关键词: 光合参量, 偏最小二乘, 支持向量机, 多元线性回归, 神经网络, 高光谱

Abstract:

【Objective】Photosynthesis is the basis of crop yield and quality formation. Accurate quantitative remote sensing inversion of crop photosynthetic parameters can not only understand the growth and development of crops and the accumulation of organic matter, but also can provide reference for the ecosystem process model based on remote sensing. In order to estimate the photosynthetic characteristic parameters quickly and accurately, the hyperspectral inversion model of three photosynthetic parameters of winter wheat was constructed by combining the original spectrum, three traditional spectral transformation techniques and four simulation methods. The feasibility of hyperspectral inversion of photosynthetic parameters of winter wheat was discussed, and the applicability of different spectra and simulation methods were compared. 【Method】The maximum net photosynthetic rate (Amax), PSⅡ effective photochemical quantum yield (Fv'/Fm') of different leaf ages was obtained under the support of gas exchange and hyperspectral field experiments of winter wheat under different nitrogen application conditions. The photochemical quenching coefficient (qP) and hyperspectral reflectivity were obtained, and the reciprocal, logarithmic and first-order differential transformations of the original hyperspectrum were carried out. According to the results of correlation analysis of three photosynthetic parameters and four spectra, the band whose significant level was better than 0.01 was selected as input variable, and then the partial least square (PLS), support vector machine (SVM), multivariate linear regression (MLR) and artificial neural network (ANN) were used to establish the inversion model of photosynthetic parameters of winter wheat leaves. Based on the determination coefficient (R ²) and root mean square error (RMSE) of modeling and validation process, the simulation accuracy of different models was compared and analyzed.【Result】(1) The correlation analysis of the three photosynthetic parameters and four spectra showed that the sensitive spectral regions of the primitive, reciprocal and logarithmic spectra to the three photosynthetic parameters (Amax, Fv′/Fm′ and qP) were concentrated in the 400-750 nm spectrum range. The sensitive spectral regions of the first derivative spectrum to the three photosynthetic parameters were 470-560, 630-700 and 700-770 nm, respectively. (2) The optimal inversion model of Amax, Fv'/Fm' and qP was composed of MLR model based on reciprocal spectrum, MLR model based on first derivative spectrum and MLR model based on original spectrum, respectively. The R ² of the modeling was 0.75, 0.65 and 0.65, respectively, and the R ² of the validation was 0.73, 0.59 and 0.44, respectively. The results showed that the simulation of Amax and Fv'/Fm' based on hyperspectral method was feasible, the effectiveness of simulated qP needed further be verified. (3) The spectral performance of different transformations was different. Taking PLS simulation Amax as an example, the order of spectral performance was original spectrum > reciprocal spectrum > logarithmic spectrum > first derivative spectrum. (4) The estimation ability of different models was also different. Taking Amax simulation based on original spectrum as an example, the order of estimation ability of different models was MLR > PLS > ANN > SVM.【Conclusion】By comparing four spectra and four simulation methods, the hyperspectral inversion results of three photosynthetic parameters of winter wheat showed that Amax and Fv'/Fm' could be well simulated by hyperspectral method, but hyperspectral interpretation ability to qP was low and further study was needed. The hyperspectral information was sensitive to the photosynthetic parameters of winter wheat and affected by spectral types and simulation methods. It could be used to monitor the dynamic changes of photosynthetic capacity of winter wheat and to provide a basis for understanding the growth of crops.

Key words: photosynthetic parameters, partial least squares, support vector machine, multivariate linear regression, neural network, hyperspectral