TEA-LWNet: a lightweight multispectral feature enhancement network based on RGB-NIR fusion for detection of tea leaf diseases and pests.

To address the low contrast and mild blur in aerial tea plantation imagery, this study introduces an RGB-NIR dataset composed of UAV-acquired RGB and near-infrared images. The RGB-NIR representation provides complementary visible and near-infrared information, enhancing the contrast between damaged and healthy leaf regions while reducing interference caused by shadows and highlights. This improves the separability of small pest-damaged areas and early disease lesions in weakly textured tea canopy scenes. Based on this, we propose TEA-LWNet, a lightweight multi-scale detection framework. Its backbone, Multi-FENet, leverages a DS-MBS module for simultaneous high-order semantic expansion and low-level detail preservation. The neck features an Adaptive Enhancement (AE) module for spatial-semantic alignment and an MDFF structure with embedded Gated Fusion (GF) to suppress background diffusion during upsampling. Optimization is driven by SAB-Loss-combining IoU, Distributed Focal Loss, and BCE-enforcing scale-invariant geometric constraints and sub-pixel boundary calibration to prioritize small target learning. Evaluations on a self-built RGB-NIR dataset demonstrate that TEA-LWNet achieves an mAP@0.5 of 93.48% under strict parameter and FLOP constraints. Ablation studies confirm the functional complementarity of the proposed modules in cross-layer noise suppression and boundary refinement. Overall, TEA-LWNet achieves a favorable balance between detection accuracy and model complexity, indicating its potential for resource-constrained agricultural monitoring systems. However, its runtime performance on UAV or embedded edge hardware remains to be further validated.