Research on the BiSpikeNet classification model for tea plant pests and diseases.

Background: Fast and accurate identification of tea pests and diseases with lightweight, efficient, and resource-friendly computation is a key challenge in smart tea gardens. While deep learning models like ResNet extract strong features, their high computational cost limits application. Conversely, lightweight models like BiRealNet reduce complexity but suffer from unclear intermediate features and weak discrimination. Methods: To address these limitations, we propose BiSpikeNet, a lightweight, high-efficiency classification model integrating Convolutional Neural Networks (CNNs) and Spiking Neural Networks (SNNs) under a "Diagnostic" design paradigm. We introduced an attention mechanism to enhance key feature representation and optimized residual structures. Furthermore, a cross-modal fusion framework was designed, combining a binary SNN meta-network (BiSNN) with a pulse-based residual block (SNNMultiShortcutBlock) to capture temporal and spatial features. Results: The proposed model achieved 92.30% accuracy with fewer parameters. With the cross-modal fusion framework, the accuracy improved to 99.92%. Compared with existing lightweight models such as TLDDM and LiSA-MobileNetV2, our method demonstrates superior performance with lower computational costs. Conclusion: Experimental results demonstrate that BiSpikeNet provides an effective balance between accuracy and efficiency, offering a robust solution for lightweight, resource-efficient tea pest and disease detection. Code is available at https://github.com/huyalin070-svg/BiSpikeNet.