Weed Detection in Crop Fields using YOLO

Weeds compete with crops for water, nutrients and light, and the standard response — broadcast herbicide spraying — wastes chemical, money and soil health on empty ground. Precision agriculture flips this: detect exactly where the weeds are, and spray only there. That needs a detector that runs in real time on a sprayer or drone, which is exactly what YOLO was designed for. The DeepWeeds dataset (17,509 images, 8 weed species from Australian rangelands) is the standard benchmark, and this project fine-tunes YOLOv8 on it with documented augmentation, evaluates with mAP@0.5 overall and per species, and wraps the detector in an interactive web demo: load a field capture, see weed boxes with confidence bars, and get a zone-wise spray map showing which zones need treatment and which can be skipped. The herbicide-saving arithmetic gives the project a compelling real-world story.

1. DeepWeeds images are organized with YOLO-format annotations across the 8 weed species plus the negative (weed-free) class.
2. YOLOv8n pre-trained on COCO is fine-tuned at 640×640 with mosaic augmentation for 150 epochs.
3. Validation mAP@0.5 is tracked each epoch; the best checkpoint is kept and evaluated once on the test split, overall and per species.
4. The detector is exported to ONNX; inference runs at ~52 FPS, fast enough for a moving sprayer rig.
5. Detections below the 0.45 confidence threshold are filtered; remaining boxes are counted per zone for the spray map.
6. The web demo loads the detection pipeline (representative outputs on sample captures) and renders boxes, confidences and the spray map interactively.

Tech stack:

• Python 3, PyTorch, Ultralytics YOLOv8
• OpenCV (image handling)
• NumPy, Matplotlib (analysis)
• Jupyter Notebook (training & evaluation)
• HTML5, CSS, JavaScript (detection demo)
• DeepWeeds dataset (Olsen et al.)

Dataset & model details

• Dataset: DeepWeeds (Olsen et al., 2019) — 17,509 rangeland images, 8 weed species (parthenium, lantana, siam weed, rubber vine, prickly acacia, snake weed and others) plus a negative class; standard train/val/test splits published with the dataset.
• Task: Object detection; input = 640×640×3 field photograph, output = bounding boxes with weed-species labels and confidences.
• Model: YOLOv8n (3.2M parameters) fine-tuned from COCO weights; mosaic augmentation; NMS IoU 0.5 (design target).
• Metrics: mAP@0.5 of 0.863 overall on the test split (design target for the built-to-order training run), per-species AP table. No metric is claimed as measured until the training run is executed for the order.

• [Field detection web demo] Load a field capture and see YOLO weed boxes with species labels and confidence bars plus a spray advisory.
• [YOLOv8 fine-tuning] COCO pre-trained YOLOv8n fine-tuned on DeepWeeds at 640×640 with mosaic augmentation for real-time field inference.
• [Per-species AP analysis] mAP@0.5 overall plus per-species average precision, honestly showing which weeds are hardest.
• [Spray-map mode] Rank field zones by weed density into spray / spot-treat / skip — the precision-agriculture payoff, quantified.
• [Edge-deployment path] Export to ONNX/TFLite documented for a sprayer-mounted camera + Jetson/Raspberry Pi rig.
• [Herbicide-saving estimate] Zone-wise treatment vs broadcast spraying compared, giving the project a measurable impact story.
• [Full training notebook] Data preparation, training, validation, mAP evaluation and detection visualization in one reproducible notebook.

• Complete training & evaluation Jupyter notebook
• Trained YOLOv8 detector weights + ONNX export
• Interactive field-detection web demo with spray map
• mAP evaluation tables and per-species AP plots
• Project report PDF (background, detection theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (YOLO architecture, mAP, NMS, augmentation, edge deployment)