Satellite Object Detection using YOLOv11

A computer vision system for detecting and classifying ESA spacecraft and space debris from high-resolution satellite imagery (1024×1024), built using a fine-tuned YOLOv11-Medium model.

🔗 Links

GitHub Repository

🛠 Tech Stack

Python PyTorch YOLOv11 OpenCV Albumentations

📌 Problem

Detect and classify multiple satellite objects and debris in space imagery, where objects vary in orientation, lighting conditions, and noise levels.

📊 Dataset

SPARK 2022 Challenge (University of Luxembourg – SnT)

1024×1024 satellite images
11 object classes (ESA spacecraft + debris)
Custom bounding box annotations

⚙️ Pipeline Overview

Data Preparation: Structured dataset into YOLO format
Label Conversion: CSV → normalized YOLO bounding boxes
Data Augmentation: Applied domain-specific augmentations
Model Training: Fine-tuned YOLOv11-Medium

🔄 Data Augmentation

Flips & rotations (orbital perspectives)
Brightness & contrast changes (sunlight variation)
Gaussian noise (sensor noise simulation)
Blur effects (motion & optical distortion)
Coarse dropout (sensor failure simulation)

🧠 Model Training

Model: YOLOv11-Medium
Epochs: 35
Batch size: 8
Image size: 1024×1024
Learning rate: 0.01

📈 Results

Accuracy: ~86%
Strong generalization (no overfitting observed)
Improved precision and recall over training

📊 Training Metrics

💻 Training Environment

HPC Iris Cluster (University of Luxembourg)
GPU-based training
SLURM job scheduling

🚀 Inference

Generates predictions for test images and outputs bounding boxes in CSV format.