Abstract: Continuous monitoring of disease- and pest-induced damage in rice canopies is fundamental for early detection, temporal dynamic analysis, and scientific pest and disease management. However, manual surveys are labor-intensive, satellite remote sensing is constrained by spatial resolution, and unmanned aerial vehicles are unsuitable for unattended long-term monitoring. In this study, a Pan-Tilt-Zoom (PTZ) camera-based monitoring platform was developed, and an illumination-robust detection algorithm, RiceDPD-YOLO, was proposed for automated detection of disease- and pest-induced damage in rice canopies under complex field conditions. The platform consists of three PTZ monitoring units and automatically collects rice canopy images at 30-min intervals for three consecutive years. Based on these long-term observations, a custom field dataset covering seven categories of major rice disease- and pest-induced damage was constructed, with samples collected across different years, growth stages, and illumination conditions. RiceDPD-YOLO is an improved YOLOv11-based detection model incorporating four task-oriented components to enhance illumination robustness, background suppression, multi-scale feature representation, and redundant bounding-box suppression. The saturation perception enhancement module (SPEM) exploits illumination-insensitive saturation information in the HSV (hue, saturation, value) color space to generate attention masks, thereby enhancing damage-related responses under variable illumination. The spatial and channel collaborative attention (SCCA) module uses the spatial distribution of damage regions to guide channel recalibration and suppress background interference. The position-enhanced dual-mode attention (PDMA) module introduces relative and dynamic position biases to model long-range dependencies and multi-scale spatial relationships of irregular damage regions. The cluster anchor fusion non-maximum suppression (CAF-NMS) method reduces duplicate detections caused by discontinuous damage distributions. Experimental results showed that RiceDPD-YOLO achieves a precision of 81.1%, a recall of 80.2%, an F1-score of 80.6%, an mAP₅₀ of 85.1%, and an inference speed of 6 frames per second. Compared with YOLOv11, RiceDPD-YOLO improves mAP₅₀ by 5.9 percentage points and outperforms representative models, including DINO and RT-DETR. The developed platform integrates scheduled image acquisition, cloud storage, automatic detection, historical retrieval, and statistical visualization, providing a practical foundation for long-term monitoring of rice canopies under complex field environments.

Key words: rice disease- and pest-induced damage, automatic monitoring, Pan-Tilt-Zoom (PTZ) camera, illumination-robust, YOLO model