基于深度迁移学习的木虱取食行为识别方法

Deep Transfer Learning-based Approach for Recognition of Psyllid Feeding Behaviors

摘要:

刺吸电位图谱技术 (Electricalpenetrationgraph,EPC) 是解析木虱取食行为的重要手段。然而，目前对 EPC 信号的分析仍严重依赖人工判读，效率低下。因此，本研究提出一种基于深度迁移学习的木虱 EPC 自动识别方法，融合 Transformer 与双向门控循环单元 (BiCRU) 以提升模型在有限样本下的识别性能。首先对原始 EPC 信号进行变分模态分解 (VMD) 以降低噪声干扰，随后构建基于 Transformer 编码器与 BCRU 的混合模型，充分利用其全局感知与局部时序建模能力，实现波形特征的协同提取。为缓解木虱样本不足的问题，先在规模大的虫 EPC 数据集上构建预训练模型，然后通过领域自适应策略迁移至目标木虱数据集，并通过微通机制完成跨物种知识迁移。结果表明，当单类别样本量为 40 且采用全参数微调策略时，模型识别准确率可达到 98.50%; 而在资源受限条件下 (单类别样本量 20 且仅微调末 2 层), 模型准确率仍能保持 97.54%。本研究提出的融合迁移学习技术与深度学习特征提取方法的思路，可用于木虱取食行为的自动分析，也可为其他刺吸式昆虫 EPC 信号的分析提供参考。

Abstract:

The electrical penetration graph (EPG) technique serves as a critical tool for analyzing psyllid feeding behaviors, investigating psyllid-citrus host interactions, and exploring pathogen transmission mechanisms. Nevertheless, the interpretation of EPG signals still heavily relies on manual effort, which remains inefficient and labor-intensive. To address this issue, an automated EPG recognition method was proposed based on deep transfer learning, which integrated a Transformer and a bidirectional gated recurrent unit (BiGRU) to enhance recognition performance under limited sample availability. The proposed method began with variational mode decomposition ( VMD) to suppress noise in raw EPG signals, followed by a hybrid model, combining a Transformer encoder and a BiGRU network. This architecture leveraged the Transformer's global perception and the BiGRU's local temporal modeling capability to collaboratively extract discriminative waveform features. To mitigate the scarcity of labeled psyllid data, the model was firstly pre-trained on a larger aphid EPG dataset and then transferred to the target psyllid dataset via a domain adaptation strategy. Cross-species knowledge transfer was accomplished through a fine-tuning mechanism. Experimental results demonstrated that the model achieved a recognition accuracy of 98. 50% when using 40 samples per category with full-parameter fine-tuning. Under resource-constrained conditions (20 samples per category with only the last two layers fine-tuned), the model still maintained an accuracy of 97.54%. The research result demonstrated the viability of integrating transfer learning with deep learning-based feature extraction for automated analysis of psyllid feeding behavior, while also offering a valuable reference for analyzing EPG signals in other piercingsucking insects.