基于GA-BP神经网络高光谱反演模型分析玉米叶片叶绿素含量

Analysis of Chlorophyll Contents in Maize Leaf Based on GA-BP Neural Network Hyperspectral Inversion Model

叶绿素是评价玉米健康状况的重要生理生化参数,而快速、准确检测玉米叶片叶绿素含量,是实现玉米长势及健康状况精准诊断的关键。为提高玉米叶片叶绿素含量的高光谱反演精度,以玉米试验小区为基础,测定了东北地区玉米不同生长期的叶片光谱反射率及其对应的叶绿素含量。首先采用一阶微分方法提取光谱特征,构建9种高光谱特征参数(Db、Dy、Dr、λb、λy、λr、SDb、SDy和SDr),并分析一阶微分光谱、高光谱特征参数与叶绿素含量间的相关关系,优选出与叶绿素含量相关性较高的3种特征参数作为自变量,分别为535nm处的一阶微分值、蓝边内最大一阶微分值Db、蓝边面积SDb,叶绿素含量实测值作为因变量,随后采用遗传算法对BP神经网络进行优化,建立BP神经网络(BPNN)和遗传算法优化的BP神经网络(GA-BPNN)反演模型,并对模型进行验证;再结合主成分回归(PCR)和偏最小二乘回归(PLSR)模型进行比较。结果表明:叶绿素含量与一阶微分光谱在535nm处具有最大相关系数(R=-0.738),并且与特征参数Db、SDb呈显著相关,相关系数R分别为-0.732和-0.728;遗传算法可以有效地对BPNN初始权值随机化、易陷入局部极值等不足实现优化,并为其定位出理想的搜索空间;GA-BPNN模型的建模集与验证集R2分别为0.878和0.898,RMSE为0.731,与其他反演模型相比,GA-BPNN模型的稳定性和预测能力均表现最好,可为定量预测玉米叶片叶绿素含量提供一定的理论和技术依据。

Since Chlorophyll is an important physiological and biochemical parameter for evaluating the health status of maize,the rapid and precision detection of chlorophyll contents in corn leaves is the key to accurate diagnosis of maize growth and health status.In order to improve the hyper spectral inversion accuracy of chlorophyll contents in corn leaves based on the maize experimental plot,the spectral reflectance of leaf and its corresponding chlorophyll contents in different growth stages of maize in Northeast China were determined.Firstly,the first-order differential method was used to extract the spectral signature,and nine hyper spectral characteristic parameters(Db,Dy,Dr,λb,λy,λr,SDb,SDy and SDr)were constructed.And the first-order differential spectroscopy,the relationship between the hyper spectral characteristic parameters and the chlorophyll contents were analyzed.Three characteristic parameters with high correlation were selected with chlorophyll contents as independent variables,which were first-order differential value at 535 nm,maximum first-order differential value Db in blue edge,blue-edge area SDb.The measured value of chlorophyll contents was used as the dependent variable,and then BP neural network was optimized by genetic algorithm,BP neural network(BPNN)and genetic algorithm optimized BP neural network(GA-BPNN)inversion model were established,then the models were validated and combined with principal component regression(PCR)and partial least squares regression(PLSR)models for comparison.The results showed that the chlorophyll content value and the first-order differential spectrum had the largest correlation coefficient at 535nm(R=-0.738),and were significantly correlated with the characteristic parameters Db and SDb.The correlation coefficients R were-0.732 and-0.728,respectively.The genetic algorithm could effectively optimize the BPNN initial weight randomization and easy to fall into local extremum,and to locate the ideal search space for it;The modeling set and validation set R2 of the GA-BPNN model were 0.878 and 0.898,respectively,and the RMSE was 0.731.Compared with other inversion models,GA-BPNN model had the best stability and prediction ability,which can provide a theoretical and technical reference for quantitative prediction of corn leaf chlorophyll contents.