Toward smart agriculture: a hybrid mamba-transformer vision framework for plant disease detection

Plant disease detection under complex field conditions remains a critical challenge for precision agriculture due to varying illumination, scale variations, subtle lesion patterns, and inter-class visual ambiguity. This study proposes MAFusionNet, a disease-aware hybrid vision framework integrating Mamba and Transformer architectures, with components explicitly designed for plant disease-specific challenges. The MAFusion Mixer operates parallel CS-Mamba and self-attention branches to simultaneously capture sequential lesion boundary evolution and global diseasecontext spatial relationships. The CS-Mamba branch employs the SS2D-LS Block with twodimensional selective scanning and Local-Selective enhancement for linear-complexity longrange modeling while preserving 2D lesion morphology. The PConv operator uses asymmetric directional kernels forming cross-shaped receptive fields to capture anisotropic disease patterns such as vein-aligned blights and directional rust streaks. We constructed PD40, a large-scale dataset with 80,369 expert-verified annotated images across 40 disease categories spanning eight major crops, with inter-annotator agreement Cohen's κ = 0.874. Extensive experiments demonstrate that MAFusionNet achieves 94.7% mAP⁵⁰ and 81.8% mAP^50:95 on PD40, surpassing 25 state-of-the-art baselines including recent hybrid Mamba-Transformer detectors (CropMamba, HybridMamba, Mamba-DETR), with comprehensive ablation studies validating each component's non-redundant contribution. Edge deployment analysis on NVIDIA Jetson hardware demonstrates practical feasibility: the compressed MAFusionNet-T-Lite variant (8.7M parameters) achieves 89.3% mAP⁵⁰ at 18.4 FPS on Jetson Nano with 8.3W power consumption. The dataset and code are available at PD40-Dataset GitHub Repository.