基于主动学习的玉米种子纯度检测模型更新

种子纯度反映种子品种在特征特性方面典型一致的程度，提高种子纯度检测的准确性和可靠性对保证种子的质量具有重要的意义。高光谱图像技术可以同时反映种子的内部特征和外部特征，在农产品无损检测中已经得到广泛应用。利用近红外高光谱图像实现农产品无损检测的实质就是建立光谱信息与农产品品质参数之间的数学模型关系。但光谱信息易受环境、时间的影响，当待测样本的产地或者年份发生改变时光谱信息也随之改变，导致建立的模型的稳定性变差、泛化能力减弱。针对这一问题，采用主动学习算法选择具有代表性的待测样本，最终以添加最少最优的样本数来扩大原模型的样本空间，从而实现模型的快速更新，提高模型的稳定性，同时与基于随机选择算法（RS）和Kennard-Stone算法(KS)的模型更新效果进行比较。实验结果表明：在不同样本集划分比例下（1∶1, 3∶1, 4∶1），利用主动学习添加40个新样本更新后的2010年的玉米种子纯度检测模型对2011年新样本的预测精度由47%，33.75%，49%提高到98.89%，98.33%，98.33%;利用主动学习添加56个新样本更新后的2011年的玉米种子纯度检测模型对2010年新样本的预测精度由50.83%，54.58%，53.75%提高到94.57%，94.02%，94.57%;同时基于主动学习算法的模型更新效果明显优于RS和KS。因此基于主动学习算法实现玉米种子纯度检测模型的更新是可行的。

Purity Detection Model Update of Maize Seeds Based on Active Learning

Seed purity reflects the degree of seed varieties in typical consistent characteristics, so it is great important to improve the reliability and accuracy of seed purity detection to guarantee the quality of seeds. Hyperspectral imaging can reflect the internal and external characteristics of seeds at the same time, which has been widely used in nondestructive detection of agricultural products. The essence of nondestructive detection of agricultural products using hyperspectral imaging technique is to establish the mathematical model between the spectral information and the quality of agricultural products. Since the spectral information is easily affected by the sample growth environment, the stability and generalization of model would weaken when the test samples harvested from different origin and year. Active learning algorithm was investigated to add representative samples to expand the sample space for the original model, so as to implement the rapid update of the model’s ability. Random selection (RS) and Kennard- Stone algorithm (KS) were performed to compare the model update effect with active learning algorithm. The experimental results indicated that in the division of different proportion of sample set (1∶1, 3∶1, 4∶1), the updated purity detection model for maize seeds from 2010 year which was added 40 samples selected by active learning algorithm from 2011 year increased the prediction accuracy for 2011 new samples from 47%, 33.75%, 49% to 98.89%, 98.33%, 98.33%. For the updated purity detection model of 2011 year, its prediction accuracy for 2010 new samples increased by 50.83%，54.58%，53.75% to 94.57%，94.02%，94.57% after adding 56 new samples from 2010 year. Meanwhile the effect of model updated by active learning algorithm was better than that of RS and KS. Therefore, the update for purity detection model of maize seeds is feasible by active learning algorithm.