优质棉种是全面推广棉花精量播种技术的基础。 采用近红外高光谱成像技术实现微破损棉种可视化识别, 为棉种精选设备的研制奠定理论基础。 以未破损和微破损两类棉种各540粒作为样本(其中405粒作为建模集, 135粒棉种作为预测集), 分批采集874~1 734 nm范围的样本高光谱图像, 提取光谱数据并去除首尾两端明显噪声保留955~1 659 nm范围内光谱为棉种样本的光谱。 首先使用Kennard-Stone(KS)算法进行样本划分, 并通过平滑算法Savitsky-Golay(SG)对光谱进行预处理。 采用二阶导数光谱(2nd spectra)方法、 连续投影算法(SPA)和主成分载荷(PCA-loading)方法分别选取10, 14和11个特征波长。 基于全部光谱数据和特征波长建立偏最小二乘判别分析(PLS-DA)模型、 K最邻近(KNN)模型和支持向量机(SVM)模型, SPA-PLS-DA模型取得了较好的结果, 建模集和预测集的鉴别率分别为91.50％和90.33％。 基于SPA-PLS-DA模型分别对未破损样本和微破损样本及其混合样本图像进行识别, 取得了较好的识别结果, 微破损棉种的识别率达90%以上。 结果表明, 结合近红外高光谱成像和图像处理技术, 能够实现微破损棉种的可视化识别。

High quality cotton seeds are the basis of precision seeding technique. In this paper, near-infrared hyperspectral imaging technology is used to realize the visible identification of micro-damaged cotton seeds, which lays a theoretical foundation for the development of cotton seeds selection equipment. Near-infrared hyperspectral images of two kinds of 540 cotton seeds, undamaged and micro-damaged, were acquired, of which 405 samples were used as the calibration set, and 135 samples were used as prediction set. After analyzing the original spectral curve of the full wave band, the noise at both ends was removed. Firstly, KS algorithm was used to divide samples, and the spectra was pretreated by smoothing algorithm( Savitsky-Golay), respectively using the second derivative spectra (2nd spectra) method, principal component analysis loading (PCA-loading) method and successive projection algorithm (SPA) method to extract the feature wavelength, then partial least squares discriminant analysis (PLS-DA) model, K nearest neighbor (KNN) model and support vector machine (SVM) model ware used to analyze the characteristic spectrum. By comparing the analysis results, SPA-PLS-DA was selected as the model, the discrimination rate of the calibration set and the prediction set is up to 91.50% and 90.33%, respectively. Finally, the SPA-PLS-DA model is used to identify the mixed images of undamaged and micro-damaged cotton seeds. The identification results were identified by different colors,the corresponding visual identification figure is generated, and good recognition results were obtained. Moreover, the recognition rate of micro-damaged cotton seeds was above 90%. The result indicates that the near-infrared hyperspectral technology and image processing technology can be used to realize the visual identification of the micro-damaged cottonseeds.