RiceMamba-DETR: efficient rice disease detection via hybrid Mamba-transformer architecture with cross-scale feature aggregation and large-scale dataset

Abstract

Accurate rice disease detection in field imagery remains challenging because disease symptoms are often small, visually ambiguous, and embedded in cluttered scenes with overlapping leaves, variable illumination, and complex backgrounds. Existing CNN-based detectors offer fast inference but have limited capacity for long-range context modeling, whereas Transformer-based detectors capture global dependencies at a quadratic attention cost. Progress is further constrained by the lack of large-scale rice-specific detection benchmarks with expert bounding-box annotations. To address these challenges, we present RiceMamba-DETR, a hybrid Mamba-Transformer detector designed for efficient and context-aware rice disease localization. The model introduces Cross-Scale Feature Aggregation (CSFA), where Selective Feature Scanning (SFS) uses SS2D state-space scanning to capture bidirectional spatial context with linear sequence complexity at the scanning component, and Multi-Scale Context Aggregation (MSCA) strengthens lesion representations across scales. We further integrate a Separable Dynamic Decoder for adaptive localization and propose Hierarchical Focal Quality Loss to jointly handle class imbalance, localization quality, and scale variation. Alongside the model, we construct RiceDisease-55K, a large-scale rice disease detection dataset containing 55,070 images across 9 disease categories and 1 healthy category (Normal). Under our evaluation protocol, RiceMamba-DETR achieves 82.6% mAP@50 on RiceDisease-55K, 3.2 percentage points higher than YOLOv13-L, while running at 72.6 FPS on an RTX 4090 under the desktop profiling protocol. Supplementary evaluations yield 78.3% mAP@50 on PlantDoc and 99.2% mAP@50 on PlantVillage; the latter is interpreted as leaf-level localization under controlled laboratory conditions rather than lesion-level field detection. These results suggest that selective state-space modeling, when coupled with cross-scale aggregation and quality-aware supervision, is a promising design direction for practical rice disease detection.