直接校正算法的柑橘溃疡病高光谱模型传递

Hyperspectral Model Transfer for Citrus Canker Detection Based on Direct Standardization Algorithm

针对目前所建立的柑橘溃疡病高光谱模型普适性差、预测精度低的问题,提出了基于不同仪器间高光谱模型传递来提高模型稳健性的方法。以脐橙52和卡拉卡拉红肉脐橙为研究对象,利用实验室高光谱成像平台(System 1,S1)和便携式高光谱成像仪(System 2,S2)采集了健康和染病柑橘的高光谱图像,建立了独立的柑橘溃疡病判别模型,并分析了不同预处理方法和判别模型对模型预测性能的影响。将S1和S2分别作为源机和目标机,利用直接校正算法对目标机获取的高光谱图像进行校正,分析模型传递前后的模型判别能力。结果表明,采用二阶导数预处理,极限学习机预测性能最佳,基于S1和S2检测的预测集识别率分别为97.5%和98.3%;以S1数据建立主模型,对经直接校正算法校正后的S2高光谱图像进行识别,预测集的识别率从校正前的38.1%提高到了86.2%。说明该方法可用于不同型号高光谱成像仪之间的定标模型传递,对于建立稳健可靠的柑橘溃疡病判别模型具有重要意义。

There is existence of poor universality and low prediction precision in citrus canker hyperspectral models in previous research .It is necessary to investigate an approach to improve the robustness of hyperspetral model transfer between different instruments which proposed to improve the robustness of the calibration model .Hyperspectral images of two different varieties including Navel Orange 52andCaraCarawere acquired using a laboratory hyperspectral imaging system （System 1 ,S1） and a portable hyperspectral imaging system （System 2 ,S2） .The discriminant models for the citrus canker detection were developed based on the images from S1 and S2 ,respectively ,and different pretreatment and classification methods were also investigated .Mean-while ,direct standardization （DS） algorithm was used to calibrate hyperspectral images collected by S2 which was considered as the slave while S1 as the master ,and the performance of the discriminant model were evaluated before and after the model transfer .It was shown that the best discriminant results were achieved by the extreme learning machine （ELM ） combined with the second-order derivative with the classification accuracies of 97.5% by S1 and 98.3% by S2 ,respectively .By using DS ,the clas-sification accuracy increased from 38.1% to 86.2% after the model transfer .It is demonstrated that the DS algorithm is useful for the calibration model transfer between different instruments ,which would be helpful for developing a robust method for the citrus canker detection .