🌊 Flood Detection using Vision Transformers (ViT)

A computer vision application that detects flooded vs non-flooded scenes in images using a fine-tuned Vision Transformer (ViT).

The model is trained using LoRA (Low-Rank Adaptation) for parameter-efficient fine-tuning and includes an interactive Gradio interface for real-time inference.

---

🖼️ App Demo Screenshot

---

✨ Features

• Vision Transformer Backbone: Uses google/vit-base-patch16-224-in21k.
• Parameter-Efficient Training: Fine-tunes ~0.6% of model parameters using LoRA.
• Automated Dataset Handling: Downloads the Louisiana Flood 2016 dataset automatically via kagglehub.
• Interactive Inference: Upload images and get predictions via a Gradio UI.
• Deployment Ready: Compatible with Hugging Face Spaces.

---

🛠️ Tech Stack

• Model: Vision Transformer (ViT-Base)
• Fine-Tuning: LoRA (PEFT)
• UI: Gradio
• Frameworks: PyTorch, Transformers, PEFT
• Dataset: Louisiana Flood 2016 (Kaggle)

---

📦 Installation

1. Clone the Repository

git clone https://github.com/arman1o1/flood-detection-ViT.git
cd flood-detection-ViT

2. Install Dependencies

pip install -r requirements.txt

---

▶️ Usage

Run the main script:

python flood_detection_vit.py

What happens next?

1. Model Check

   • Looks for trained LoRA adapters in ./flood_detection_vit_lora

2. Training (if needed)

   • Downloads the dataset automatically
   • Fine-tunes the ViT model for 3 epochs
   • Saves adapters locally

3. Inference

   • Launches a Gradio web interface (local or public link)
   • Upload images to classify flood vs non-flood scenes

---

⚙️ Technical Details

• Base Model: google/vit-base-patch16-224-in21k

• Task: Binary Image Classification

• LoRA Configuration:

  • Rank: 16
  • Alpha: 16
  • Target Modules: Query / Value

• Execution: GPU recommended, CPU supported

• Caching: Trained adapters reused on subsequent runs

---

⚠️ Notes & Limitations

• First run may take time due to dataset download and training
• GPU significantly speeds up training
• Intended for research and experimentation, not production deployment

---

📄 License

This project is licensed under the MIT License.