A lightweight RPB-YOLO11-based detector improves mobile phenotyping of rice panicle blast

Abstract

Rice panicle blast detection is an important task in plant disease phenotyping. Field-based detection remains challenging because infected spike regions are often small, sparse, elongated, and affected by overlapping panicles, complex backgrounds, and variable illumination. In this study, we propose RPB-YOLO11, a lightweight YOLO11-based detector designed for rice panicle blast detection. The model uses a Lightweight Ghost Backbone (LGB) to reduce redundant computation. It uses Anisotropic Axial Stripe Attention (A2SA) to represent elongated panicle structures. It also uses Focal Multi-Scale Attention (FMSA) for multi-scale feature refinement and Adaptive Geometric Shape IoU (AGS-IoU) for geometry-aware localization. The model was trained and evaluated on a rice panicle image dataset containing 1,055 training images, 69 validation images, and 169 test images. On the test set, RPB-YOLO11 achieved 76.09% mAP50, 45.44% mAP50-95, 73.98% precision, and 72.75% recall with 6.21 GFLOPs. Compared with the YOLO11n baseline, it improved mAP50, mAP50-95, precision, and recall by 2.73, 2.12, 1.64, and 2.11 percentage points, respectively. An Android-oriented inference application supports local image inference, detection visualization, class counting, and diseased-panicle incidence estimation. These results suggest that RPB-YOLO11 provides a practical approach for image-based rice panicle blast survey.

Code and data availability

The paper links a public Hugging Face dataset used to establish the rice panicle blast detection dataset and a public GitHub release (data availability statement) containing the study's datasets/models.