【摘要】 针对传统玉米茎秆强度破坏式检测方法费时费力的问题，采用高光谱成像数据结合统计学习方法对灌浆期和蜡熟期的19个玉米品种的茎秆穿刺强度和折断力进行检测，给出适于进行玉米茎秆强度检测的特征提取和建模方法。试验在5 000株·亩^-1种植密度下种植了19个玉米品种，采集灌浆期和蜡熟期茎秆基部的高光谱图像，使用目标区域分割的方式自动进行光谱图像反射率校正和目标光谱曲线提取。对采集的样本数据使用主成分分析法（PCA）和包裹式特征提取法提取光谱特征，并分别建立了主成分回归（PCR）和偏最小二乘回归（PLSR）的茎秆强度预测模型。通过对各特征提取方法和各模型交叉验证预测结果的对比，找到适于进行玉米茎秆强度检测的特征提取和建模方法。试验结果表明：PCA方法提取光谱特征具有明显的降维效果，但用PCA方法提取特征建立的PCR模型对玉米茎秆强度的预测效果一般，用包裹式特征提取方法建立的PLSR模型在灌浆期和蜡熟期的模型预测效果均优于PCR模型，模型的剩余预测残差（RPD）在2.90～3.93之间，可以用于定量分析预测茎秆强度。更多还原

【Abstract】 Given the time-consuming and labor-consuming problem of traditional maize stalk strength destructive detection methods, this study used hyperspectral imaging data combined with statistical learning methods to detect the puncture strength and breaking force of the stalks of 19 maize varieties in the filling stage and wax maturity stage. Moreover, the feature extraction and modeling methods suitable for detecting corn stalk strength are given. In the experiment, 19 corn varieties were planted at a planting density of 5 000 plants·mu^-1. The hyperspectral images of the base of the stalks at the filling stage and wax maturity stage were collected, and the target area segmentation method was used to automatically perform spectral image reflectance correction and target spectral curve extraction. Principal Component Analysis（PCA） and wrapped feature extraction were used to extract spectral features from the collected sample data, and principal component regression（PCR） and partial least squares regression（PLSR） were developed for the prediction of stalk strength. By comparing each feature extraction method and the cross-validation prediction results of each model, we found suitable feature extraction and modeling methods for maize stalk strength detection. The experimental results showed that the PCA method extracted spectral features had obvious dimensionality reduction effect. However, the PCR model built with PCA method extracted features had average prediction effect on maize stalk strength, and the PLSR model built with wrapped feature extraction method had better prediction effect than the PCR model at both the filling and waxing stages. The residual predictive deviation（RPD） of the PLSR model was higher than that of the PCR model. The RPD of the PLSR model ranged from 2.90 to 3.93, which could be used for quantitative analysis to predict stalk strength.更多还原