Project Title: AI-Driven Crop Identification for Smallholder Indian Farms Using Multi-Modal Satellite Data
Abstract :
Addressing India's agricultural complexity is critical for food security, yet accurate crop identification remains a major challenge due to fragmented farms and diverse cropping practices. To solve this, our project develops a robust AI model by fine-tuning a state-of-the-art geospatial foundation model on multi-modal satellite data specifically for the Indian context.
Our core technical approach is to fine-tune the NASA-IBM Prithvi foundation model, chosen for its proven effectiveness with complex Earth observation data. We will integrate Sentinel-1 SAR imagery with multispectral data from Sentinel-2 and ResourceSAT-2/2A, leveraging the ground-truthed AgriFieldNet India dataset. This multi-modal fusion is central to our strategy, ensuring high accuracy even during monsoon seasons by mitigating the impact of cloud cover. The performance of this advanced model will be benchmarked against conventional CNN/ViT architectures.
Our pipeline includes a comprehensive preprocessing stage (normalization, augmentation, and handling of missing bands) and rigorous cross-district evaluation. Success will be measured using accuracy, per-class F1-scores, and confusion matrix analysis to ensure fairness across all crop types.
The outcome will be a scalable model providing near-real-time crop monitoring, empowering policymakers with accurate yield forecasts and farmers with data-driven insights. By tailoring a generative AI solution to India’s unique agricultural landscape, our project delivers a tangible tool for enhancing national food security.
