SiPLS-CARS与GA-ELM对哈密瓜冠层叶片含水率的反演估测

传统的叶片含水率检测方法效率低、操作繁琐且是有损的检测，不利于大田哈密瓜叶片含水率的快速获取。为实现对大田哈密瓜生长期进行更精细的田间灌水管理，利用光谱技术分别获取了哈密瓜植株在成长期（M₁）、开花期（M₂）、结果期（M₃）、成熟期（M₄）四个时期内的冠层叶片样本，采用烘干法测得叶片样本的含水率。为提高预测模型的精度和稳定性，首先开展并讨论极限学习机（ELM）模型中的核函数与隐含层神经元个数的选择对ELM模型精度的影响。随后分别利用联合子区间偏最小二乘法（SiPLS）及其与竞争性自适应重加权采样法（CARS）、遗传算法（GA）、连续投影算法（SPA）的组合算法对全波段光谱数据中与叶片含水率相关性高的特征波长进行筛选提取。再分别使用GA与粒子群算法（PSO ）对已经确定最佳核函数与隐含层神经元个数的ELM模型中的输入层与隐含层间的连接权值（W）和隐含层神经元阈值（B）进行优化选择，获取最优且稳定的W与B值，进一步提高模型的稳定性和预测精度。最后将四种特征波长提取算法优选出的特征波长分别进行ELM，GA-ELM，PSO-ELM建模分析，以校正集和预测集的相关系数R_c与R_p为模型评价指标，经过对比分析优选出能准确预测哈密瓜冠层叶片含水率的反演估测模型。采用SiPLS及其与CARS，GA和SPA的组合算法提取特征波长，筛选出的变量数分别为273，20，32和6，占全光谱变量的15.6%，1.2%，1.9%和0.03%。进一步将筛选出的特征波长作为自变量，叶片的含水率作为因变量，建立了ELM的预测模型，最佳预测精度R_p值为0.845 0，预测精度不是很理想。故引入GA与PSO对ELM中随机产生的W与B值进行优化选择。最终，经过研究发现，利用GA优化后的ELM模型结合SiPLS-CARS筛选出的特征波长建立的哈密瓜冠层叶片含水率预测精度最优，故反演叶片含水率的最优建模方式为SiPLS-CARS-GA-ELM，R_c值为0.928 9，R_p值为0.903 2，所建模型精度较高，可为大田哈密瓜冠层叶片的含水率进行快速检测，为田间灌溉管理提供科学依据。

Estimation of Leaf Moisture Content in Cantaloupe Canopy Based on SiPLS-CARS and GA-ELM

To realize more precise irrigation management during the growing period of Hami Melon in the field. The traditional methods for measuring leaf moisture content are inefficient, complicated and destructive, which is not conducive to obtaining moisture content of Hami melon leaves in the field. In this study, the leaf samples of cantaloupe in four periods of growth (M₁), flowering (M₂), fruit (M₃) and maturity (M₄) were obtained by spectral technology, and the moisture content of the leaf samples was measured by drying method. The influence of the choice of kernel function and the number of hidden neurons on the precision of the ELM model is discussed. Then SiPLS and its combined algorithm with CARS, GA and SPA were used to extract the characteristic wavelengths with a high correlation with leaf moisture content. GA and PSO algorithms are used to optimize the connection weights (W) between the input layer and the hidden layer of the ELM model, and the threshold (B) of the hidden layer of the ELM model, the optimal and stable W and B values are obtained further to improve the stability and prediction accuracy of the model. Finally, four feature wavelength extraction algorithms are combined with ELM, GA-ELM and PSO-ELM to analyze the model, and the Correlation Coefficient between the correction set and the prediction set is taken as the evaluation index of the model. Through the comparison and analysis, the inversion estimation model of cantaloupe canopy leaf moisture content was optimized. The results show that the number of SiPLS and its combination with CARS, GA and SPA are 273, 20, 32 and 6 respectively, accounting for 15.6%, 1.2%, 1.9% and 0.03% of the total spectrum variables. Taking the selected characteristic wavelength as the independent variable and the moisture content of the leaves as the dependent variable, the prediction model of ELM is established, but the prediction accuracy is not very ideal. Therefore, GA and PSO are introduced to optimize the randomly generated W and B values in ELM. Finally, it is found that the precision of predicting water content of cantaloupe canopy leaves based on the ELM model optimized by GA and SiPLS-CARS is the best. Therefore, the optimal modeling method of leaf moisture content retrieval is SiPLS-CARS-GA-ELM, R_C value is 0.928 9, R_P value is 0.903 2, the precision of the model is high, which can be used to detect the leaf moisture content in cantaloupe canopy, the research provides the theoretical basis for the field irrigation management.