Mamba-YOLO: A Hybrid Architecture with Linear-Complexity Selective Scan Mechanisms for Enhanced Microscopic Rice Disease Detection

Abstract

Abstract While automated visual inspection facilitates large-scale crop disease management, its deployment in field environments remains challenging. The morphological similarity of early-stage symptoms, combined with severe canopy occlusion, frequently degrades model accuracy. When applied to these unconstrained datasets, standard lightweight Convolutional Neural Networks (e.g., the YOLOv5-v11 series) tend to overfit, yielding an accuracy of only around 46% mAP@0.5 on microscopic targets. Even advanced NMS-free architectures like YOLO26 struggle to capture the long-range spatial dependencies required to decouple highly ambiguous, discrete lesions like early-stage Rice Blast.We address this fundamental structural limitation by introducing Mamba-YOLO. This hybrid architecture integrates the Visual State Space Model (VMamba) directly into a lightweight YOLO26-Nano baseline. We replaced standard deep bottleneck layers with Visual State Space (VSS) modules, leveraging a Selective Scan Mechanism (SSM) to model global context with linear computational complexity (𝑂(𝑁)). Our network perceives fragmented pathological features across the entire image without the massive quadratic computational burden typical of Vision Transformers.Our empirical evaluations on a multi-class rice disease dataset yield compelling insights. Mamba-YOLO achieves a highly competitive overall mAP@0.5 of 92.36%, performing on par with the heavily optimized pure-CNN baseline (92.41%). More importantly, under the strictly penalized mAP@0.5:0.95 metric, our architecture establishes a new peak of 55.6%. We recorded a critical +0.9% accuracy breakthrough for Rice Blast, the most challenging microscopic category. Beyond static accuracy, analysis of the training dynamics proves that the selective scan mechanism acts as a robust global regularizer, effectively collapsing the massive generalization gap that plagues traditional lightweight detectors.We achieve these structural breakthroughs with near-zero overhead. Mamba-YOLO maintains an ultra-low computational footprint of 5.9 GFLOPs and requires only 2.69 million parameters. This Pareto-optimal balance positions our architecture as a highly robust, field-ready solution for deploying high-precision diagnostics on resource-constrained agricultural edge devices.

Plant phenotyping relevance

イネ病害の症状・病斑を画像から検出する新規深層学習アーキテクチャを開発し、複数の評価指標で性能検証しているため、植物表現型取得手法が中心である。

Code and data availability

The paper describes a curated rice disease dataset (1,860 images) and a custom Mamba-YOLO architecture, but no public dataset URL, code repository, or trained model checkpoint is provided. The data availability statement only offers data on request, and the only URLs cited (Ultralytics, Mamba, MambaVision) are generic/

No evidence-backed public reproduction asset is currently recorded.