README.md

metadata

title: Afri Wildlife Classify
emoji: 📊
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.38.2
app_file: app.py
pinned: false
license: mit
short_description: Classifies pictures of four African wildlilfe animals

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

Wild Animal Prediction App 🦓🐘🦏🐃

A deep learning-powered web application for classifying African wildlife images using DenseNet-201. This application can identify four key African savanna species: Buffalo, Elephant, Rhinoceros, and Zebra.

🔬 Research Background

This application is based on research from "Evaluating Deep Learning Models for African Wildlife Image Classification: From DenseNet to Vision Transformers" published at DeepLearningIndaba 2025. The work addresses the critical need for automated wildlife monitoring tools in African conservation contexts, where traditional field surveys are labor-intensive and time-consuming.

Key Research Findings

• DenseNet-201 achieved 67% accuracy on the African Wildlife dataset
• Best performing CNN among tested architectures (ResNet-152, EfficientNet-B4)
• Optimized for deployment in resource-constrained conservation settings
• Trained on balanced dataset of 1,504 images (376 per species)

🏆 Model Performance

Metric	DenseNet-201 Performance
Overall Accuracy	67.0%
Macro F1-Score	0.67
Buffalo F1	0.72
Elephant F1	0.61
Rhino F1	0.60
Zebra F1	0.76

Model Comparison

While Vision Transformer (ViT-H/14) achieved 99% accuracy in our experiments, DenseNet-201 was selected for deployment due to:

• Efficiency: 20M parameters vs 632M for ViT
• Speed: 92.5s training time vs 6574.2s for ViT
• Deployability: 4.29 GFLOPs vs 1016.72 for ViT
• Resource Requirements: Suitable for edge deployment and offline use

🚀 How to Use

Online Demo

🤗 Try it now on Hugging Face Spaces: Simply upload an image of a buffalo, elephant, rhinoceros, or zebra and click Submit!

Requirements

gradio
torch
torchvision
Pillow

📸 Best Practices for Image Upload

For optimal results:

• Use clear, well-lit images of the animal
• Ensure the animal is prominently featured in the frame
• Avoid heavily cropped or blurry images
• JPG or PNG formats work best
• File size under 5MB recommended

🌍 Conservation Impact

This tool supports:

• Biodiversity Monitoring: Automated species identification from camera traps
• Anti-Poaching Efforts: Rapid wildlife population assessment
• Citizen Science: Enabling non-experts to contribute to conservation data
• Field Research: Reducing manual image classification workload for researchers

⚠️ Limitations & Considerations

Current Limitations

• Domain Shift: Performance may decline on images from different environments or cameras
• Species Scope: Limited to 4 species (buffalo, elephant, rhino, zebra)
• Image Quality: Sensitive to lighting conditions and image resolution
• 67% Accuracy: Not suitable for critical conservation decisions without human verification

Ethical Considerations

• Human-in-the-loop: Always verify AI predictions with expert knowledge
• Bias Awareness: Model trained on curated dataset may not generalize to all conditions
• Privacy: Ensure no human subjects in uploaded images
• Responsible Use: Tool designed to assist, not replace, conservation professionals

🔧 Technical Details

Architecture

• Base Model: DenseNet-201 pretrained on ImageNet
• Classification Head: Custom 4-class classifier with dropout (p=0.2)
• Input Processing: Images resized to 64×64 pixels, normalized to [0,1]
• Framework: PyTorch with Gradio interface

Training Configuration

• Dataset: African Wildlife Dataset (Ferreira, 2020)
• Training Split: 80% (1,203 images) / Test: 20% (301 images)
• Optimizer: Adam (lr=0.001)
• Loss Function: CrossEntropyLoss
• Batch Size: 32
• Epochs: 10

📊 Dataset Information

The model was trained on the public African Wildlife Dataset containing:

• 1,504 total images (balanced across species)
• 376 images per class (buffalo, elephant, rhino, zebra)
• High-quality color images from African nature reserves
• Representative of savanna ecosystem species

🤝 Contributing

We welcome contributions to improve this conservation tool:

• Dataset expansion: Adding more species or diverse image conditions
• Model improvements: Testing new architectures or training techniques
• UI enhancements: Improving user experience and accessibility
• Documentation: Helping others understand and use the tool

📝 Citation

If you use this application in your research or conservation work, please cite:

@article{aliyu2025wildlife,
  title={Evaluating Deep Learning Models for African Wildlife Image Classification: From DenseNet to Vision Transformers},
  author={Aliyu, Lukman Jibril and Muhammad, Umar Sani and Ismail, Bikisu and Wakili, Almustapha A and Yimam, Seid Muhie and Muhammad, Shamsuddeen Hassan and Abdullahi, Mustapha},
  journal={DeepLearningIndaba 2025 Conference},
  year={2025},
  pages={1-13}
}

🌟 Future Development

Planned improvements include:

• Expanded species coverage (40+ species like Snapshot Serengeti)
• Enhanced robustness through data augmentation and domain adaptation
• Mobile deployment for offline field use
• Active learning integration for continuous model improvement
• Camera trap integration for automated monitoring systems

📞 Support & Contact

For questions, issues, or collaboration opportunities:

• 🤗 Hugging Face Space: Try the live demo and report issues in the community tab
• 🔬 Research: Contact the development team for academic collaboration
• 🌍 Conservation: Reach out for deployment in conservation projects

---

Disclaimer: This is a research prototype designed to assist conservation efforts. Always verify AI predictions with expert knowledge before making critical conservation decisions.

License: This project supports open, Africa-centric AI research and follows ethical AI practices.