基于高光谱图像技术的大豆品种无损鉴别  

Nondestructive identification of soybean seed varieties based on hyperspectral image technology

作　　者：柴玉华[1] 毕文佳 谭克竹[1] 张春雷[1] 刘春涛[1] 

机构地区：[1]东北农业大学电气与信息学院,哈尔滨150030

出　　处：《东北农业大学学报》2016年第3期86-93,共8页Journal of Northeast Agricultural University

基　　金：国家自然科学基金(31271911);黑龙江省自然科学基金(ZD201303)

摘　　要：为解决传统大豆品种检测方法存在的效率低和精度差等问题，应用高光谱图像分析技术展开大豆品种甄别研究。采集10个品种（每品种100粒，共1000粒）大豆样本400．92—999．53nm的高光谱反射图像，分别进行中值平滑、多元散射校正和数据标准归一化预处理去噪，提取样本图像中心30×30pixels感兴趣区域的平均光谱曲线和标准差曲线。分别以样本平均光谱值主成分得分、标准差光谱值主成分得分及两者结合作为模型输入，基于T—S模糊神经网络和随机森林思想组合分类器构建鉴别模型。经中值平滑的光谱平均值和标准差作输入，结合随机森林思想的组合分类模型鉴别效果最佳，训练集、测试集的平均鉴别率分别达99．6％和97．6％。结果表明，采用高光谱图像技术可实现大豆品种高精度无损鉴别。In order to improve the efficiency and accuracy of identification, hyperspectral image technology was employed to determine the varieties of soybean seed. In this study, 10 soybean seed varieties with 1 000 grains were selected as inspection samples. Hyperspectral reflectance data of soybean samples were collected in spectral region of wave length of 400.92-999.53 nm, followed by the image denoised by median smoothing, correction of multiplicative scatter and normalization to extract the regions of interesting average spectral curves and standard deviation curves. Scores of average spectral data from PCs, scores of standard deviation spectral data from PCs and both of them together were used as inputs respectively. The combined classification model was developed based on T-S fuzzy neural network or random forest classifier. Taken the combination of median smoothed average and standard deviation as inputs, the random forest model presented the highest identification with 99.6% in training and 97.6% in testing. Hyperspectral image technology is feasible for non- destructive identification of soybean seed varieties.

关 键 词：大豆 高光谱图像 品种甄别 T—S模糊神经网络 随机森林思想组合分类器 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]