红边光谱谐波分析的神经网络法叶绿素含量反演研究

叶绿素含量测定对于了解作物生长状况具有重要意义。为实时、快速、准确获取叶绿素含量,研究了玉米叶片叶绿素含量的BP神经网络(BPNN)法高光谱反演模型;而BPNN输入因子的选择是建立反演模型的关键。已有研究证明作物红边光谱与叶绿素含量有较强的相关性,为避免红边参数提取的不确定性,提高建模精度与效率,运用红边光谱的频率域谐波分析(HA)技术获得谐波余项、振幅和相位等能量谱特征分量(ESCC);并选择具有强相关性的10个ESCC进行主成分分析后,取前4位主分量作为BPNN的输入因子,进而进一步强化其相关性来构建叶绿素含量反演模型。同时,分别用遗传算法(GA)和小波基(wavelet-based)函数优化BPNN结构,建立GA-BPNN、WNN反演模型。实验通过比较BPNN、GA-BPNN、WNN模型和常规的多元线性回归(MLR)模型的玉米叶片叶绿素含量反演结果,得出非线性的BPNN模型要明显优于线性的MLR模型;而在神经网络模型中,GA-BPNN优化模型的反演精度最高。

Research on Inversing Chlorophyll Content Based on Neural Network methods Optimized by the Harmonic Analysis of Red Edge Spectrum

Chlorophyll content determination is of great significance to understand the status of the crop growth. Some hyperspectral inversing models were researched to real-timely, rapidly and accurately predict the chlorophyll content of corn leaf based on BP neural network (BPNN) in the paper, and the choice on input factors of BPNN is the key to build the inversing models in the study. Existing studies have shown that the correlation is strong between crop red edge spectrum and chlorophyll content. In order to avoid the uncertainty in extracting the red edge parameters and improve the modeling precision and efficiency, some energy spectrum characteristic components (ESCC) of red edge spectrum decomposed in frequency domain were obtained like the harmonic remainder, amplitudes and phases by the harmonic analysis (HA) technology, then there are 10 ESCCs that had higher correlation with chlorophyll content were selected and processed by the principal component analysis(PCA), finally the first 4 principal components after PCA processing were chosen as the input factors of BPNN for further strengthening the correlation in order to better build the inversing models. Meanwhile, the BPNN model was optimized using genetic algorithm (GA) and wavelet-based function respectively, so that the GA-BPNN and WNN inversing models were established, too. The experimental results proved that the nonlinear BPNN models were better than multiple linear regression model (MLR) by comparing the chlorophyll content of corn leaf inversed by the BPNN, GA-BPNN, WNN and normal MLR models, and the inversing accuracy of the optimized GA-BPNN model is the highest in all of the BPNN methods.