Abstract: As a key link in the process of tobacco harvest, maturity measurement should be immediate, scientific and accurate. There are some problems in the current implementation of tobacco maturity recognition. For example, the measurement instruments are expensive and cumbersome to operate, which cannot be promoted for use in the field; the embedded image processing algorithms in phone cameras interfere with the effective features of images and the complex weather environment in the field affects the consistency of image acquisition; existing recognition algorithms ignore botanical domain information, which affects model accuracy and universality. Accordingly, this paper proposes a low-cost and effective method to predict the SPAD value and the maturity of fresh tobacco leaves in the field immediately, which ensures the quality of subsequent modulation of tobacco leaves by improving the prediction accuracy. Firstly, this paper develops a portable shooting device to achieve high-quality image acquisition in the field and proposes a segmentation method adapted to tobacco images in the field taking the CX-26 as the research object. After that, this paper extracts the feature data of the image target area and builds a SPAD value prediction model and maturity identification model in turn using the XGBoost algorithm. Then, a model integration idea for two models is proposed, which is able to use the strong correlation between SPAD value and maturity to improve maturity identification accuracy by predicting SPAD values. The method performs well in all evaluation indexes, where the mean absolute error reaches 0.470 3 for SPAD value prediction and macro F1-Score reaches 95.27% for maturity identification. Finally, this paper develops an APP to realize the transmission of tobacco images and prediction results between the portable shooting device and the model to achieve fast, objective and accurate immediate prediction of tobacco maturity in the field. The results can provide effective technical support for the accurate harvest of crops in the field.

Key words: maturity of tobacco leaves, SPAD value, machine learning, image features, immediate prediction

摘要： 成熟程度判定作为烟叶采收时的重要工作，需要满足即时性、科学性和准确性要求。目前烟叶成熟度识别实施存在测量仪器价格昂贵且操作繁琐，无法在田间推广使用；手机摄像头的内嵌图像处理算法干扰图像有效特征，田间复杂天气环境影响图像采集一致性；现存识别算法忽略植物学领域信息，影响模型准确性和普适性等问题。据此提出一种低成本且有效的田间即时鲜烟叶SPAD值预测和成熟度识别方法，通过提高识别准确率保障烟叶后续调制质量。研发便携式拍摄装置，实现田间高质量图片采集；并以CX-26烟叶品种为研究对象，提出一种适配于田间烟叶图像的分割方法，通过提取图像目标区域特征数据，利用XGBoost算法依次搭建鲜烟叶SPAD值预测模型和成熟度识别模型。提出两个模型的集成思路，集成模型能够利用SPAD值和成熟度的强相关关系，通过预测SPAD值提高成熟度识别准确性。该方法在各项评价指标中均表现优秀，其中SPAD值预测平均误差为0.470 3，成熟度识别宏F1-Score为95.27%。研发手机APP完成拍摄装置和云端模型之间烟叶图像和预测结果的传输，实现在田间对烟叶成熟程度快速、客观、准确的即时预测。该成果可为田间农作物精准采收提供有效技术支持。

关键词: 烟叶成熟度, SPAD值, 机器学习, 图像特征, 即时预测