Foundation models for agricultural sciences - Challenges and opportunities

Ioannis N. Athanasiadis

Keywords: agricultural sciences, foundation models

TL;DR: Foundation models for advancing agricultural sciences

Abstract:

Recent advances in artificial intelligence (AI) are changing fast the way we conduct research in agricultural sciences, and address global challenges related to food security. Despite a first wave of success, there has been relatively less impact to agricultural and food sciences from AI advances, especially when compared to other scientific domains and global challenges, such as biodiversity conservation and climate change. In part, this is because agriculture is inherently a complex, multifaceted problem: it concerns not only plants and animals growing in, and interacting with, their natural environment; it also concerns humans who intervene in these processes, as farmers, policymakers, technology providers, or consumers.
Agricultural systems interactions across humans, plants, animals and nature are both hard to capture and hard to understand. Hard to capture, as most experimental and observational datasets are (relatively) small, expensive and slow to collect, and offer only a partial view of the full interaction space. Hard to understand, as domain knowledge is fragmented across several disciplines, from genetics to agroecology and food systems science. Together, these challenges, i) complex human-environment interactions, ii) limited and fragmented data and iii) siloed knowledge, have constrained the transformative potential of AI for agriculture.
Foundation models, developed with self-supervised methods using unlabeled and multimodal multidisciplinary corpora, offer a new pathway for AI-driven discoveries in agricultural sciences.
Foundation models can help overcome the issues inherent to AI research in agriculture, in two distinct ways. First, by performance improvements in terms of accuracy or less uncertain decision-making, and, second, by lowering the data barriers for developing performant models in new conditions or tasks. Benchmarking the performance of foundation models in both challenges is necessary for monitoring progress.

Submission Number: 123