Research On Hyperspectral Estimation Of SPAD Of Rice Canopy In Cold Region Based On Ensemble Learning

Rice is one of the most important food crops in China,accounting for half of the total grain yield.Accurate management of rice production is of great significance for ensuring national food security.Nitrogen fertilizer is an indispensable nutrient element in the growth process of rice.The precise application of nitrogen fertilizer can increase rice yield,reduce production costs,reduce environmental pollution,and achieve sustainable development of agricultural production.There is a significant linear correlation between the nitrogen content of rice and the relative chlorophyll content(SPAD)of rice leaves,which can effectively reflect the nutritional status of rice.Accurate measurement of SPAD content in rice canopy has important guiding significance for rice nutrition monitoring,fertilization regulation,and agricultural irrigation.Spectral technology,especially hyperspectral technology,provides effective means for rapid,non-destructive,and real-time monitoring of rice,and provides a method for large-scale monitoring of rice nutritional status.This paper takes Heilongjiang rice under different growth cycles and nitrogen levels as the research object,and takes the Northeast Agricultural University Base in Acheng District,Harbin City as the research area.From 2021 to 2022,field experiments on rice in cold regions were carried out.The UAV was equipped with a hyperspectral imager to obtain rice hyperspectral image data and synchronously obtain rice ground SPAD.Based on hyperspectral data,the quantitative relationships between characteristic band combination,vegetation index and canopy SPAD were explored,and ensemble learning algorithms were used to construct a prediction model for the relative chlorophyll content of rice canopy in cold regions,and spatial inversion of rice canopy SPAD in cold regions was carried out.This paper aims to provide a scientific basis for real-time non-destructive monitoring of chlorophyll nutritional status and fertilization management of rice canopy in cold regions.The main research contents and conclusions of this paper are as follows:(1)Acquisition and processing of hyperspectral data and ground sample data.According to the research tasks of the project,the research plan of the project was formulated.During the rice tillering stage,jointing stage and heading stage,the UAV equipped with the S185 hyperspectral imager obtained the rice canopy hyperspectral image data,and the handheld chlorophyll meter SPAD-502 obtained the rice ground samples.Multivariate Scatter Correction(MSC)was used to preprocess the spectral data,and the evaluation indicators of the prediction model established by the preprocessed data were all better than the original spectral model.(2)Extraction of hyperspectral data feature variables.In order to solve the problems of large amount of full-band spectral data,high dimensionality,and high cost of training models,this paper used three methods of SPA,CARS,and Vegetation Index to obtain characteristic variables as the input of the prediction model.At the tillering stage,jointing stage and heading stage of rice,the number of characteristic bands extracted based on the SPA algorithm were 9,8 and 5 respectively;the number of characteristic bands extracted based on the CARS algorithm were 19,25 and 20 respectively.For the Vegetation Index,this paper selected 13 commonly used Vegetation Indexes,used the hyperspectral information of the canopy leaves in the three growth stages of cold rice to calculate the Vegetation Index value,and analyze the correlation between the 13 Vegetation Indexes and SPAD.The relatively sensitive Vegetation Indexes were obtained to form a sensitive Vegetation Index combination at different stages.The results showed that the number of characteristic bands selected by the CARS algorithm is less than 5%of the total bands,effectively retaining useful band information,eliminating useless redundant bands,and showing the best prediction effect in the model.(3)Based on four ensemble learning algorithms,prediction models of rice SPAD in cold regions were established.In this paper,the selected SPA characteristic band combination,CARS characteristic band combination,and sensitive Vegetation Index combination were used as input variables,and SPAD was used as the response variable to construct cold region rice SPAD inversion models based on XGBoost,LightGBM,GA-XGBoost,and GA-LightGBM respectively.,and conducted model comparison analysis.The research results showed that:in the three growth stages,the model with the best performance is CARS-GA-LightGBM,in which the R2 of the prediction model verification set based on the tillering stage data is 0.857,and the RMSE is 0.887;the jointing stage verification set R2 is 0.916,RMSE is 0.687;the R2 of the heading date validation set is 0.877,and RMSE is 0.849.In summary,this paper proposed an effective prediction method for rice SPAD in cold region based on hyperspectral technology,realized precise monitoring of rice nutrition in different growth stages.This technology provided support for rice precision and information-based planting,and also provided a new method for rice fertilization and fertilizer control management.