Null Pointers

Satellite Eyes, AI Insights — Transforming Indian Agriculture

Comment

Inspiration

Agriculture in India is powered by millions of smallholder farmers with fragmented and diverse fields. Unlike the large homogeneous farms in the US or Europe, this makes crop identification using satellite imagery much harder.

We were inspired by the vision that combining AI + satellite data can provide accurate crop maps, supporting food security, crop insurance, and sustainable farming policies.

What it does

Our idea is to build an AI-powered crop classification system using:

  • SAR data (Sentinel-1, EOS-4): Captures soil moisture and crop structure even during cloudy monsoon seasons.
  • Multispectral data (Sentinel-2, ResourceSAT): Provides vegetation indices such as NDVI and EVI to monitor vegetation health.
  • AI models: Fuse SAR and optical data to classify crops (e.g., rice, wheat, maize, sugarcane) across Indian districts.

The system will produce field-level crop maps for yield forecasting, insurance validation, and policy-making.

How we built it (planned)

  1. Data Preparation

    • Use the AgriFieldNet India dataset with crop labels.
    • Preprocess SAR + multispectral imagery: resampling, cloud masking, temporal alignment.

  2. Feature Engineering

    • Compute vegetation indices:

$$ NDVI = \frac{NIR - RED}{NIR + RED} $$

$$ EVI = \frac{2.5 \cdot (NIR - RED)}{(NIR + 6 \cdot RED - 7.5 \cdot BLUE + 1)} $$

  • Extract SAR backscatter features (VV, VH).
  • Build phenological curves to capture crop growth cycles.
  1. Modeling
    • Baseline: Random Forest, XGBoost.
    • Advanced: CNN + LSTM for spatio-temporal learning.
    • Foundation Models: Fine-tune IBM Prithvi or SatMAE for crop classification.

Challenges we ran into (expected)

  • Fragmented farms → Mixed pixels at 10m resolution.
  • Seasonal overlaps → Similar spectral signatures across crop cycles.
  • Generalisation → Ensuring model accuracy across UP, Bihar, Rajasthan, and Odisha.
  • Compute efficiency → Training deep models within hackathon constraints.

Accomplishments that we’re proud of

  • Designing a fusion pipeline for SAR + optical imagery.
  • Leveraging geospatial foundation models for Indian agriculture.
  • Building a scalable AI solution with real-world impact.
  • Aligning with national food security and farmer welfare goals.

What we learned

  • Preprocessing techniques for multi-source satellite datasets.
  • Importance of temporal signals in crop classification.
  • Strengths of foundation models in geospatial AI.
  • How to design AI systems that balance accuracy and scalability.

What’s next for Null Pointers

  • Expand with ISRO datasets (ResourceSAT-2/2A).
  • Deploy as a cloud-based service/API for farmers, insurers, and policymakers.
  • Scale across all major crop-growing states in India.
  • Extend to yield prediction, irrigation needs, and climate resilience analysis.

Built With

Share this project:

Updates