Foundation models have accelerated scientific discoveries for weather prediction, protein folding and natural language processing. Meanwhile, agricultural sciences, where technical breakthroughs could help eradicate world hunger while mitigating environmental pollution and climate impacts, have yet to achieve any substantial gains from foundation models.

In AgriscienceFM, we argue that this is caused by two factors: first, the complexity and multidisciplinary nature of agricultural sciences, and second, the diversity and locality of available agricultural datasets.

To overcome these challenges, AgriScienceFM will develop three modular FMs, each tackling a major driver of agriculture: biological material (G), environment (E) and management (M), which are later aligned to enable multimodal and agentic AI pipelines.

To achieve this ambition:

(a) we contribute with agriculture-specific methods for foundation model development, and bring together, annotate and align existing public datasets to facilitate FM training.

(b) we develop a first-of-its-kind suite of benchmarking tasks that are designed to be widely applicable in the field.

AgriscienceFM benchmarking tasks are grouped thematically into four use cases that each moves towards addressing a major challenge in the field: monitoring crops and water from space, farmer advice for soil health, phenotyping and breeding resilient crops, and precision crop and livestock farming for pest and diseases.

Collectivity, the development of the three foundation models, tailored to the three main drivers in agriculture, and their thorough evaluation on our four use cases, demonstrates how AI can enable advancement in agricultural sciences.