HarvestEye

Inspiration

The inspiration for Harvest Eye emerged from witnessing the disconnect between advanced satellite technology and its practical application in Indian agriculture. While space agencies generate terabytes of agricultural data daily, 86% of Indian farmers—managing plots smaller than 2 hectares—make critical crop decisions with limited scientific insights.

The central question that drove me was: “How can we democratise precision agriculture using generative AI and foundation models to bridge the gap between satellite intelligence and farmer decision-making?”

This led to the conceptual framework of Harvest Eye—an AI system that transforms raw satellite observations into actionable agricultural intelligence through multi-modal learning and conversational interfaces.

What it does

HarvestEye is an AI-powered system that identifies crop types across fragmented farmlands in Northern India using satellite imagery (SAR + multispectral data). It converts raw satellite signals into field-level insights such as crop type, health, and growth stage. The system also provides an interactive and conversational interface, making complex geospatial intelligence accessible to farmers, researchers, and policymakers.

How we built it

Collected SAR and optical imagery from Sentinel-1, Sentinel-2, ResourceSAT-2/2A, and EOS-4.
Preprocessed data using Rasterio, GDAL, and NumPy to align spatial resolutions, normalise values, and remove noise.
Fine-tuned a geospatial foundation model with the AgriFieldNet India dataset, enhanced by IBM’s SAR + optical additions.
Implemented multi-modal fusion (spectral + radar + temporal data) to improve crop classification accuracy.
Designed visual dashboards and conversational query support to make results intuitive and farmer-friendly.

Challenges we ran into

Fragmented farmland: Indian fields are small and irregular, making boundary detection complex.
Data noise: Clouds, shadows, and radar speckle interfered with clean signals.
Class imbalance: Abundant data for wheat and rice, but limited for niche crops.
Generalisation: Ensuring the model performed consistently across multiple states and growing seasons.

Accomplishments that we're proud of

Successfully built a multi-modal crop classification system combining radar and optical data.
Improved classification accuracy through region-specific fine-tuning.
Created an interactive visualisation layer that bridges the gap between advanced AI outputs and practical farming needs.
Demonstrated how conversational AI can make satellite intelligence accessible to non-technical users.

What we learned

SAR is a game-changer during monsoons because it can “see through” clouds.
Generative AI (GANs) can help balance datasets by generating synthetic samples for underrepresented crops.
Temporal consistency is critical for reliable crop monitoring across growth stages.
Multi-scale attention mechanisms help scale from sub-hectare plots to district-level analysis.

What's next for HarvestEye

Expand to real-time monitoring by integrating near-daily satellite feeds.
Deploy a mobile-friendly conversational assistant in multiple Indian languages for farmer queries.
Partner with government agencies and NGOs to provide actionable insights for food security and policy-making.
Incorporate climate modeling to simulate “what-if” scenarios (e.g., reduced rainfall, subsidy shifts).
Explore quantum computing and causal AI for next-generation agricultural intelligence.