YOLOv12 · Research Deployment
A minimal, research-driven AI system for early detection of Huanglongbing disease and nutrient deficiencies in citrus crops.
Leaf images are captured using mobile devices or field cameras under natural lighting conditions.
Images are resized to 416×416 and normalized before inference to ensure consistent detection quality.
YOLOv12s model analyzes visual features and localizes disease patterns in real time.
Detected class and confidence score are returned for decision-making and treatment planning.
The dataset was developed as part of an academic research project funded by the University Grants Commission (UGC) and conducted at Sylhet Agricultural University. A total of 3,000 manually annotated citrus leaf images were collected under controlled and field conditions, with an equal distribution across three classes.
Healthy citrus leaves exhibit uniform dark-green coloration across the lamina and veins, indicating adequate nutrient uptake and physiological balance. These images serve as the control class and baseline reference during model training and evaluation.
Zinc-deficient leaves display symmetrical interveinal chlorosis, reduced leaf size, and narrow leaf structure. This class is critical as zinc deficiency symptoms are often visually confused with citrus greening, leading to misdiagnosis in the field.
HLB-infected leaves show irregular blotchy mottling, asymmetrical chlorosis, and vein corking in advanced stages. These visual patterns were carefully annotated to enable the model to distinguish disease symptoms from nutrient deficiencies.
To ensure unbiased evaluation and strong generalization, the curated citrus leaf dataset was systematically divided into training, testing, and validation subsets following standard deep learning research practices.
| Subset | Percentage | Number of Images | Purpose |
|---|---|---|---|
| Training | 70% | 2,100 | Model learning and parameter optimization |
| Testing | 20% | 600 | Unseen data evaluation during development |
| Validation | 10% | 300 | Final performance assessment and model selection |
Leaf images collected from SAU orchards and nearby subdistricts under controlled and natural conditions.
Manual bounding-box annotation using LabelImg following YOLO text format standards.
Background normalization, image resizing (416×416), and dataset formatting.
YOLOv12s fine-tuned for 20 epochs using optimized hyperparameters and disk caching.
Performance measured using precision, recall, F1-score, and mAP metrics.
YOLOv12s is a single-stage object detection model optimized for real-time inference. Its attention-based backbone and multi-scale feature aggregation make it well suited for fine-grained agricultural disease detection tasks.
Extracts spatial and semantic features such as leaf texture, vein structure, and color variation.
Aggregates multi-scale features, enabling accurate detection of leaves with varying sizes and orientations.
Predicts bounding boxes, class probabilities, and confidence scores in a single forward pass.
Architecture: YOLOv12s
Image Size: 416 × 416
Epochs: 20
Optimizer: AdamW
IoU Threshold: 0.7
Research-backed technology designed for growers, researchers, and policymakers.