GeoAgriNet

About GeoAgriNet

Inspiration

The motivation behind GeoAgriNet was to leverage cutting-edge satellite imagery and AI to address the critical challenge of accurate, real-time crop classification in India. The diverse agricultural landscapes coupled with varying climatic conditions require a robust system that combines multiple data types and scalable processing to generate actionable insights. GeoAgriNet aims to empower policymakers, researchers, and advisory services with reliable information to improve crop monitoring and decision-making.

What We Learned

Building GeoAgriNet deepened our understanding of multi-source remote sensing data fusion, especially combining Synthetic Aperture Radar (SAR) and optical satellite imagery. We mastered asynchronous streaming data ingestion in Rust and optimized machine learning pipelines for efficient inference. Additionally, we learned how to apply a foundation geospatial model fine-tuned with localized ground truth for improved accuracy, adapting advanced AI for practical agricultural applications.

How We Built It

The architecture involves ingesting multi-source data—SAR from Sentinel-1/EOS-4, optical imagery from Sentinel-2 and ResourceSAT, and ground labels from AgriFieldNet India. Data ingestion uses Rust’s asynchronous streaming for efficient handling. Preprocessing pipelines extract features like NDVI and EVI from optical images and perform georeferencing and cloud masking for SAR. Temporal and spatial features are fused and fed into a pre-trained geospatial foundation model. This model is then fine-tuned with Indian ground truth data to adapt to local conditions. Finally, Rust-based ML inference generates district-wise crop classification maps. The results are accessible to end users through APIs and dashboards, with a feedback loop for continuous improvement.

Challenges

Key challenges included managing heterogeneous data sources with differing formats and temporal resolutions, designing a scalable yet low-latency ingestion and processing pipeline, and ensuring that the foundation model accurately adapts to diverse Indian agro-ecological zones through fine-tuning. Handling cloud cover in optical data and integrating SAR’s distinct imaging modality required careful preprocessing and feature fusion strategies.

LaTeX Example:

The vegetation indices were computed using formulas like:

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

$$ EVI = 2.5 \times \frac{NIR - Red}{NIR + 6 \times Red - 7.5 \times Blue + 1} $$

These indices serve as critical features for crop classification.

This implementation combines the strengths of remote sensing, AI, and local expertise into GeoAgriNet, offering an innovative solution to India’s agricultural monitoring needs.