Abstract: Cotton is an important economic crop. However, the impurity content of cotton will have an impact on the cotton ginning and spinning effects, ultimately affecting the quality of cotton products such as yarn and fabrics. The impurity content of cotton is one of the main indicators for cotton grading and pricing, so the detection of cotton impurities is meaningful and valuable in practical application. The research on cotton impurity detection is rich in content and diverse in methods, mainly including hand selection, image methods, and spectral methods. Hyperspectral imaging technology was first applied in the field of remote sensing, and in recent years, it has been sprouting up in cotton impurity detection. In order to detect and identify cotton impurities, the unginned cotton was firstly picked by machine; the cotton fibers were then rolled down from the unginned cotton and impurities were removed to get the raw cotton. Then, the raw cotton was purified again and separated to obtain four types of substances: pure cotton, cotton branches and leaves, broken seeds, and mudstone. Each sample weighed 10g, with 30 samples for each substance, and their hyperspectral images were collected in the hyperspectral imaging system. Black-and-white plate correction and smoothing processing were applied to the collected hyperspectral images, then the region of interest was selected and spectral data were extracted. The spectra of the same substance were averaged, and the standard spectral curve of the corresponding substance was obtained and analyzed. By using three methods of multiplicative scatter correction (MSC), standard normal variate transformation (SNV), and first derivative (FD) to preprocess the extracted spectra of pure cotton and three types of impurities, the paper analyzed the effects of different preprocessing methods. Principal component analysis was used to reduce the dimensionality of all original spectra and preprocessed spectra of the samples. The top eight principal components were selected as factors based on cumulative contribution rate to establish a discriminant analysis model and determine the optimal impurity identification and classification model. The results showed that the trends of spectral curves of pure cotton, cotton branches and leaves, and broken seeds were similar, while the spectral curves of mudstone were flat and significantly different from the other three types of spectra. The above three methods could effectively eliminate baseline drift in the spectral curves, making the spectra smoother. Spectral trends preprocessed by MSC and SNV were similar. The FD preprocessing amplified the original small characteristic peaks, making them more prominent. Dimensionality reduction by principal component analysis (PCA) could effectively solve the problem of huge information content and wide spectral bands in hyperspectral data. When the number of principal components reached eight, the cumulative contribution rates of the original spectrum and the three preprocessed spectra could all reach 85%; the training sets' accuracy of the discriminant analysis model corresponding to the original spectrum and three spectra preprocessed by MSC, SNV, and FD was 100%, 96.67%, 98.89%, and 100%, respectively. The accuracy of the test sets was 100%, 90%, 93.33%, and 100%, respectively. In conclusion, hyperspectral imaging technology can effectively detect and identify impurities in raw cotton.

Key words: cotton impurities, hyperspectra, spectral preprocessing, principal component analysis, discriminant analysis

摘要： 棉花杂质含量是棉花定级和定价的主要指标之一，实现棉花杂质的快速、无损检测具有重要的现实意义和应用价值。对原棉进行除杂，并分离得到纯棉、棉枝叶、破籽和泥石4类物质，在高光谱成像系统中采集这4类物质高光谱图像，提取并分析他们的高光谱数据及特征，采用多元散射校正、标准正态变量变换和一阶导数作为数据预处理方法，选用主成分分析进行数据降维及可区分性研究，结合主成分个数的优选建立判别分析模型，建立了原棉杂质检测模型，并对比分析所建立的不同检测模型的分类效果。结果表明：多元散射校正、标准正态变量变换和一阶导数3种方法均能有效消除光谱曲线的基线漂移现象，使原有特征峰更加明显；检测模型的主成分个数优选为8；对应原始光谱、多元散射校正、标准正态变量变换和一阶导数预处理的判别分析模型测试集准确率分别为100%、90%、93.33%和100%，证实了利用高光谱成像技术能够对原棉杂质进行有效检测识别。

关键词: 棉杂质, 高光谱, 光谱预处理, 主成分分析, 判别分析