基于冠层图像煤污识别的温室白粉虱危害分级方法

Method for Greenhouse Whitefly Damage Grading Based on Sooty Mold Recognition in Canopy Images

摘要:

针对温室白粉虱成虫体型微小且具有隐蔽性，导致冠层监控难以实现直接监测的问题，利用其诱发的煤污病征进行危害程度识别具有重要意义。然而，由于不同作物背景本底颜色空间的重叠干扰，以及卷积神经网络易捕获背景非因果特征导致跨场景泛化难题，本文提出了一种基于冠层图像煤污识别的温室白粉虱危害分级方法。确立了以煤污面积比与视觉显著度为核心的4级危害指数（HI-0~HI-3）评价标准，并构建了覆盖多物候期与异质背景的专家标注数据集，以此提出一种颜色先验引导的双通道融合分级架构（CPG-Fusion）。以ResNet-18为骨干网络，通过嵌入颜色先验注意力模块（CPAM），将具备亮度不变性的逐像素相对暗度转换为空间注意力权重，通过残差调制方式强制模型特征向煤污病征区域聚焦，从而抑制背景信号干扰。试验结果显示，在独立温室的跨场景评估中，集成CPAM的模型将二次加权Kappa（Quadratic weighted Kappa, QWK）从ResNet-18基线的0.401提升至0.551，且所有误判均严格限制在相邻等级内。横向试验证实，本方法在保持高精度的同时实现了本组对比中最低的折间标准差，可靠性优于主流自学习注意力机制。梯度加权类激活映射（Grad-CAM）可视化证实，模型特征激活区已与真实煤污物理掩膜实现空间对齐。研究结果证明了物理约束在提升定性识别到定量分级中的关键作用，为设施蔬菜精准植保提供了低成本、高鲁棒性的自动监测方法。

Abstract:

Aiming to address the cross-scene generalization deficit of end-to-end CNN classifiers on the task of greenhouse whitefly damage grading in facility vegetable production, where standard models tended to fix on background-correlated cues such as illumination and plant growth stage rather than the physical signature of sooty mold, a color-prior-guided two-channel grading method (CPG-Fusion) was proposed.The core of the method was a color prior attention module (CPAM): the per-image relative darkness, computed as a scene-brightness-independent physical cue, was fed as the attention input signal of a ResNet-18 backbone so that the CNN was consistently guided toward the region that directly corresponded to sooty mold coverage and was steered away from background spurious features.A dataset of 260 images with 4-level hazard index annotations was collected from two solar greenhouses in Changping, Beijing;the CNN was trained under a 5-fold stratified cross-validation protocol, and the deployed prediction was obtained by softmax probability averaging across folds.On 58 cross-scene OOD images independently collected in a second greenhouse (GH2), CPAM raised the quadratic weighted Kappa (QWK) from 0.401 of the ResNet-18 baseline to 0.551, improved the recall of the light-damage level by 22 percentage points, and confined all misclassifications to adjacent levels.In the same-protocol horizontal comparison with four representative attention modules (SE-Net, CBAM, ECA and Coordinate Attention), CPAM matched the upper bound of accuracy of the four counterparts on the OOD ensemble while attaining the lowest across-fold standard deviation among the four self-attention modules, indicating better single-model reliability under small-sample weak supervision.Grad-CAM visualization and the near-zero marginal gain of late fusion between CPAM and the pixel prior channel jointly confirmed that the pixel color prior was internalized into the CNN feature map.The proposed method showed that injecting interpretable physical cues into the attention pathway can improve the robustness of deep learning models under complex agricultural environments, and can be directly deployed on existing canopy surveillance cameras without additional hardware or target-scene parameter tuning.