information of hyperspectral,and then combined with the partial least squares regression algorithm to fit,can realize the high-precision inversion of nitrogen content in maize leaves based on the sensitive band of hyperspectral data(R²_V=0.96,RMSE=0.17).（2）Using the gradient boosting regression tree algorithm,a nonlinear inversion model（abbreviated as GBRT）of leaf chlorophyll content and hyperspectral full-band of winter wheat at different growth stages was established.Since the ”max-features” parameter in the GBRT model can adjust the number of features of the input data to achieve the purpose of dimensionality reduction,the inversion effect of the GBRT model is better.The inversion accuracy at the mature stage is R²_V=0.95,RMSE=3.64.（3）Using the first-order differential method to convert the hyperspectral reflectance data can achieve the purpose of removing background noise.The Stacking ensemble learning algorithm is applied to build a model for remote sensing monitoring based on the first-order differential spectrum.Carry out ground-scale and satellite-scale inversion research on chlorophyll content of winter wheat to provide theoretical basis and technical support for crop growth and development and field fine regulation.