YOLOv8n Fine-tuned on VHR-10 Remote Sensing Dataset

This model is a fine-tuned YOLOv8n (nano) model trained on the NWPU VHR-10 (Very High Resolution) remote sensing dataset for detecting ground equipment and vehicles in satellite imagery.

Model Description

This model demonstrates the feasibility of using YOLOv8 for detecting various pieces of ground equipment through satellite imagery, serving as a proof-of-concept for commercial applications in competitive intelligence, fleet monitoring, and automated equipment detection.

Model Details

Model Type: YOLOv8n (nano) - Object Detection
Training Dataset: NWPU VHR-10 Remote Sensing Dataset
Model Size: ~6MB (3M parameters)
Input Resolution: 640x640 pixels
Training Duration: 50 epochs
Framework: Ultralytics YOLOv8

Detected Classes

The model can detect 10 classes of objects commonly found in satellite imagery:

airplane - Aircraft on airfields and airports
ship - Naval vessels and boats
storage_tank - Industrial storage tanks
baseball_diamond - Baseball fields and diamonds
tennis_court - Tennis courts and facilities
basketball_court - Basketball courts
ground_track_field - Athletic tracks and fields
harbor - Harbor facilities and ports
bridge - Bridges and overpasses
vehicle - Ground vehicles and equipment

Performance Metrics

Overall Performance

mAP@0.5: 98.0% (exceptional)
mAP@0.5:0.95: 68.2% (good across IoU thresholds)
Overall Precision: 94.1%
Overall Recall: 96.5%
Inference Speed: 9.9ms per image

Vehicle Detection Performance (Primary Focus)

Vehicle F1 Score: 79.2%
Vehicle Precision: 87.5%
Vehicle Recall: 81.5%
Vehicle mAP@0.5: 88.8%

Class-wise Performance (F1 Scores)

Ground Track Field: 100.0%
Airplane: 98.0%
Ship: 95.8%
Baseball Diamond: 94.3%
Tennis Court: 91.7%
Basketball Court: 90.9%
Bridge: 87.0%
Storage Tank: 84.2%
Harbor: 81.8%
Vehicle: 79.2%

Intended Use

Primary Applications

Proof-of-concept for satellite-based equipment detection
Competitive intelligence and market analysis
Fleet monitoring and logistics optimization
Infrastructure inventory management
Automated lead generation based on equipment detection

Commercial Potential

This model demonstrates that AI can reliably detect vehicles and equipment in satellite imagery, laying the groundwork for specialized commercial applications such as:

Hostler detection for logistics companies
Construction equipment monitoring
Fleet tracking and analysis
Market research and competitive analysis

Usage

Loading the Model

from ultralytics import YOLO

# Load the model
model = YOLO('best.pt')

# Run inference
results = model('satellite_image.jpg')

# Process results
for result in results:
    boxes = result.boxes
    for box in boxes:
        class_id = int(box.cls)
        confidence = float(box.conf)
        print(f"Detected: {model.names[class_id]} (confidence: {confidence:.3f})")

HuggingFace Usage

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

# Download model from HuggingFace
model_path = hf_hub_download(
    repo_id="omgbobbyg/satellite-equipment-detection-yolov8n-vhr10",
    filename="best.pt"
)

# Load and use model
model = YOLO(model_path)
results = model('your_satellite_image.jpg')

Training Details

Dataset

NWPU VHR-10 Dataset: 800 very high-resolution remote sensing images
Training Split: 70% (559 images)
Validation Split: 20% (160 images)
Test Split: 10% (81 images)
Image Sources: Google Earth and Vaihingen dataset

Training Configuration

Model: YOLOv8n (nano)
Epochs: 50
Batch Size: 8 (memory optimized)
Image Size: 640x640
Optimizer: AdamW (auto-selected)
Learning Rate: 0.000714 (auto-selected)
GPU: NVIDIA RTX 4090

Limitations and Considerations

Strengths

Excellent overall detection performance (98% mAP@0.5)
High recall rate ensures minimal missed detections
Fast inference suitable for real-time applications
Good generalization across different object types

Limitations

Vehicle detection shows 49% over-prediction rate (false positives)
Performance varies with object size and complexity
Generic model - specialized training could significantly improve accuracy
Limited to 10 predefined classes

Recommendations for Production Use

Implement post-processing filtering for specific use cases
Consider ensemble methods for higher accuracy
Use larger YOLOv8 variants (s/m/l) for better precision
Develop specialized models for specific equipment types

Citation

If you use this model in your research, please cite the original VHR-10 dataset:

@article{cheng2014multi,
  title={Multi-class geospatial object detection and geographic image classification based on collection of part detectors},
  author={Cheng, Gong and Han, Junwei and Zhou, Peicheng and Guo, Lei},
  journal={ISPRS Journal of Photogrammetry and Remote Sensing},
  volume={98},
  pages={119--132},
  year={2014},
  publisher={Elsevier}
}

License

This model is released under the MIT License. The underlying YOLOv8 framework is licensed under GPL-3.0.