cocoa-plant-detector 0.1.1

YOLOv8 Cocoa Plant Detection + GIS AI Pipeline with CLI

🌱 Cocoa Plant Detection + GIS Pipeline with YOLOv8 + Streamlit

---

Table of Contents

• Overview
• Project Structure
• Full Detection Pipeline
• QGIS Integration
• Streamlit Web Demo
• Command-Line Interface (CLI)
• Geo-Referencing Details
• Dataset Notes
• Model Performance
• Reproducibility
• Example Visuals
• License
• Maintainer

---

Demo Notebook

📌 Note: GitHub may truncate full notebook preview due to image-heavy output — download or open locally for complete view.

• 📓 Training, Inference, Mapping with YOLOv8

---

Overview

This repository delivers a production-ready AI pipeline for detecting cocoa plants and planting holes from high-resolution aerial imagery using YOLOv8 and QGIS.

It includes:

• 🧩 Tiling of large orthomosaic images
• 🧠 Batch object detection on tiles
• 🧭 Spatial deduplication and coordinate conversion
• 🗂️ GeoJSON/GPKG export for GIS tools

---

🚀 Features

• 🧱 Orthomosaic tiling with overlap
• 🧠 YOLOv8 CLI + Streamlit interface
• 🌍 GeoJSON/GPKG GIS-ready outputs
• 🎯 UTM coordinate conversion + deduplication
• 🧪 Notebook: training, inference, and QGIS integration

---

🧪 Installation

Clone the repo and install dependencies:

git clone https://github.com/NiiOsa1/cocoa-plant-detector.git
cd cocoa-plant-detector
pip install -r requirements.txt

---

🧠 Model Auto-Download

The trained YOLOv8s model will be downloaded automatically when you run the CLI or Streamlit app — no manual setup needed.

🔗 Backup link (optional): Download from Google Drive

---

Project Structure

cocoa-plant-detector/
├── cocoa_cli_pipeline/         # 🔧 CLI entrypoints (tile, infer, geoconvert)
│   ├── cli.py                  # ├── CLI command dispatcher
│   ├── infer.py                # ├── CLI: batch inference
│   ├── geoconvert.py           # ├── CLI: label → GeoJSON conversion
│   └── tiler.py                # └── CLI: tile large .tif orthomosaics
├── cocoa_pipeline_core/        # 🧠 Core reusable logic (non-CLI)
│   ├── tile_creator.py         # ├── Manual tile generator (1024px w/ overlap)
│   ├── batch_infer_yolov8.py   # ├── Scripted batch inference (auto-downloads model)
│   ├── deduplicate_and_filter.py  # ├── Deduplicate + convert YOLO txt to GeoJSON
│   └── utils.py                # └── Utility functions (e.g., download model)
├── streamlit_app/              # 🌐 Web UI (live tile prediction)
│   └── app.py
├── notebooks/                  # 📓 Training + inference walkthrough
│   └── CocoaPlant_YOLOv8_Training_QGIS_Deployment.ipynb
├── examples/                   # 📸 Visual results: metrics, outputs, overlays
│   ├── val_batch0_pred.jpg     # ├── Predicted validation samples
│   ├── StreamlitApp.png        # └── UI + overlay visualizations
├── test_data/                  # 🧪 Sample tiles + labels for quick testing
│   └── tile_31200_22400.tif
├── runs/                       # 📂 Outputs from inference (auto-created, .gitignored)
├── requirements.txt            # 📦 Dependencies for pip install
├── setup.py                    # 📦 Enables pip install . and CLI registration
├── .gitignore                  # 🧼 Cleans output, model weights, build artifacts
└── README.md                   # 📖 You're looking at it

---

Full Detection Pipeline

1. 🧱 Tile Large Aerial Image

   python cocoa_pipeline_core/tile_creator.py
   

   Input: Image_4.tif

   Output: tiles1/ (1024×1024 px tiles with 22% overlap)

2. 🧠 YOLOv8 Batch Inference

   You can now run the inference script with your own tile folder like this:

   python cocoa_pipeline_core/batch_infer_yolov8.py \
    --tiles test_data/ \
    --output runs/test_output
   

🔁 The script will automatically:

• Download best.pt from Google Drive if missing

• Predict on all .tif tiles in the folder

• Save detection labels and visual outputs

  Inference across all tiles

  Confidence: conf=0.22, IoU: 0.73

  Outputs YOLO .txt format

3. 🌍 Convert to GeoJSON/GPKG

   python cocoa_pipeline_core/deduplicate_and_filter.py
   

   python deduplicate_and_filter_final.py

   Converts to map-based UTM coordinates

   Deduplicates detections from overlapping tiles

   Saves:

   qgis_ready6_detections.geojson

QGIS Integration

This repo includes a ready-to-open QGIS project:

📄 Project_Cocoa.qgz

Includes:

Image_4.tif as the basemap

Pre-styled detection layer (holes + cocoa plants)

Correct CRS, colors, symbology

🚀 To use: Open QGIS (v3.28+ recommended)

Load Project_Cocoa.qgz

All layers and styles will auto-load 🎯

🗺️ GIS Integration Output

Final Detection Overlay in QGIS Bounding boxes converted to UTM, deduplicated, and displayed on the original orthomosaic.

🚀 Streamlit Web Demo

A live demo is available to test cocoa plant and planting hole detection directly in your browser:

👉 Launch the Streamlit App

You can upload .tif, .jpg, or .png tiles and get YOLOv8-predicted outputs in real time — powered by the same model used in the pipeline.

🚀 Streamlit Demo

Live Tile Prediction (via Streamlit) This tile was uploaded through the Streamlit UI and predicted live using the trained model.

Predicted Detection Output

Command-Line Interface (CLI)

The full pipeline is also usable via a single terminal command. After installing the package with:

pip install .

You get access to:

cocoa-cli --help

Available Subcommands:

Command	      Description
tile	      Tile large aerial image
infer	      Run YOLOv8 batch inference
geoconvert    Convert YOLO labels to deduplicated GeoJSON

Example Usage:

 1. Tile large .tif image
cocoa-cli tile --input path/to/Image_4.tif --output tiles/

 2. Run batch inference
cocoa-cli infer --tiles tiles/ --output predictions/

 3. Convert YOLO txt → GeoJSON (deduplicated, GIS-ready)
cocoa-cli geoconvert --input predictions/ --output output.geojson

💡 The trained model is automatically downloaded if missing. You do not need to manually specify --weights unless using a custom model.

Geo-Referencing Details

These values were extracted from the original orthomosaic (Image_4.tif) and used to convert pixel detections into map-based coordinates (UTM):

Origin X      = 525765.6597     ← Upper-left X in UTM meters
Origin Y      = 8702143.3199    ← Upper-left Y in UTM meters
Pixel Width   = 0.01298099167   ← ~1.3 cm per pixel
Pixel Height  = -0.01298099167  ← Negative = North-up

📌 Important: These values are dataset-specific. Always extract your raster metadata when working with a different image.

Dataset Notes

Total training images: 1500

1089 expertly labeled

411 cleaned pseudo-labeled

Classes:

0 = cocoa-plant

1 = hole

Model Performance

| Set        | mAP@0.5 | Precision | Recall | F1 Score |
|------------|---------|-----------|--------|----------|
| Train      | 0.907   | 0.861     | 0.876  | 0.868    |
| Validation | 0.844   | 0.846     | 0.797  | 0.821    |
| Test       | 0.779   | 0.835     | 0.738  | 0.784    |

✅ Model: YOLOv8s (22.6MB)

⚡ Fast inference: ~16ms per image

🎯 Balanced detection for both classes

🔒 No signs of overfitting despite small dataset

Reproducibility

Run the pipeline locally:

1. Tile image python cocoa_pipeline_core/tile_creator.py

2. Run inference python cocoa_pipeline_core/batch_infer_yolov8.py

3. Convert & deduplicate python cocoa_pipeline_core/deduplicate_and_filter.py

Optional: Streamlit UI streamlit run streamlit_app/app.py

---

Example Visuals

Here’s a quick walkthrough of the pipeline in action:

A few visual highlights from model performance and predictions:

📊 Training Metrics & Model Behavior

F1 Score vs Confidence

Precision vs Confidence

---

🧠 Sample Predictions on Validation Tiles

Sample Prediction – val_batch0

Sample Prediction – val_batch1

---

🧪 Try It Yourself

Want to test the pipeline locally with included sample tiles?

streamlit run streamlit_app/app.py

Upload from test_data/ and see the same prediction pipeline running on your machine.

License

This repository is licensed under the MIT License. You’re free to use, modify, and distribute — with credit to the author.

Maintainer

Michael Mensah Ofeor 🔗 GitHub – @NiiOsa1 📩 [e-mail]

⚡ Built for real-world cocoa analysis. Powered by geospatial logic, YOLOv8, and a deep love for clean AI pipelines.