基于TabPFN的分数阶高光谱橡胶树叶片白粉病诊断研究

doi:10.3964/j.issn.1000-0593(2025)12-3332-10

摘要
参考文献
相关文章 (15)

全文: PDF (7462 KB)
输出: BibTeX | EndNote (RIS)

摘要：白粉病(PM)是影响橡胶树健康生长和天然橡胶产量的常见叶部病害之一，快速准确地进行病害诊断分级对实施精准防控和保障天然橡胶产量具有重要意义。利用高光谱成像技术对海南橡胶种植区的染病叶片进行检测分析，采集了不同病害等级的橡胶叶片样本，通过高光谱成像设备获取965.4～1 668.0 nm波段范围的反射光谱数据。高光谱数据存在噪声和信息冗余，采用支持向量机(SVM)、随机森林(RF)、多层感知机(MLP)三种传统模型和具有自动加权特征能力的表格化先验数据拟合网络(TabPFN)对原始光谱及Savitzky-Golay平滑、标准正态变换(SNV)、分数阶微分(FOD)预处理后的全波段数据进行建模分析，通过多模型评估筛选出最优光谱校正方法。为了验证TabPFN的特征加权能力，基于最优预处理数据，分别采用主成分分析(PCA)、ReliefF、最大相关性最小冗余(mRMR)及HSICLasso算法进行特征选择，提取与白粉病分级相关的敏感波段。对比全波段与特征子集在不同分类模型上的性能差异，确定最优模型架构。最后采用SHAP可解释性分析方法，揭示最优模型中关键特征波段对白粉病分级的主要影响。结果表明：在所有诊断模型中TabPFN均能达到最优，展现其较高的鲁棒性和较好的特征加权选择能力。其中，FOD预处理能有效处理光谱噪声并放大有效细节特征，提升数据质量的效果最优，基于该方法的全波段TabPFN模型分类准确率达95.27%，较传统方法提升3.24%~13.24%，HSICLasso筛选20个关键特征后准确率仍保持94.31%，模型复杂度降低近90%，精度仅降低1.01%。SHAP分析显示1 160和1 400 nm附近波段为关键判别区域，该波段区域与C—H、O—H化学键振动高度相关，对应叶片碳水化合物、木质素以及水分含量，表明该模型能够有效捕捉由白粉病引起的叶片理化参数变化所导致的光谱响应特征。验证了FOD结合TabPFN算法检测白粉病的可行性，所建模型可精准判定病害等级，为胶园精准施药、保障橡胶树健康以提升橡胶产量提供参考。

关键词：高光谱成像；橡胶树白粉病；分数阶微分；机器学习；可解释性分析

Abstract：Powdery mildew (PM) is a common foliar disease that negatively impacts the health of rubber trees and the yield of natural rubber. Rapid and accurate disease diagnosis is essential for implementing precise control measures and ensuring optimal rubber production. This study employed hyperspectral imaging technology to analyze infected leaves in Hainan rubber plantations. Samples of rubber leaves at various infection levels were collected, and hyperspectral reflectance data ranging from 965.4 to 1 668.0 nm were obtained using hyperspectral imaging equipment. The hyperspectral data contained noise and redundant information. Three traditional models, namely Support Vector Machine (SVM), Random Forest (RF), and Multi-Layer Perceptron (MLP), as well as the Tabular Prior Data Fitting Network (TabPFN), which incorporates automatic feature weighting, were used to model and analyze both the raw spectral data and the full-band data preprocessed by Savitzky-Golay smoothing, Standard Normal Variate (SNV), and Fractional Order Differentiation (FOD). A multi-model evaluation identified the optimal spectral preprocessing method. To assess the feature weighting capability of TabPFN, Principal Component Analysis (PCA), ReliefF, Maximum Relevance Minimum Redundancy (mRMR), and HSICLasso algorithms were employed for feature selection, extracting sensitive bands associated with powdery mildew grading. The performance of the whole band and feature subsets was compared across different classifiers to determine the optimal model architecture. Finally, Shapley Additive Explanationswas used to analyze the key features and their influence on disease grading. The results showed that TabPFN outperformed all other models, demonstrating superior robustness and effective feature weighting selection. FOD preprocessing effectively reduced spectral noise and enhanced the extraction of essential detail features, resulting in the highest data quality improvement. The full-band TabPFN model with FOD preprocessing achieved a classification accuracy of 95.27%, surpassing traditional methods by 3.24%～13.24%. After applying HSICLasso to select 20 critical features, the accuracy remained at 94.31%, while reducing model complexity by nearly 90% and only decreasing accuracy by 1.01%. SHAP analysis identified the 1 160 nm and 1400 nm regions as key discriminatory bands, linked to C—H and O—H chemical bond vibrations. These bands correspond to the leaf's carbohydrate, lignin, and water content, indicating the model's ability to capture spectral responses related to physicochemical changes caused by powdery mildew. This study validates the integration of FOD and TabPFN for PM detection, providing an accurate model for assessing disease severity, which can aid in precise pesticide application and promote the health of rubber trees, ultimately improving rubber production.

Key words：Hyperspectral imaging; Rubber tree powdery mildew; Fractional order differential; Machine learning; Interpretability analysis

收稿日期: 2025-05-05 修订日期: 2025-09-17

中图分类号:

O433.4

基金资助: 国家自然科学基金项目(32560418，32572204)，海南省重点研发计划项目(ZDYF2022XJS008)资助

通讯作者: 唐荣年 E-mail: rn.tang@hainanu.edu.cn

作者简介: 胡文锋，1973年生，海南大学机电工程学院副教授 e-mail: 990531@hainanu.edu.cn

引用本文:

胡文锋，陈周洋，李创，罗小川，赵永臣，何勇，唐荣年. 基于TabPFN的分数阶高光谱橡胶树叶片白粉病诊断研究[J]. 光谱学与光谱分析, 2025, 45(12): 3332-3341.
HU Wen-feng, CHEN Zhou-yang, LI Chuang, LUO Xiao-chuan, ZHAO Yong-chen, HE Yong, TANG Rong-nian. Research on Fractional-Order Hyperspectral Diagnosis of Rubber Tree Leaf Powdery Mildew Based on TabPFN Model. SPECTROSCOPY AND SPECTRAL ANALYSIS, 2025, 45(12): 3332-3341.

链接本文:

https://www.gpxygpfx.com/CN/10.3964/j.issn.1000-0593(2025)12-3332-10 或 https://www.gpxygpfx.com/CN/Y2025/V45/I12/3332

[1] SUN Yong-shuai, TIAN Wei-min, ZHAI De-li, et al(孙永帅, 田维敏, 翟德利, 等). Bulletin of Chinese Academy of Sciences(中国科学院院刊), 2024, 39(1): 191.
[2] LI Bo-xun, LIU Xian-bao, CHEN Li-qiong, et al(李博勋, 刘先宝, 陈丽琼, 等). Scientia Sinica Vitae(中国科学: 生命科学), 2024, 54(10): 1798.
[3] Zeng T, Li C, Zhang B, et al. Frontiers in Plant Science, 2022, 13: 829479.
[4] Bao D, Zhou J, Bhuiyan S A, et al. Journal of Agriculture and Food Research, 2024, 18: 101369.
[5] BAI Da-yu, SHI Qing-hua, WANG Jian-quan, et al(白大昱, 史庆华, 王建全, 等). Journal of Chinese Agricultural Mechanization(中国农机化学报), 2024, 45(11): 172.
[6] BAO Hao, HUANG Li, ZHANG Yan, et al(鲍浩, 黄莉, 张艳, 等). Progress in Biochemistry and Biophysics(生物化学与生物物理进展), 2025, 52(2): 513.
[7] Xuan G, Li Q, Shao Y, et al. Computers and Electronics in Agriculture, 2022, 197: 106921.
[8] Cheng X, Feng Y, Guo A, et al. Remote Sensing, 2024, 16(1): 105.
[9] Cheng X, Huang M, Guo A, et al. Remote Sensing, 2024, 16(9): 1634.
[10] MAMAT Sawut, LI Wu-yao, CUI Jin-tao, et al(买买提·沙吾提, 李武耀, 崔锦涛, 等). Cotton Science(棉花学报), 2024, 36(4): 296.
[11] ZHANG Yun, MAO Hong, FANG Hao-fan, et al(张运, 茆红, 方浩帆, 等). Remote Sensing Information(遥感信息), 2024, 39(5): 20.
[12] Chen K, Li C, Tang R. Industrial Crops and Products, 2017, 108: 831.
[13] Zhang J, Jing X, Song X, et al. Computers and Electronics in Agriculture, 2023, 206: 107671.
[14] Rogers M, Blanc-Talon J, Urschler M, et al. Journal of Food Measurement and Characterization, 2023, 17(6): 6039.
[15] WU Meng-hong, DOU Sen, LIN Nan, et al(吴梦红, 窦森, 林楠, 等). Spectroscopy and Spectral Analysis(光谱学与光谱分析), 2025, 45(1): 204.
[16] Nguyen D H D, Tan A J H, Lee R, et al. Pest Management Science, 2025, 81(1): 149.
[17] Su J, Liu C, Coombes M, et al. Computers and electronics in agriculture, 2018, 155: 157.
[18] Swaraj S, Aparna S. Infrared Physics and Technology, 2024, 142: 105522.
[19] Hollmann N, Müller S, Purucker L, et al. Nature, 2025, 637(8045): 319.
[20] Verma B, Prasad R, Srivastava P K, et al. Computers and Electronics in Agriculture, 2022, 192: 106581.
[21] Lundberg S M, Lee S I. Proceedings of the 31st International Conference on Neural Information Processing Systems, 2017, 4768.
[22] Zhao Z, Anand R, Wang M. Maximum Relevance and Minimum Redundancy Feature Selection Methods for a Marketing Machine Learning Platform, 2019 IEEE International Conference on Data Science and Advanced Analytics (DSAA). IEEE, 2019: 442.
[23] CHENG Feng-wei, WANG Wen-jian, ZHANG Zhen-zhen(程凤伟, 王文剑, 张珍珍). Journal of Nanjing University (Natural Sciences)[南京大学学报(自然科学版)], 2023, 59(6): 928.
[24] Yamada M, Jitkrittum W, Sigal L, et al. Neural computation, 2014, 26(1): 185.
[25] Curran P J. Remote Sensing of Environment, 1989, 30(3): 271.
[26] HU Peng-fei, TANG Rong-nian, HU Wen-feng, et al(胡鹏飞, 唐荣年, 胡文锋, 等). Jiangsu Agricultural Sciences(江苏农业科学), 2024, 52(18): 222.