基于高光谱和BP神经网络的玉米叶片SPAD值遥感估算

为了进一步提高玉米叶绿素含量的高光谱估算精度，该文测定了西北地区玉米乳熟期叶片的光谱反射率及其对应的叶绿素相对含量（soil and plant analyzer development，SPAD）值，分析了一阶微分光谱、高光谱特征参数与 SPAD的相关关系，构建了基于一阶微分光谱、高光谱特征参数和 BP 神经网络的 SPAD 估算模型，并对模型进行验证；再结合主成分回归（principal component regression，PCR）、偏最小二乘回归（partial least squares regression，PLSR）以及传统回归模型与 BP 神经网络模型进行比较。结果表明：SPAD 值与一阶微分光谱在763nm 处具有最大相关系数（R=0.901）；以763 nm 处的一阶微分值、蓝边内最大一阶微分为自变量建立的传统回归模型可用于玉米叶片 SPAD 估算；将构建传统回归模型时筛选到的光谱参数作为输入，实测 SPAD 值作为输出，构建 BP 神经网络模型，其建模与验模 R2分别为0.887和0.896，RMSE 为2.782，RE 为4.59%，与其他回归模型相比，BP 神经网络模型预测精度最高，研究表明 BP 神经网络对叶绿素具有较好的预测能力，是估算玉米叶片 SPAD 值的一种实时高效的方法。

Estimation of maize leaf SPAD value based on hyperspectrum and BP neural network

Leaf chlorophyll content provides valuable information about the productivity, physiological status of vegetation. Measurement of hyperspectral reflectance offers a rapid, nondestructive method for leaf chlorophyll content estimation. In order to improve the accuracy of hyperspectral estimation about the leaf chlorophyll content, in this paper, the modeling of chlorophyll content of maize leaves based on the hyperspectrum was developed. The field experiments were conducted in the testing farm of Northwest Agriculture and Forest University, Yangling City, Shaanxi Province. During the maize growth period of milk stage, hyperspectral reflectance measurements were collected in wavelength of 350 to 2500 nm using spectrometer (SVC HR-1024i), and at the same time, chlorophyll content of maize leaves was obtained by using SPAD-502. There were totally 120 samples collected, two thirds of which were utilized as the training set and remaining one third as the validation set. The model constructed relied on the training set and the validation set was evaluated, respectively. The correlation between first derivative spectra, hyperspectral characteristic parameters and SPAD values were analyzed. Then single variable linear and nonlinear fitting traditional regression models respectively based on first derivative spectra and hyperspectral characteristic parameters were established to estimate the SPAD values. Besides, taken the first derivative values at 763 nm, the maximum first derivative values within blue edge (Db), red edge position (λr) and blue edge area (SDb) as the input parameters, the measured SPAD values as the output parameters, BP neural network model was built. By using the same input parameters, principal component regression (PCR) and partial least squares regression (PLSR) were used to estimate the SPAD values, too. Then we compared the predictive power of traditional regression models, PCR and PLSR models to BP neural network model. Some critical conclusions were made based on the study. First, the maximum correlation coefficient between SPAD values and first derivative spectra located at 763 nm (R=0.901) and the polynomial model was better than the linear model. As to the hyperspectral characteristic parameters, the variable among which the maximum first derivative values within blue edge (Db) was significantly related with SPAD values(R=-0.850) and its linear model was the best model of SPAD estimation models established by the hyperspectral characteristic parameters. The coefficients of determination for the calibration set of the two traditional regression models were 0.868 and 0.711, and the corresponding values of root mean square error (RMSE) were 3.069 and 4.340; for the validation set, the coefficients of determination were 0.864 and 0.743, and the values of RMSE were 3.186 and 4.317. Second, when using the BP neural network, PCR and PLSR established estimation models, the coefficients of determination for the calibration set were 0.887, 0.813 and 0.673 respectively, and the corresponding values of RMSE were 3.169, 3.495 and 5.797, respectively; for the validation set, the coefficients of determination were 0.896, 0.854 and 0.704, and the corresponding values of RMSE were 2.782, 3.221 and 6.034. Third, compared the five SPAD estimation models, BP neural network model achieved the best result in this research and the coefficients of determination of the calibration set and the validation set were highest, the value of RMSE of the validation set was lowest. The traditional regression model based on first derivative values at 763 nm performed second to BP neural network. Finally, BP neural network model had better predictive power of the chlorophyll content. The results showed that the method was a real-time and efficient method for maize leaf SPAD estimation. Our research may provide a theoretical basis for the improvement of remote sensing inversion accuracy of maize chlorophyll content.