GuavaMamba‐YOLOv11: A State‐Space Enhanced Lightweight Detection Network for Early Guava Disease Prediction

ABSTRACT

Early‐stage guava leaf diseases are usually present in the form of micro‐lesions with low contrast, irregular boundaries and high inter‐class similarity, so the detection of early‐stage diseases in complex orchard environments is very difficult. Conventional lightweight object detectors show low ability to model long‐range dependencies and cross‐scale feature interactions, leading to missed detections in the critical stage of infection. To overcome the limitations of these methods, this research proposes a State‐Space Enhanced YOLOv11 framework, which combines the Selective State Space Model (Mamba) to restore the continuity of global contextual features, the cross‐scale dynamic attention mechanism to refine the feature of different scales and the micro‐lesion detection head responsible for detecting micro‐lesion to optimise by Wise‐IoU v3 loss. Experimental evaluation on a field‐acquired guava disease dataset containing 14,800 images and eleven categories shows that the proposed model achieves 97.1% mAP@0.5, 77.6% mAP@0.5:0.95, 96.4% precision, 94.8% recall and 95.6% F1‐score. Compared with the baseline YOLOv11, GuavaMamba‐YOLOv11 improves mAP@0.5 by 4.7%, mAP@0.5:0.95 by 9.5%, recall by 5.9% and small‐lesion AP by 11.5%, while maintaining lightweight deployment with 3.8 M parameters, 8.9 GFLOPs and 130 FPS. These improvements are significant because higher recall and small‐lesion AP reduce missed early infections, supporting timely disease intervention and real‐time orchard monitoring. These results validate that the integration of state‐space modelling with dynamic attention is a useful method to improve the sensitivity of micro‐lesions and the practical deployment for monitoring guava disease in smart agriculture systems.