Inspiration
Satellite imagery is powerful, but most AI models act like black boxes. We wanted to create a tool that not only classifies land use but also explains why it made a decision, building trust for real-world applications like farming, urban planning, and environmental monitoring.
What it does
GeoLens classifies satellite images into 10 land-use categories (Forest, River, Residential, etc.) and generates heatmaps showing the exact regions that influenced the prediction — making AI transparent and reliable.
How we built it
Preprocessed the EuroSAT dataset.
Trained a ResNet18 model for land-use classification.
Integrated Grad-CAM for visual explanations.
Deployed the solution as a Streamlit app for interactive use.
Challenges we ran into
Cleaning noisy/incomplete datasets.
Balancing accuracy vs. lightweight deployment.
Handling Grad-CAM integration errors and visualization mismatches.
Accomplishments that we're proud of
Achieved strong accuracy across all 10 classes.
Made AI predictions explainable and user-friendly.
Successfully deployed a working end-to-end web app.
What we learned
Model explainability is as important as accuracy.
Proper preprocessing and noise reduction drastically improve results.
Deployment bridges the gap between research and real-world usability.
What's next for GeoLens
Expand to more datasets (climate, disaster monitoring).
Add real-time inference for drone/satellite streams.
Collaborate with policymakers for practical adoption.
Log in or sign up for Devpost to join the conversation.