码上扫一扫！ |
| 基于递归特征消除的烟草花叶病高光谱精准识别 |
|
盛鑫1, 郝兴安2, 杨照3, 李顺欣1, 刘星1, 冯明晨1, 毛锐1,4
|
|
1.西北农林科技大学 信息工程学院，陕西 杨凌 712100;2.西北农林科技大学 植物保护学院，陕西 杨凌 712100;3.陕西省烟草公司咸 阳市公司，陕西 咸阳 712000;4.陕西省农业信息智能感知与分析工程技术研究中心，陕西 杨凌 712100
|
|
| 摘要: |
| 【 目的】针对烟草花叶病毒（tobacco mosaic virus，TMV）早期侵染引发的微弱光谱变化难以有效捕捉 与可靠判别等问题，构建基于生理敏感特征优选与可解释性机器学习的TMV高光谱精准识别方法体系，为TMV的早期识别与精准防控提供技术支撑。【 方法】采集健康与TMV早期侵染烟叶的高光谱（400~1 000 nm）反射数据，经白板校正、感兴趣区域提取等预处理后，计算归一化色素叶绿素指数（normalized pigment chlorophyll index，NPCI）、水 分指数（water index，WI）等52个候选植被指数；采用递归特征消除（recursive feature elimination，RFE）筛选最优特征 子集；在此基础上构建并对比评估 K 近邻（K-nearest neighbors，KNN）、支持向量机（support vector machine，SVM）等4类机器学习模型的性能，结合随机搜索与网格搜索优化超参数，并利用独立测试集验证模型泛化能力；进一步引入沙 普利加性解释（shapley additive explanations，SHAP），对最优模型进行可解释性分析，量化关键特征对分类决策的贡 献。【 结果】经RFE筛选得到NPCI与WI双特征子集，证明色素与水分变化是TMV早期识别的关键信息，有效解决了高光谱数据的维数灾难问题。基于双特征子集构建的最佳识别模型KNN，其受试者工作特征曲线下面积（receiver operating characteristic-area under curve，ROC-AUC）、5折交叉验证准确率（accuracy，ACC）及F1分数（balanced F score，F1-score）分别为0.987，0.955和0.961；在独立测试集上，ROC-AUC仍达0.960，泛化性能良好。SHAP分析表明，NPCI与WI为驱动模型决策的核心因子，其值升高会正向推动样本被判定为“TMV侵染”，其变化与病毒侵染诱导的叶绿素降解及水分胁迫过程相一致。【 结论】仅利用NPCI与WI 2个植被指数即可实现对TMV的早期高精度 识别，显著降低了模型复杂度。所提出的“RFE特征优选+KNN+SHAP”方法体系在提升识别性能的同时，揭示了 “色素变化为主、水分变化为辅”的早期光谱响应机理，为TMV的早期快速识别及低成本、便携式烟草TMV监测装备 的研发提供了参考依据。 |
| 关键词: 烟草花叶病毒 早期识别 高光谱成像 植被指数 递归特征消除 SHAP 可解释性分析 |
| DOI：10. 13207/j. jnwafu. 2026. 09. 007 |
| 分类号: |
| 基金项目:中国烟草总公司陕西省公司 2024 年度科技项目（KJ-2024-06）；陕西省秦创原“科学家+工程师”队伍建设项目（2025QCYKXJ-070） |
|
| Hyperspectral precise detection of tobacco mosaic disease using recursive feature elimination |
|
SHENG Xin1, HAO Xing’an2, YANG Zhao3, LI Shunxin1, LIU Xing1, FENG Mingchen1, MAO Rui1,4
|
|
1.College of Information Engineering，Northwest A&F University，Yangling，Shaanxi 712100，China;2.College of Plant Protection ， Northwest A&F University，Yangling，Shaanxi 712100，China;3.Shaanxi Tobacco Company Xianyang City Company，Xianyang，Shaanxi 712000，China;4.Shaanxi Engineering Research Center of Agricultural Information Intelligent Perception and Analysis，Yangling，Shaanxi 712100，China
|
| Abstract: |
| 【 Objective】Aiming to address the challenge of effectively capturing and reliably discriminating the subtle spectral changes caused by early infection of tobacco mosaic virus（TMV），a precise hyperspectral re-cognition framework based on physiologically sensitive feature selection and interpretable machine learning was developed，to provide technical support for early TMV detection and precise control.【Method】Hyperspectral （400-1 000 nm） reflectance data of healthy and early TMV infected tobacco leaves were collected. After pretreatment with whiteboard correction and region of interest extraction，fifty-two candidate vegetation indices， including normalized pigment chlorophyll index（NPCI） and water index（WI），were calculated. Recursive feature elimination（RFE） was used to select the optimal feature subset. Based on the selected subset，four categories of machine-learning models，including k-nearest neighbors（KNN） and support vector machine（SVM），were developed and comparatively evaluated. Hyperparameters were optimized using random search and grid search，and model generalization performance was assessed using an independent test set. Furthermore，shapley additive explanations（SHAP） was introduced to interpret the best-performing model by quantifying the contribution of key features to classification decisions.【Result】A dual-feature subset consisting of NPCI and WI was obtained through RFE screening，demonstrating that pigment and water-related changes are key information for early TMV identification，which effectively mitigates the curse of dimensionality in hyperspectral data. The KNN model built on the dual-feature subset achieved the best performance，with receiver operating characteristic-area under curve（ROC-AUC） of 0.987，five-fold cross-validation accuracy（ACC） of 0. 955 and balanced F score （F1-score） of 0.961. On the independent test set，the ROC-AUC remained as high as 0.960，indicating good generalization performance. SHAP analysis further showed that NPCI and WI were the dominant drivers of model decision-making：an increase in their values positively drove samples to be classified as“ TMV-infected”，and its variation was consistent with TMV-induced chlorophyll degradation and water-stress processes.【Conclusion】 Early and precise identification of TMV can be achieved using only two vegetation indices（NPCI and WI），substantially reducing model complexity.The proposed“ RFE-based feature selection+KNN+SHAP” methodological framework not only improves identification performance but also reveals an early spectral response mechanism dominated by pigment changes and supplemented by water status variations.This provides a reference for the early and rapid identification of TMV，as well as for the development of low-cost，portable TMV monitoring equipment for tobacco. |
| Key words: tobacco mosaic virus early detection hyperspectral imaging vegetation indices recursive fea-ture elimination shapley additive explanations analysis |
