Abstract: To meet the demand on rapid and damage-free rice origin confirmation, hyper-spectral images(400~1 000 nm) of a total number of 990 rice samples from primary rice production regions of Meihekou city, Jilin province, as well as other rice producing regions such as Songyuan, Daan and Huinan, were collected as the object of study, and the spectra were preprocessed by multiplicative scatter correction (MSC) handling method. With the adoption of multilayer perceptro (MLP), extreme learning machine (ELM) and online sequence extreme learning machine (OS-ELM) algorithms, the origin confirmation model was established on the basis of the full-band hyperspectral data and the data undergoing dimensionality reduction by the multidimensional scaling (MDS) method. The results indicated that the OS-ELM model based on full-band hyperspectral data had the best classification performance with an accuracy rate of 98.3%. After the MDS processing, the input data variables were reduced by 96.6%, while the accuracy rate of the MDS-OS-ELM model was stable at 97.4%. The comparative analysis was performed against the training time among three models, and it was concluded that the training time of OS-ELM was significantly more than that of MLP, and the training time on data acquisition in batches was more than ELM, which provided an efficient, accurate and stable method for rice origin confirmation.

Key words: hyper-spectral, multidimensional scaling, extreme learning machine, online sequential-extreme learning machine, multi-layer perceptron

摘要： 为满足快速无损的大米产地确证需求，采集吉林省梅河口市水稻主产区及松原、大安、辉南等其他水稻产区共990个大米样本的高光谱图像（400~1 000 nm）作为研究对象，利用多元散射校正（MSC）处理方法对光谱进行预处理。采用多层感知机（MLP）、极限学习机(ELM)与在线序列极限学习机(OS-ELM)算法，分别基于全波段高光谱数据以及经多维尺度分析（MDS）方法降维后的数据建立产地确证模型。结果表明，基于全波段高光谱数据的OS-ELM模型分类性能最好，准确率达到98.3%。经MDS处理后，输入的数据变量减少了96.6%，MDS-OS-ELM模型准确率稳定在97.4%。对三种模型的训练时间进行对比分析，OS-ELM训练时间明显优于MLP，在分批次获取数据时训练时间优于ELM。为大米产地确证提供了一种高效、准确、稳定的方法。

关键词: 高光谱图像技术, 多维尺度分析, 在线序列极限学习机, 极限学习机, 多层感知机