Leveraging Causality and Explainability in Digital Agriculture

Ilias Tsoumas

Keywords: 5. AI for Science, 6. Other

Abstract:

Sustainable agricultural practices are increasingly vital due to escalating environmental concerns and the urgency of climate change mitigation. Digital agriculture offers a promising avenue for promoting these practices through advanced data analysis. A critical example is pesticide use, which, while essential for pest control and food security, also contributes significantly to environmental harm. Integrated Pest Management (IPM) offers a climate-smart alternative but suffers from low adoption, partly due to farmers’ skepticism about its effectiveness. We propose a causal and explainable framework to enhance digital agriculture, using IPM as a focal case to demonstrate the value of causality and explainability. Our framework includes (i) robust pest population predictions across diverse settings through invariant and causal learning, (ii) explainable pest presence predictions via transparent models, (iii) counterfactual-based actionable advice for in-season interventions, (iv) field-specific treatment effect estimations, and (v) assessments of the effectiveness of our advice using causal inference. Supporting this fifth component, a prior study pioneered the use of causal inference for empirically evaluating recommendation systems in agriculture, demonstrating that sowing-time advice based on weather predictions led to yield increases of 12–17%. This evidence-based approach strengthens the case for using causal inference to validate digital agricultural tools, enhancing farmer trust, supporting policy decisions, and fostering the broader adoption of sustainable practices.

Enviromental Data Science: https://doi.org/10.1017/eds.2025.14
AAAI 2023: https://doi.org/10.1609/aaai.v37i12.26697

Submission Number: 44