Spectrally optimised YOLOv10s-SeqOpt framework for real-time UAV-based early detection of avocado foliar diseases in indian orchards.

The YOLOv10s-SeqOpt framework achieved 96.0% accuracy in detecting avocado foliar diseases, demonstrating potential for real-time monitoring in orchards.

Summary and findings

This study evaluated a UAV-based multispectral imaging and deep learning detection framework for early detection of avocado foliar diseases in Indian orchards. The framework, YOLOv10s-SeqOpt, was tested on a dataset of 5,953 annotated images. It achieved 96.0% accuracy and 69.5ms real-time inference on NVIDIA Jetson Orin Nano.

Abstract

Avocado orchards in southern India experience annual yield losses of 20-40% due to foliar pathogens, including Anthracnose (Colletotrichum Gloeosporioides), Root Rot (Phytophthora Cinnamomi), Algal Leaf Spot (Cephaleuros Virescens) and Scab (Elsinoe Perseae). Conventional field scouting and RGB based imaging are unable to detect early-stage infections beneath dense canopy cover, delaying timely intervention and contributing to excessive fungicide use. To address these limitations, this study introduces a UAV based multispectral-RGB imaging pipeline paired with a deep learning detection framework built from the YOLO family of models (YOLOv5s-YOLOv10s) evaluated on a field-collected dataset of 5,953 annotated images comprising 35,422 bounding box instances across four foliar disease classes. A Sequential SGD-AdamW (SeqOpt) optimizer was designed to stabilize gradient dynamics multispectral noise by combining the global exploration capability of SGD in early training phases with the adaptive refinement precision of AdamW in later phases. As a result, SeqOpt achieved an 8% reduction in final loss and a 14% improvement in stability index compared to individual baseline optimisers, confirming superior convergence quality under spectrally heterogeneous training conditions. Systematic benchmarking across all six architectures demonstrated that YOLOv10s-SeqOpt delivered the highest detection performance, achieving 96.0% accuracy, F1-score = 0.911, and mAP@0.5 = 0.937 on multispectral validation set. YOLOv10s-SeqOpt consistently outperformed YOLOv5s-YOLOv9s across both RGB and multispectral datasets, with multispectral OCN input providing measurable detection advantages over RGB across five of six architectures evaluated The optimized model was quantized and deployed on two embedded AI platforms - NVIDIA Jetson Orin Nano, achieving real-time inference at 69.5ms per frame under PyTorch runtime, and the Raspberry Pi 5, supporting scheduled survey inference at 1.545s per frame via ONNX runtime - both without thermal throttling across extended evaluation sessions. The proposed framework demonstrates field-ready capability for large-scale orchard monitoring, bridging the gap between research-grade detection models and real-world precision agriculture deployment through a validated, scalable pipeline integrating multispectral imaging, deep learning and embedded edge computing for sustainable disease management in Indian horticulture.

Key findings

• Not reported in abstract.
• Not reported in abstract.
• Not reported in abstract.

Limitations

• Not reported in abstract.
• Not reported in abstract.

Elsewhere in the Adamax corpus