采集东北和非东北产地大米样本高光谱图像, 筛选多个特征波长图像并提取图像特征, 结合模式识别方法建立判别模型, 并联合多个模型构成模型集群快速、 准确判别东北/非东北大米产地。 东北大米以粳米为主, 主要涵盖长粒香, 圆粒香, 稻花香和小町米4个品种。 为建立实用性强、 适用范围广的东北/非东北大米产地判别模型, 实验主要收集了国内粳米代表性产区且以上述4个品种为主的样本, 构成原始样本集: 其中东北产地5份, 包括黑龙江(1)、 吉林(2)、 辽宁(2), 非东北产地5份, 包括河北(1)、 浙江(1)、 江苏(2)、 安徽(1)。 每个产地样本随机选取100粒, 共计100×10粒大米样本。 采用芬兰Specim公司的SisuCHEMA高光谱成像系统采集样本900~1 700 nm高光谱图像。 按照大米轮廓选取感兴趣区域提取出单粒大米样本的平均光谱, 采用Kennard-Stone法按照4∶1划分训练集和测试集。 应用连续投影算法筛选得到原始样本集光谱的8个特征波长: 1 460.30, 1 400.20, 1 424.92, 945.98, 1 315.62, 1 220.87, 1 705.91和942.53 nm; 采用方向梯度直方图分别提取8个波长下的图像特征, 结合支持向量机建立基于单特征波长图像的东北/非东北大米产地鉴别模型, 识别准确率分别为85.5%, 77.5%, 76.5%, 73.5%, 71%, 68.5%, 67%和65.5%; 鉴于单模型识别率不高的现状, 提出建立基于特征波长图像模型集群综合判别大米产地的策略, 即按照单模型识别率从高到低排序后分别联合3个、 5个和7个特征波长图像模型的预测结果, 当预测样本判定为真的比率>50%, 则判定样本为真, 反之则为假。 联合1 460.30, 1 400.20, 1 424.92, 945.98, 1 315.62, 1 220.87和1 705.91 nm七个波段的模型集合对测试集样本的识别率可达90.5%。 实验结果表明高光谱结合模型集群策略可为建立性能稳健、 适用范围广的东北/非东北大米产地快速检测模型提供切实可行的思路和方法参考。

Hyperspectral images of rice from northeast/non-northeast regions were collected, and spectral images at characteristic wavelengths were screened. The clustering combination of image features and pattern recognition method was established to quickly and accurately identify northeast/non-northeast rice origin. Northeast rice is mainly japonica rice,and the typical northeastern rice varieties include long-grain, round-grain, rice flower and Xiaoding rice. Considering the practicability and applicability of rice origin identification model,samples of 10 origins and 4 varieties above were collected to form the original sample set. Among them, there are five northeastern origins, including Heilongjiang (1), Jilin (2), Liaoning (2), and five non-northeastern origins, including Hebei (1), Zhejiang (1), Jiangsu (2) and Anhui (1). 100 samples were selected randomly from each producing area. Hyperspectral images of 100×10 rice samples were collected using SisuCHEMA hyperspectral imaging system (Specim, Finland)in the range of 900~1 700 nm. Extracting the average spectra of a single rice sample by selecting the region of interest according to the rice contour,Kennard-Stone method was used to divide training set and test set according to the ratio of 4∶1. Eight characteristic wavelengths were screened by Successive Projections Algorithm(SPA): 1 460.30, 1 400.20, 1 424.92, 945.98, 1 315.62, 1 220.87, 1 705.91, 942.53 nm. The eight models were built respectively by HOG features extracted from single characteristic wavelength Image and SVM to identify the rice origin whether it was from northeast or non-northeast China. The recognition accuracy was as follows: 85.5%, 77.5%, 76.5%, 73.5%, 71%, 68.5%, 67%, 65.5%. In view of the low recognition rate of single model, a strategy of establishing model cluster based on single characteristic wavelength image model to synthetically discriminate rice origin was proposed. According to the recognition rate of single model from high to low, the cluster models were established by respectively combining three, five and seven the signal models above. While the probability of the sample judged to be true predicted by the conjunctive model is greater than 50%, the sample will be judged to be true, otherwise it will be false. The experimental results showed that the recognition rate of the test set samples can reach 90.5% by combining the model sets of 1 460.30, 1 400.20, 1 424.92, 945.98, 1 315.62, 1 220.87 and 1 705.91 nm bands. This study shows that hyperspectral technology combined with the strategy of conjunctive model consensus can provide feasible and effective methods to establish a robust and wide applicability model to recognize the rice origin (northeast/non-northeast) rapidly.