Answers are grounded in this paper and cite their sources.
• A lightweight YOLOv9t-DyE + MobileSAM framework was developed for field-scale detection and quantification of the chilli leaf curl complex caused by ChiLCV and mite infestation. • The YOLO architecture integrates Efficient Channel Attention (ECA), DySample, and MPDIoU loss, enhancing feature representation, spatial detail preservation, and localization accuracy. • A semi-supervised labeling pipeline expanded the dataset from 2275 to 5519 images, significantly reducing manual annotation effort while improving label consistency. • The proposed model achieved strong performance (Precision: 0.769±0.007, Recall: 0.748±0.004, mAP50: 0.813±0.005, and mAP50–95: 0.45±0.004. This represents relative improvements of 3.17 % in Precision, 3.63 % in Recall, 3.09 % in mAP50, and 3.26 % in mAP50–95) with ∼18.9 % faster inference than the YOLOv9t baseline. • Integration with the Segment Anything Model (SAM) enabled pixel-level severity estimation and grading of virus- and mite-induced symptoms, providing a practical tool for precision pest–disease monitoring. For chilli leaf curl complex accurate field-scale assessment is challenging because both chilli leaf curl virus (ChiLCV) and mite infestation generate visually similar curling symptoms leading to frequent misdiagnosis during routine scouting. Such overlap, coupled with reliance on manual scoring, often causes inconsistent severity estimation and improper management decision. To address this, we propose a lightweight deep learning framework for reliable detection and quantification of chilli leaf curl complex. Our model, termed YOLOv9t-DyE, incorporates Efficient Channel Attention (ECA), DySample, and MPDIoU loss and is trained using a semi-supervised labeling workflow. Under five-fold cross-validation, it has achieved a strong detection performance with a Precision: 0.769±0.007, Recall: 0.748±0.004, mAP50: 0.813±0.005, and mAP50–95: 0.45±0.004. This represents relative improvements of 3.17% in Precision, 3.63% in Recall, 3.09% in mAP50, and 3.26% in mAP50–95 ( p < 0.05), along with ∼18.9% faster inference and negligible impact on model size or computation compared to the baseline YOLOv9t. Predictions are further processed using variants of the Segment Anything Model (SAM) and after comparative evaluation the lightweight MobileSAM model is selected as the recommended option for integration into the pipeline. The framework is computationally efficient and deployable on mobile or embedded devices for field surveillance. Overall, this integrated pipeline provides a robust, scalable, and interpretable solution for distinguishing visually confounding stresses and delivering precise severity estimation. This capability facilitates timely and data-driven pest-disease diagnostics for sustainable management while offering a field-ready practical tool for deployment across large agricultural fields.
No takes yet. Share an insight, caveat, or question.
Paul et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: