CropVision India: AI-Powered Satellite Crop Mapping

Inspiration

India’s agricultural sector feeds more than a billion people but still faces challenges like crop misreporting, lack of timely yield forecasts, and unreliable monitoring due to weather and cloud cover. Traditional survey methods are slow, costly, and often inaccurate. We were inspired by the potential of satellite data and AI to bring transparency, efficiency, and scalability to crop monitoring and support farmers, policymakers, and insurers with better information.

What it does

CropVision India is an AI-powered satellite crop mapping system that classifies crop types using multimodal satellite data. By combining Sentinel-1 SAR (VV, VH) and Sentinel-2 multispectral imagery, the model generates accurate district- and field-level crop maps across Indian states like Uttar Pradesh, Rajasthan, Odisha, and Bihar. These maps can be used for:

Real-time crop monitoring
Early yield forecasting
Agricultural policy planning and crop insurance support ## How we built it
Data preprocessing: Applied radiometric correction, SAR speckle filtering, cloud masking, and resampling all inputs to 10m resolution.
Feature engineering: Constructed temporal vegetation indices (NDVI, EVI, NDWI) and SAR backscatter features.
Modeling: Used a multimodal temporal transformer architecture, leveraging the Prithvi geospatial foundation model for transfer learning, fine-tuned on the AgriFieldNet India dataset.
Evaluation: Performed district-level split testing using F1-score, accuracy, and per-class precision/recall.
Tools & Libraries: Python, PyTorch, Hugging Face Transformers, Rasterio, GDAL, Google Earth Engine. ## Challenges we ran into
Handling cloud cover in optical imagery required robust cloud masking
Processing large satellite datasets efficiently under resource constraints
Balancing imbalanced crop type labels during training
Ensuring generalization across different states and agro-climatic zones
## Accomplishments that we're proud of
Successfully implemented a multimodal AI pipeline that fuses SAR and optical data
Achieved high accuracy and F1-scores across diverse regions
Scaled the model to produce reliable crop maps at both field and district levels
Built a solution aligned with real-world applications like crop insurance, yield prediction, and government policy support
## What we learned
The importance of multimodal fusion for overcoming limitations of single-source satellite data
How temporal transformers effectively capture crop growth cycles
Practical aspects of handling big geospatial data (resampling, corrections, cloud filtering)
The value of foundation models (like Prithvi) for transfer learning in geospatial AI ## What's next for CropVision India: AI-Powered Satellite Crop Mapping ## What's next for CropVision India: AI-Powered Satellite Crop Mapping
Expand coverage across all Indian states and integrate more crop types
Incorporate real-time Google Earth Engine pipelines for continuous monitoring
Develop a farmer-accessible dashboard for ground-level insights
Collaborate with government agencies and insurers for deployment in agricultural schemes
Extend the model for yield estimation and climate resilience studies