FCMamba: A frequency-enhanced state space network for edge-based plant disease recognition

Accurate plant disease recognition on edge devices requires models that can preserve fine-grained lesion details, capture long-range spatial dependencies, and maintain low computational cost. However, existing lightweight networks often sacrifice high-frequency texture information for efficiency, whereas global modeling architectures usually improve contextual representation at the cost of increased parameters and computational overhead. To address this accuracy-efficiency dilemma, this study proposes FCMamba, a frequency-enhanced visual state space network for efficient fine-grained plant disease recognition. The proposed framework is built upon a Frequency-Enhanced Visual State Space (FC-VSS) module, in which a two-dimensional Selective Scan (SS2D) branch is used to model long-range spatial dependencies and a wavelet transform convolution branch is introduced to enhance frequency-sensitive lesion textures. These two complementary representations are further integrated through spatially guided adaptive fusion, enabling the model to strengthen disease-related structural and textural cues while suppressing background interference. In addition, a dual-task training strategy is designed by using image reconstruction as an auxiliary supervision signal, which encourages the encoder to preserve lesion morphology and structural information during training without increasing inference cost. Extensive experiments were conducted on the PlantVillage benchmark, a fine-grained disease confusion setting, a real-world field dataset, and a cross-crop apple dataset. FCMamba achieves 99.20% accuracy on PlantVillage, 81.50% accuracy on field images, and 89.69% accuracy on the apple dataset. The fine-grained confusion analysis further shows that the proposed method improves the discrimination of visually similar disease categories, while field and cross-crop experiments demonstrate its generalization ability under complex practical conditions. Compared with representative Transformer- and state space model (SSM)-based baselines, FCMamba reduces the number of parameters and FLOPs by up to approximately 76% and 78%, respectively. These results indicate that FCMamba achieves a favorable balance between recognition accuracy, model compactness, and computational efficiency, providing an effective solution for edge-based fine-grained plant disease recognition.