logo

Inspiration

India’s farms are small, fragmented, and diverse, making crop monitoring difficult compared to countries with large homogeneous farms. Reliable crop identification is crucial for food security, crop insurance, policymaking, and farmer advisories. With the rise of geospatial foundation models and availability of Indian satellite data, we saw a chance to tackle this challenge using AI + satellite fusion.

What it does

AgriVision uses satellite data (SAR + multispectral) and AI models to classify crop types across fields in Northern India. Our solution: Processes satellite images from Sentinel-1, Sentinel-2, EOS-4, and ResourceSAT-2. Combines time-series, spectral indices, and SAR backscatter for robust predictions. Generates crop maps at field level, displayed on an interactive dashboard. Provides confidence scores + explainability so results are trustworthy.

How we built it

Data Prep – Used AgriFieldNet dataset (Radiant Earth + IBM enhancements), clipped fields, extracted vegetation indices (NDVI, EVI, SAVI), and normalized SAR bands. Baseline – Started with Random Forest/XGBoost for quick benchmarking. Advanced Models – Built CNN + Transformer-based time-series models, fusing SAR + optical data. Explainability – Added Grad-CAM/SHAP for field-level interpretability. Visualization – Created a Streamlit dashboard to map crop predictions across districts.

Challenges we ran into

Data diversity: Farms differ greatly across states → harder to generalize. Class imbalance: Major crops (wheat/rice) dominate, while pulses/minor crops had less data. SAR + Optical fusion: Aligning multi-modal data required heavy preprocessing. Time constraints: Limited time to fully fine-tune large geospatial foundation models.

Accomplishments that we're proud of

Successfully fused SAR + multispectral + time-series data for higher accuracy. Built a working dashboard for crop classification, not just code. Achieved strong performance on minor crops, not just dominant ones. Designed an explainable AI workflow, so farmers and policymakers can trust outputs.

What we learned

How multi-modal fusion (SAR + optical) improves robustness under cloudy conditions. Importance of time-series learning in agriculture (growth cycles matter!). Hands-on with geospatial AI tools & foundation models for agriculture. How to balance innovation vs. feasibility in a time-bound hackathon.

What's next for AgriVision

Scale pan-India → extend beyond 4 states to cover the entire country. Integrate real-time weather and soil data for richer predictions. Deploy as a mobile/web app for farmers & policymakers. Collaborate with government & agri-tech startups for adoption in insurance, yield forecasting, and advisory services. Fine-tune geospatial foundation models (like Prithvi-100M, SatViT) for even better generalization.