AgroVisionAI

Project Title : AgroVisionAI

Abstract

Accurate crop identification across India's fragmented smallholder farms is a critical challenge with high societal impact. While geospatial foundation models (GFMs) trained on homogeneous Western farms reach 85%+ accuracy, performance in Indian conditions drops to 60–70% due to fragmented fields, intercropping, and seasonal variability.

Our project, AgroVisionAI, aims to bridge this gap by building a practical and explainable crop classification system tailored for India. The hackathon deliverable will focus on:

Baseline Model Fine-Tuning – Using Sentinel-1 SAR and Sentinel-2 multispectral data with the AgriFieldNet India dataset to build a robust crop classification model fine-tuned from foundation backbones (Prithvi, Clay).

Explainable Predictions – Developing an interactive dashboard with color-coded crop maps, field-level confidence scores, attention-based heatmaps, and SHAP-based feature attribution to increase trust among farmers and policymakers.

Seasonal Testing – Evaluating the model across Kharif and Rabi seasons with a target of ≥75% accuracy in Kharif and ≥70% in Rabi, demonstrating adaptability.

As future extensions beyond the hackathon, we envision integrating active learning pipelines (for label efficiency) and domain adaptation methods (for stronger seasonal generalization).

The expected outcome is a generalizable and transparent crop identification tool, deployable as a lightweight web dashboard for agricultural departments, insurance agencies, and agri-tech startups. By combining accuracy with interpretability, AgroVisionAI ensures practical adoption in India's diverse farming ecosystem.

Built With

• huggingface-(foundation-models)
• libraries/frameworks:-pytorch/transformers-(modeling)
• rasterio/gdal-(geospatial-preprocessing)
• scikit-learn
• shap-(explainability)
• streamlit-(dashboard)