基于聚类引导权重学习的高光谱苹果叶片病害图像波段筛选方法

Hyperspectral Band Selection for Apple Leaf Disease Images Based on Clustering-guided Weight Learning

摘要:

针对苹果叶片病害高光谱诊断中存在的数据波段冗余度高、计算负荷重以及现有波段筛选方法难以同时兼顾筛选效率、识别性能和波段物理意义等问题，传统特征工程方法虽然能够保留原始波段的物理意义，但对波段间结构相关性的利用不足，且在筛选效率和泛化能力方面存在局限;端到端深度学习方法虽能提升分类性能，却难以量化单波段对病害诊断的实际贡献，不利于筛选出具有物理可解释性的波段。为此，本文提出一种基于聚类引导权重学习的高光谱苹果叶片病害图像波段筛选方法，采用层次聚类对高光谱波段进行结构化分组，构建双分支卷积神经网络，分别学习波段的重构保真度权重与分类判别性权重，实现波段重要性的多维度定量评估;设计簇内自适应选择策略，从各聚类簇中筛选代表性核心波段。以苹果叶片5类病害高光谱数据为对象开展试验，结果表明，本文方法可从216个有效波段中筛选出9个核心波段，样本平均筛选耗时为15.2s，苹果叶片病害诊断任务分类准确率达98.38%。所筛选波段覆盖叶绿素吸收带、红边过渡区及近红外水分响应等病害敏感光谱区间，在显著压缩光谱维度的同时，能较好保留与病害识别相关的关键光谱信息。研究结果为轻量化病害诊断模型构建及专用多光谱传感器波段设计提供参考。

Abstract:

Aiming to address the issues of high band redundancy, heavy computational burden, and the difficulty of existing band selection methods in simultaneously balancing screening efficiency, recognition performance, and band physical interpretability in hyperspectral diagnosis of apple leaf diseases, a hyperspectral apple leaf disease image band selection method was proposed based on clustering-guided weight learning.Although traditional feature engineering methods can preserve the physical meaning of original bands, they insufficiently exploited the structural correlations among bands and had limitations in screening efficiency and generalization capability.End-to-end deep learning methods, while improving classification performance, struggle to explicitly quantify the actual contribution of individual bands to disease diagnosis, making it difficult to select bands with physical interpretability.To address these issues, the hierarchical clustering was firstly employed to structurally group hyperspectral bands.Secondly, a dual-branch convolutional neural network was constructed to learn the reconstruction fidelity weight and classification discriminability weight of bands respectively, achieving multi-dimensional quantitative assessment of band importance.Finally, an intra-cluster adaptive selection strategy was designed to select representative core bands from each cluster.Experiments conducted on hyperspectral data of five types of apple leaf diseases showed that the proposed method can select nine core bands from 216 effective bands, with an average screening time of 15.2s per sample, achieving a classification accuracy of 98.38% in apple leaf disease diagnosis.The selected bands covered disease-sensitive spectral intervals such as the chlorophyll absorption band, red-edge transition region, and near-infrared water response, demonstrating that the method can significantly compress spectral dimensionality while effectively retaining key spectral information relevant to disease identification, and can provide a reference for the construction of lightweight disease diagnosis models and the band design of dedicated multispectral sensors.