UAV-based RGB and multispectral mango leaf disease detection with benchmarking of YOLOv5 to YOLOv10 and SeqOpt-optimised YOLOv8 for real-time edge deployment.

The article presents a real-time mango leaf disease detection framework with embedded edge deployment, using UAV-based multispectral imaging combined with optimised deep learning models. A custom dataset of 6,334 high-resolution RGB and multispectral images representing four common diseases was collected under natural orchard conditions using MAPIR RGB and multispectral OCN cameras mounted on a UAV. A controlled benchmarking of YOLOv5-YOLOv10 architectures was performed under identical training configurations. Although YOLOv10 achieved the highest detection accuracy, YOLOv8 offered a more favourable balance between detection performance and deployment efficiency on edge devices. To further enhance robustness and deployment suitability, the proposed SeqOpt method was applied to YOLOv8, improving the F1-score by 9.7%, mAP@50 by 8.6%, and mAP@50-95 by 21.7% compared to YOLOv10 trained under identical conditions on the multispectral validation data. In addition, on the Raspberry Pi 5 using ONNX inference, latency and energy consumption were reduced by up to 25% compared to the SGD-only YOLOv8 baseline. On the NVIDIA Jetson Orin Nano, the PyTorch model achieved 72 ms per-image inference latency, demonstrating near real-time capability. Overall, the proposed pipeline outperforms single-optimiser YOLOv8 (SGD-only and AdamW-only) and YOLOv10 baselines in detection accuracy and deployment efficiency, making it suitable for practical precision agriculture applications.