Fine-Tuning the Prithvi Foundation Model for Crop-Type and Health Mapping in Smallholder Farms

Artificial intelligence is rapidly reshaping how we monitor and manage agriculture from space. In this study, we fine-tune and evaluate a state-of-the-art vision foundation model, Prithvi-EO-2.0, for crop type classification and crop health monitoring across smallholder farms in Telangana, India. Leveraging temporally stacked satellite imagery, we assess the effectiveness of different fine-tuning strategies and address challenges related to fragmented landholdings and limited ground truth data. Our results demonstrate that fully fine-tuning the model significantly improves crop classification accuracy but requires careful handling of class imbalance to prevent overfitting. We discuss limitations arising from spatial resolution constraints and propose future directions using higher-resolution alternatives, emphasizing the potential of adapting large-scale pretrained models to deliver scalable and practical precision agriculture solutions in data-limited, heterogeneous smallholder systems.