Oil Palm Tree Detection from Drone Imagery

A deep learning project for detecting and counting oil palm trees in aerial drone imagery using state-of-the-art object detection models.

Overview

This project implements and compares two object detection approaches for automated oil palm tree counting:

• YOLOv8 (You Only Look Once v8)
• Faster R-CNN with ResNet50/ResNet101 backbones

The models are trained on high-resolution drone imagery that has been tiled into 512x512 pixel images for efficient processing.

Paper: Oil Palm Tree Detection from Drone Imagery (PDF)

Results

Best Model Performance

Model	Backbone	mAP@0.5	mAP@0.5:0.95	Precision	Recall	F1-Score
Faster R-CNN	ResNet101	0.812	0.554	0.812	0.734	0.771
YOLOv8	YOLOv8-L	0.811	0.578	0.643	0.901	0.751

• Total experiments conducted: 69 (6 YOLOv8 + 63 Faster R-CNN)
• Best mAP@0.5: 0.812 (Faster R-CNN ResNet101)
• Best mAP@0.5:0.95: 0.578 (YOLOv8-L)

Sample Detection Output

Faster R-CNN detection results showing 8,083 detected palm trees

Detection results are exported as GeoJSON files for GIS integration:

• detected_palms_fasterrcnn.geojson
• detected_palms_yolo.geojson

Interactive Map

Inference results in Google Earth Engine

🔗 Direct Link: (https://daring-thought-480112-g0.projects.earthengine.app/view/newapp2)

The interactive map allows you to:

• View all detected palm trees
• Toggle between YOLO and Faster R-CNN detections
• Visualize detection confidence levels
• Analyze tree health using NDVI
• Explore zone-by-zone statistics

Project Structure

cv_project/
├── 01_raw_data/              # Original drone imagery
├── 02_final_tiles_only/      # Generated 512x512 image tiles
├── 03_datasets/              # Processed datasets for training
│   ├── oil_palm_coco_v1/     # COCO format dataset (v1)
│   ├── oil_palm_coco_v2/     # COCO format dataset (v2 - with test split)
│   ├── oil_palm_coco_v3/     # COCO format dataset (v3)
│   ├── oil_palm_yolo_v1/     # YOLO format dataset (v1)
│   └── oil_palm_yolo_v2/     # YOLO format dataset (v2)
├── 04_experiments/           # Experiment results and logs
│   ├── 01_yolov8/            # YOLOv8 experiment outputs
│   ├── 02_faster_rcnn/       # Faster R-CNN experiment outputs
│   ├── results/              # Final models and predictions
│   │   ├── yolo_best.pt      # Best YOLOv8 model weights
│   │   ├── restnet101__exp63_best.pt  # Best Faster R-CNN weights
│   │   ├── detected_palms_*.geojson   # Detection results
│   │   └── detection_*.png   # Visualization outputs
│   └── all_experiments_log.csv  # Complete experiment metrics
├── 05_notebooks/             # Training notebooks
│   ├── 01_YOLO_oil_palm_detection.ipynb
│   ├── 02_resnet50_oil_palm_detection_FasterRCNN.ipynb
│   └── 03_resnet101_oil_palm_detection_FasterRCNN.ipynb
│   └── 04_04_inference_best_models.ipynb
├── 06_utilities/             # Helper scripts
│   └── 00_create_tiles.ipynb # Tile generation script

Dataset

Statistics

Dataset v1 (used for initial experiments):

Split	Images	Annotations	Avg Objects/Image
Train	52	2,305	44.3
Valid	13	563	43.3
Total	65	2,868	44.1

Dataset v2 (expanded dataset with test split):

Split	Images	Annotations	Avg Objects/Image
Train	306	39,465	129.0
Valid	153	19,666	128.5
Test	51	6,817	133.7
Total	510	65,948	129.3

Format

• Image size: 512 x 512 pixels
• Annotation format: COCO JSON (for Faster R-CNN) and YOLO TXT (for YOLOv8)
• Class: Single class - palm-tree

Installation

Requirements

# Clone the repository
git clone https://github.com/yourusername/oil-palm-detection.git
cd oil-palm-detection

# Create virtual environment (using uv)
uv venv
source .venv/bin/activate

# Install dependencies
uv pip install torch torchvision
uv pip install ultralytics  # For YOLOv8
uv pip install albumentations pycocotools pandas matplotlib seaborn

Dependencies

• Python 3.10+
• PyTorch 2.0+
• torchvision
• ultralytics (YOLOv8)
• albumentations
• pycocotools
• pandas
• matplotlib

Usage

Training YOLOv8

from ultralytics import YOLO

# Load model
model = YOLO('yolov8m.pt')

# Train
results = model.train(
    data='path/to/oil_palm_yolo_v1/data.yaml',
    epochs=100,
    imgsz=640,
    batch=16
)

Training Faster R-CNN

See 05_notebooks/02_resnet50_oil_palm_detection_FasterRCNN.ipynb or use the training script:

# Configure experiment
exp_config = {
    'name': 'resnet101_experiment',
    'backbone': 'resnet101',  # or 'resnet50'
    'pretrained': True,
    'num_epochs': 110,
    'batch_size': 4,
    'learning_rate': 0.004,
    # ... see full config in training script
}

Configurable Architecture Parameters (Faster R-CNN)

{
    'box_score_thresh': 0.10,      # Detection confidence threshold
    'box_nms_thresh': 0.5,         # NMS IoU threshold
    'box_detections_per_img': 100, # Max detections per image
    'rpn_fg_iou_thresh': 0.7,      # RPN foreground threshold
    'rpn_bg_iou_thresh': 0.3,      # RPN background threshold
    'box_positive_fraction': 0.30  # Positive sample ratio
}

Experiments

All 69 experiments are logged in 04_experiments/all_experiments_log.csv with the following metrics:

• mAP@0.5, mAP@0.5:0.95, mAP@0.75
• Precision, Recall, F1-score
• Inference time (ms)
• Training time (hours)
• All hyperparameters and augmentation settings

Key Findings

1. Faster R-CNN ResNet101 achieved the best mAP@0.5 (0.812) and precision (0.812)
2. YOLOv8-L achieved the best mAP@0.5:0.95 (0.578) and recall (0.901)
3. Both models achieve similar mAP@0.5 (~0.81), but differ in precision/recall trade-off:
   • Faster R-CNN: Higher precision, lower recall
   • YOLOv8: Lower precision, higher recall
4. Dense object detection (129 objects/image avg) benefits from:
   • Higher box_positive_fraction (0.30 vs default 0.25)
   • Score threshold of 0.10 for better precision-recall balance
5. Default anchor aspect ratios [0.5, 1.0, 2.0] outperformed "optimized" ratios

Authors

• Nariman Tursaliev (st125983)
• Luis Medina (st124895)

License

This project is part of the Computer Vision course at AIT (Asian Institute of Technology).

Acknowledgments

• Asian Institute of Technology (AIT)