Development of a Vision-Guided Autonomous Variable-Rate Spraying System for Site- Specific Potato Disease Management Using YOLOv26

Problem Statement

Potato cultivation in India and globally faces significant yield losses due to foliar diseases, particularly late blight (Phytophthora infestans) and early blight (Alternaria solani). Conventional management relies on "blanket spraying," where fungicides are applied uniformly across the entire field regardless of disease presence. This approach is inefficient because disease incidence is inherently heterogeneous due to microclimate and canopy variations, leading to excessive chemical application on healthy plants and inadequate treatment on diseased ones. This indiscriminate use increases production costs, accelerates pathogen resistance, and poses environmental and health risks. While deep learning has advanced disease detection, a critical gap remains in translating real-time perception into field-level, autonomous variable-rate intervention.

Methodology

The study developed a unified perception–decision–actuation framework integrated into an autonomous three-wheeled sprayer (TWAS) operating at the Indian Institute of Technology Kharagpur.

1. System Architecture and Hardware

• Platform: A custom autonomous vehicle equipped with an Extended Kalman Filter (EKF) for navigation and a Pure Pursuit controller for path tracking.
• Computing: A distributed dual-processing architecture using two Raspberry Pi 5 units. Pi-1 handled navigation, localization, and primary row control, while Pi-2 acted as a dedicated perception node for the adjacent row. Communication was managed via UDP for low-latency data exchange.
• Sensing: Two downward-facing Logitech C270 RGB cameras were mounted at a height of 0.70 m, providing a field of view (FOV) of approximately 0.63 m × 0.36 m. This geometry ensured full coverage of the target row (0.30–0.35 m canopy width) while maintaining lateral isolation from adjacent rows.
• Actuation: Solenoid valves controlled by Pulse Width Modulation (PWM) allowed for variable-rate spraying. A camera-nozzle offset of 0.30 m was implemented, with a distance-based lag compensation algorithm using an inductive proximity sensor to synchronize spray actuation with vehicle speed.

2. Dataset and Model Development

• Dataset: A composite dataset of 1,561 field images containing 14,166 annotated leaf instances (7,348 diseased, 6,818 healthy) was created under dense canopy conditions.
• Model Selection: The study compared YOLOv8, YOLOv12, and YOLOv26 across multiple scales (nano to large) against conventional detectors (Faster R-CNN, RetinaNet, etc.).
• Training: Models were trained using the Ultralytics framework with CIoU loss, Binary Cross-Entropy (BCE), and Distribution Focal Loss (DFL). Data augmentation included flipping, scaling, and HSV adjustments, but Mosaic augmentation was excluded to maintain localization stability in dense canopies.
• Deployment: The lightweight YOLOv26n variant was selected for embedded deployment. The model was exported to ONNX format and optimized with FP16 precision for execution on the Raspberry Pi 5.

3. Control Logic and Severity Estimation

• Temporal Aggregation: To handle occlusion and detection noise, a 2-second temporal sliding window was used. Disease severity was estimated at the canopy level by aggregating the area of detected healthy and diseased regions across the window.
• Variable-Rate Strategy: The calculated severity index ($S(t)$) was mapped to discrete PWM duty cycles:
  • $S(t) < 0.10$: 40% duty cycle (Baseline).
  • $0.10 \le S(t) < 0.30$: 80% duty cycle.
  • $S(t) \ge 0.30$: 100% duty cycle.
  • Regions with insufficient foliage ($A_{total} < 3000$ pixels) received no spray (0%).

Key Results

1. Model Performance

• Accuracy: YOLOv26 achieved the highest detection performance with a peak mAP@0.5 of 0.961, outperforming YOLOv8 (0.924) by 3.7% and YOLOv12 (0.938) by 2.3%.
• Class Discrimination: YOLOv26 demonstrated superior class separation, with a normalized true positive rate of ~0.99 for both healthy and diseased classes, minimizing false negatives (missed diseases) and false positives (unnecessary spraying).
• Embedded Performance: On the Raspberry Pi 5, the optimized ONNX model achieved an inference speed of approximately 10 FPS. At an operating speed of 0.18 m/s, this resulted in a spatial sampling interval of 0.018 m, sufficient for continuous monitoring.

2. Field Experimentation

• Operational Validation: Field trials covered 598 spatial spray decision instances across six crop rows. The system successfully captured spatiotemporal heterogeneity, identifying high-severity zones in specific rows (e.g., Rows 1, 3, 5) while detecting low-severity zones in others.
• Correlation: A strong positive correlation ($r = 0.83$) was observed between the estimated disease severity and the commanded PWM duty cycle, confirming the system's ability to adapt spray intensity to local infection levels.
• Chemical Savings: Compared to a simulated conventional broadcast spraying baseline (constant maximum flow), the proposed variable-rate system reduced agrochemical consumption by 32.75%.
  • Conventional Rate: 458 L/ha.
  • Proposed System Rate: 308 L/ha.
  • Total volume reduced from 12.36 L to 8.30 L over the trial area.

Significance and Claims

The authors claim that this study successfully bridges the gap between intelligent disease perception and automated crop protection by integrating embedded deep learning with adaptive spray control. The significance of the work lies in:

1. Feasibility of Edge AI: Demonstrating that high-accuracy object detection (YOLOv26) can be deployed on low-cost, resource-constrained hardware (Raspberry Pi 5) for real-time field operations.
2. Precision Agriculture: Validating a closed-loop system that moves beyond simple detection to actionable, site-specific intervention, thereby reducing chemical inputs without compromising disease management.
3. Robustness in Dense Canopies: The use of temporal sliding windows and canopy-level severity estimation proved effective in handling the occlusion and variable illumination typical of mature potato crops.
4. Scalability: The framework is presented as a flexible platform that can be adapted to other row crops by retraining the detection model, offering a pathway toward sustainable, intelligent crop protection systems.

The paper concludes that the proposed perception–decision–actuation framework provides a practical technological solution for reducing agrochemical usage while maintaining effective disease control in potato cultivation.