基于校正光谱序列融合的小麦腥黑穗病籽粒分类方法

针对小麦腥黑穗病轻度患病籽粒易与健康籽粒混淆，人工识别难度大的问题，将校正光谱序列融合技术与深度学习模型相结合，实现小麦腥黑穗病籽粒快速、精准分类。以健康、轻度患病、重度患病各300粒小麦籽粒的高光谱数据为样本，通过多元散射校正算法(MSC)和标准正态变换算法(SNV)对原始光谱进行预处理，并利用二维相关光谱法(2D-COS)分析SNV与MSC算法处理后的光谱之间的互补性。使用校正光谱序列融合技术将原始光谱、SNV预处理光谱与MSC预处理光谱三者进行融合得到序列融合光谱，以充分利用不同光谱预处理数据间的互补信息。最终，利用序列融合光谱数据建立基于ResNet 50算法的小麦腥黑病分类模型。试验结果表明，序列融合光谱ResNet 50模型总体准确率最高为93.89%，F1值为93.87%，分类性能优于单一预处理光谱建立的ResNet 50模型。为进一步评估模型分类效果，使用序列融合光谱分别建立偏最小二乘判别分析(PLS-DA)、支持向量机(SVM)以及集成学习算法模型随机森林(RF)与极端梯度提升树(XGBoost)模型，并进行对比，结果显示：SVM、PLS-DA、RF与XGBoost总体准确率分别为81.67%、84.44%、89.44%与90.55%，F1值分别为81.59%、84.04%、89.49%与90.59%，ResNet 50总体准确率与F1值优于传统光谱分析模型。因此，本研究表明校正光谱序列融合技术结合深度学习模型，能够实现对不同患病程度腥黑穗病籽粒的有效分类。

Classification of Common Bunt of Wheat Kernels Based on Series Fusion of Scatter Correction Techniques

An innovative approach that integrated series fusion of scatter correction techniques with deep learning models was introduced to achieve rapid and precise classification of common bunt in wheat kernels. Manual identification of this disease can be particularly challenging, especially in cases with mild infections. To address this challenge, the high-spectral data was leveraged from a sample set comprising 300 kernels, encompassing healthy, mildly infected, and severely infected kernels. The original spectra underwent preprocessing by using the multiplication scatter correction (MSC) and standard normal variate (SNV) algorithms. Furthermore, two-dimensional correlation spectroscopy (2D-COS) analysis was employed to assess the complementarity between spectra processed by SNV and MSC. Subsequently, the series fusion of scatter correction techniques was applied to amalgamate the original spectra, SNV processed spectra, and MSC-processed spectra, resulting in fused spectral sequences that harnessed the complementary information from various spectral preprocessing methodologies. Following this, a classification model for wheat common bunt, based on the ResNet 50 algorithm, was developed by using the fused spectral data. Experimental results demonstrated that the ResNet 50 model achieved the highest classification accuracy of 93.89% and an F1-score of 93.87%, surpassing models based on individual preprocessing methods. To further evaluate the classification performance of the model, partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), and ensemble learning algorithms, random forest (RF), and extreme gradient boosting (XGBoost) models were constructed by using the fused spectral data and comparison was done. The results revealed that SVM, PLS-DA, RF, and XGBoost achieved overall recognition accuracies of 81.67%, 84.44%, 89.44%, and 90.55%, respectively, with corresponding F1-scores of 81.59%, 84.04%, 89.49%, and 90.59%. Importantly, the ResNet 50 model outperformed traditional spectral analysis models in terms of overall accuracy and F1-score. In summary, ResNet 50 outperformed traditional spectral analysis models in terms of both overall accuracy and F1-score. In conclusion, this research underscored the efficacy of combining series fusion of scatter correction techniques with deep learning models for the classification of common bunt in wheat kernels at varying infection levels. This approach held promise for the development of rapid and non-destructive detection methods for common bunt in wheat kernels.