ISRO Cross Modal Remote Sensing Image Retrieval

Cross-modal image retrieval system for remote sensing imagery using deep learning.

Features

• SAR → Optical Retrieval
• Optical → SAR Retrieval
• Optical → Optical Retrieval
• SAR → SAR Retrieval
• FAISS based fast similarity search
• Streamlit Web Interface

Dataset

ROIs1158 Spring Dataset

Model

Dual Encoder Architecture

• SAR Encoder
• Optical Encoder
• Contrastive Learning
• 256-D Embedding Space

Results

Metric	Score
Top-1 Accuracy	85.10%
Top-5 Accuracy	94.49%
Top-10 Accuracy	96.80%

Average Retrieval Time: 0.0947 sec

Project Structure

src/ ├── models/ ├── datasets/ ├── training/ └── retrieval/

Run

Train:

python src/training/train.py

Generate Embeddings:

python src/retrieval/generate_embeddings.py

Build Index:

python src/retrieval/faiss_index.py

Evaluate:

python src/retrieval/evaluate.py

Launch App:

streamlit run app.py

Demo Screenshots

Dashboard

Optical → Optical Retrieval

Optical → SAR Retrieval