dl4cv-oda 0.1.5

Coconut tree detection from drone imagery

Object Detection on Aerial Imagery for Coconut Trees Detection

Coconut tree detection from drone imagery using YOLOv8, Yolov12, and RT_DERT models with OpenStreetMap labels and OpenAerialMap imagery.

Overview

• OpenStreetMap point data: bounding boxes with buffer zones
• Tile large aerial imagery (256×256 at 9cm/pixel): Source
• Convert geographic coordinates to YOLO format
• Train multiple models of YOLOv8 (nano, small, medium) on coconut trees from Kolovai, Tonga
• Train Yolov12 and also RT-DERT.

Source: World Bank - Automated Feature Detection of Aerial Imagery from the South Pacific

Data

Statistics:

• Original: 10,631 trees (Coconut: 10,092 | Mango: 261 | Banana: 181 | Papaya: 97)
• Target: Coconut trees only Clipped labels to tiles: {'total_trees': 10092, 'total_tiles': 551, 'processed_input_tiles': 448, 'skipped_nonlabeled_tiles': 103, 'total_input_trees': 11726} Converted to YOLO format Split: train=313, val=89, test=46

70 / 20 / 10 for train/val/test

• Tiles: 256×256px at zoom 19, EPSG:4326

Structure

data/
├── raw/                # OAM imagery + OSM points
├── chips/              # 256×256 tiles (.tif)
├── labels/             # Per-tile annotations (.geojson)
└── yolo/
    ├── train/          # Training data (.png + .txt)
    ├── val/            # Validation data
    └── config.yaml     # YOLO config

notebooks/
├── experiment.ipynb         # including the dl4cv-oda package and all functions

Setup

curl -LsSf https://astral.sh/uv/install.sh | sh
git clone https://github.com/kshitijrajsharma/dl4cv-oda
cd dl4cv-oda
uv sync

Development ( version bump)

uv sync --extra dev
cz bump
git push --tags

Workflow

1. Clean OSM Data : Filter coconut trees, generate buffered bounding boxes

2. Tile Imagery : Create 256×256 tiles, clip labels to tile extents

3. YOLO Conversion : Transform coordinates (EPSG:4326 : pixels : normalized [0,1])

row, col = src.index(lon, lat)  # rasterio
x_norm = col / img_width
y_norm = row / img_height

4. Train : YOLOv8n, 100 epochs, batch 16

from ultralytics import YOLO
model = YOLO('yolov8n.pt')
model.train(data='data/yolo/config.yaml', epochs=100, imgsz=256, batch=16)

References

• OpenAerialMap
• OpenStreetMap
• Ultralytics YOLOv8
• uv Package Manager