MeravaLens

💡 Inspiration

Every day, satellites capture terabytes of imagery covering every inch of Earth’s surface — but turning that raw data into meaningful, actionable insights remains a challenge.
We built MeravaLens to change that.

Inspired by the growing urgency of urban expansion, environmental degradation, and the need for real-time geospatial intelligence, we set out to build a platform that empowers researchers, planners, and decision-makers to understand the land — pixel by pixel.

🚀 What It Does

MeravaLens is an AI-powered platform that transforms satellite imagery into rich, structured geospatial insights.

It combines:

🛰️ Open-access satellite data
🌦️ Environmental APIs (weather + air pollution)
🤖 A custom-built ResUNet for semantic segmentation
🧠 Large Language Models (LLaMA 3) for summarizing land use

With this pipeline, users can:

Retrieve satellite imagery for any global coordinate
Segment the landscape into classes (e.g. buildings, roads, water, forest, agriculture)
Add live environmental context
Generate easy-to-understand textual summaries of land use and conditions

🛠️ How We Built It

Frontend: React + MUI for an intuitive, interactive UI
Backend: Django & FastAPI power data access and model integration
AI Models:
- ResUNet (custom) with pretrained ResNet-18 encoder for segmentation
- Trained on the LoveDA dataset for robust urban/rural generalization
- Classification head for urban vs rural domain prediction
APIs:
- Google Maps Static API for satellite imagery
- OpenWeatherMap & Air Pollution API for environmental data
LLM Integration: Hugging Face’s Llama-3.3-70B-Instruct to generate human-readable insights from segmentation maps

🧗 Challenges We Ran Into

Training a dual-task model (segmentation + classification) with stable gradients was tough — balancing the losses took iteration
Ensuring consistent dimensions and skip connections when adapting UNet to a ResNet encoder required careful decoder design
Integrating multiple data sources (imagery + weather + air quality) and syncing them spatially and temporally wasn’t trivial
Turning raw segmentation into natural language summaries using an LLM required thoughtful prompt engineering

🏆 Accomplishments That We're Proud Of

Built a full-stack, production-ready platform combining computer vision, environmental data, and LLMs
Achieved strong urban-rural generalization on the LoveDA dataset
Designed a custom decoder architecture to mirror ResNet skip connections — improving segmentation performance
Seamlessly integrated LLM-generated summaries from structured pixel data

📚 What We Learned

How to build a scalable deep learning pipeline for semantic segmentation in geospatial settings
How to blend structured vision outputs with LLMs for multimodal understanding
How model performance varies between urban and rural domains — and how datasets like LoveDA help bridge that gap
The value of combining visual AI with environmental and textual intelligence for richer insights

🔭 What’s Next for MeravaLens

🌍 Launch a public demo that allows users to drop a pin and get full land + weather analysis
📦 Expand model support to temporal satellite data for change detection
🛰️ Integrate other datasets (Sentinel, Landsat, etc.)
🧠 Fine-tune the LLM summarization for domain-specific users like urban planners or ecologists
⚡ Optimize performance for faster real-time inference using quantized models and ONNX