Agricultural Data: Big or Small?

Agricultural datasets may contain extensive satellite imagery, sensor measurements, or other variables while still having relatively few high-quality observations for training and evaluating AI models. This Perspectives article examines that mismatch using agricultural examples, including a tillage-mapping project with approximately 600 ground-truth observations. Kirti Rajagopalan argues that AI models working with limited agricultural data should incorporate existing scientific knowledge rather than attempting to rediscover known relationships. Researchers should also quantify uncertainty for individual predictions and identify underrepresented crops, soils, production systems, environments, and years. Together, uncertainty and data diversity can guide researchers toward the observations most likely to improve a model. The goal is not merely to make agricultural datasets larger, but to develop AI approaches that use limited data and existing agricultural knowledge more effectively.