Near infrared hyperspectral imaging over the spectral range of 900–2500 nm was investigated for its potential to identify maize kernels inoculated with aflatoxigenic fungus (AF13) from healthy kernels and kernels inoculated with non-aflatoxigenic fungus (AF36). A total of 900 kernels were used with 3 treatments, namely, each 300 kernels inoculated with AF13, AF36 and sterile distilled water as control, separately. One hundred kernels from each treatment of 300 kernels were incubated for 3, 5 and 8 days, to obtain diverse samples. Based on the full mean spectra extracted from the same kernel side(s), the best mean overall prediction accuracies achieved were 96.3% for the 3-class (control, non-aflatoxigenic and aflatoxigenic) classification and 97.8% for the 2-class (aflatoxigenic-negative and -positive) classification using the partial least-squares discriminant analysis (PLS-DA) method. The 3-class and 2-class models using the full mean spectra extracted from different kernel sides had the best mean overall prediction accuracies of 91.5% and 95.1%. Using the most important 30, 55 and 100 variables determined by the random frog (RF) algorithm, the simplified type I-RF-PLSDA models achieved the mean overall prediction accuracies of 87.7%, 93.8% and 95.2% for the 2-class discrimination using different kernel sides’ information. Among the most important 55 and 100 variables, a total of 25 and 67 variables were consistently selected in the 100 random runs and were therefore used further for establishing the type II-RF-PLSDA models. Using these 25 and 67 variables, the type II-RF-PLSDA models obtained the mean overall prediction accuracies of 82.3% and 94.9% separately.

中文翻译：

---

一种新的基于高光谱的方法，用于鉴定感染了多种黄曲霉真菌的玉米粒

研究了 900-2500 nm 光谱范围内的近红外高光谱成像，研究其从健康玉米粒和接种非黄曲霉毒素真菌 (AF36) 的玉米粒中识别接种了黄曲霉毒素真菌 (AF13) 的玉米籽粒的潜力。共900粒，分3个处理，即每300粒分别接种AF13、AF36和无菌蒸馏水作为对照。每次处理 300 个谷粒，将 100 个谷粒培养 3、5 和 8 天，以获得不同的样品。基于从相同内核侧提取的完整平均光谱，对于 3 类（对照、非黄曲霉毒素和黄曲霉毒素）分类和 97.0 获得的最佳平均总体预测精度为 96.3%。8% 使用偏最小二乘判别分析 (PLS-DA) 方法进行 2 类（产黄曲霉毒素阴性和阳性）分类。使用从不同核侧提取的全平均谱的 3-class 和 2-class 模型具有最佳的平均总体预测准确率，分别为 91.5% 和 95.1%。使用随机青蛙 (RF) 算法确定的最重要的 30、55 和 100 个变量，简化型 I-RF-PLSDA 模型对 2 类判别实现了 87.7%、93.8% 和 95.2% 的平均总体预测准确率使用不同内核端的信息。在最重要的 55 和 100 个变量中，在 100 次随机运行中一致选择了总共 25 和 67 个变量，因此进一步用于建立 II-RF-PLSDA 模型。使用这 25 个和 67 个变量，