基于高光谱成像法对大豆种子的品种分类鉴别

Variety Classification and Identification of Soybean Seeds Based on Hyperspectral Imaging Method

摘要: 本文针对东北大豆品种分类鉴别的需求，利用高光谱成像技术获取了6种大豆样品在392.38~1011.01 nm高光谱图像，提取感兴趣(ROI)区域数据，获得大豆种子样品的反射光谱曲线。经过卷积平滑(S-G)预处理，再根据大豆光谱曲线差异选取455.54 nm、479.3 nm、604.04 nm、657.46 nm、705.72 nm、856.89 nm、918.07 nm、953.54 nm作为特征波段，分别输入极限学习机(ELM)和随机森林(RF)模型，得到的分类正确率分别为78.22% 和98.89%，模型预测时间分别为11 s和12 s。研究结果表明，经卷积平滑和高光谱波段优化的特征波段，运用随机森林(RF)模型分析是分类准确率最高、预测时间较快的分类方法，高光谱成像法可有效对大豆品种做出分类鉴别。

Abstract: According to the needs of soybean variety classification and identification in Northeast China, the hyperspectral images of six soybean samples at 392.38~1011.01 nm were obtained by hyperspectral imaging technology, the region of interest (ROI) data were extracted, and the reflection spectrum curves of soybean samples were obtained. After convolution smoothing (S-G) pretreatment, 455.54 nm, 479.3 nm, 604.04 nm, 657.46 nm, 705.72 nm, 856.89 nm, 918.07 nm and 953.54 nm were selected as the characteristic bands according to the difference of soybean spectral curves. The classification accuracy was 78.22% and 98.89% respectively, and the prediction time of the model was 11 s and 12 s respectively. The results show that the random forest (RF) model analysis is the classification method with the highest classification accuracy and faster prediction time. Hyperspectral imaging method can effectively classify and identify soybean varieties.