AI-Enhanced Climate Storylines for Agricultural Adaptation and Resilience (UK–Australia)

A farmer doesn’t plan around a “+2°C world”. They plan around a dry spring that shrinks grass growth, a wet harvest that stalls machinery, a heat spike that drops milk yield, or a flood that takes out a bridge on the only route to market.

Climate science already describes these hazards well at a global scale. The mismatch shows up at the point of action: field scale, short time windows, and hard thresholds. A probabilistic projection can be technically correct and still fail a grower, an adviser, or a regional planner.

That’s where climate storylines earn their keep. A storyline is a physically consistent narrative of “what happened” or “what could happen”, built around a sequence of weather and climate conditions that produce impacts. It doesn’t pretend to remove uncertainty; it makes uncertainty discussable. People can ask: “Is this plausible here?”, “What would have to be true?”, “What breaks first?”, “What would we do differently next time?”

What “AI-enhanced” should mean (and what it shouldn’t)

If AI just spits out nicer prose, it isn’t the point. The value is upstream:

• downscaling model output so extremes look like local weather;
• scanning large archives to find event shapes that matter to farming;
• mapping stakeholder thresholds into explicit rules that a storyline must meet or fail;
• stress-testing candidate storylines against basic physical checks (seasonality, persistence, water balance, known constraints).

And the limits are obvious. A neural downscaler can make fine-grained rainfall fields look realistic while still missing the timing of a storm that matters for spray windows. A clustering model can find patterns in drought sequences and still ignore a farm’s irrigation licence rules. The workflow has to expose these failure modes, not hide them.

A framework that starts with thresholds, not datasets

Most adaptation work starts with what data exists. I’d flip it. Start with decisions and thresholds, then work backwards:

1. Decision inventory. Pick a region and a small set of decisions that carry money, risk, and regret: sowing dates, cultivar choice, irrigation scheduling, on-farm water storage, heat stress plans, flood access routes.

2. Threshold mapping. Translate each decision into measurable triggers and tolerances. Examples: “three consecutive days above X°C during flowering”, “soil moisture below Y for Z days”, “harvestable dry days per fortnight”.

3. Event construction. Use AI-based downscaling and pattern search to build candidate event sequences that satisfy (or barely violate) those thresholds. Keep the physics checks simple and visible.

4. Storyline template. For each candidate, output a structured narrative that a practitioner can audit:

   • triggering conditions (what sets it up)
   • timeline (week-by-week, not annual means)
   • impacts mapped to thresholds (what crosses the line, when)
   • adaptation actions (who can act, with what lead time)
   • “what would change this” sensitivities (one or two variables that flip the result)

5. Co-production loop. Workshop the storylines with growers, agronomists, water managers, and local government. The goal isn’t “buy-in”. It’s correction: missing constraints, wrong thresholds, wrong language, wrong causal chain.

Agentic AI can help here as a research tool, not a replacement for judgement. Give an agent a region, a crop system, and a threshold set; then ask it to search a high-resolution climate archive for candidate sequences, log what it pulled, and show what it rejected. If it can’t explain its choices, it doesn’t get trusted.

Why the UK–Australia pairing matters

The UK and Australia sit at opposite ends of many farming problems: water surplus versus water scarcity; cool-season variability versus heat stress; different regulatory settings; different farm sizes. A cotutelle between Coventry University (UK) and Deakin University (Melbourne, Australia) makes that contrast a feature, not a complication. If a storyline method works in both contexts, it is doing something real.

References (starting points)

IPCC (2021). AR6 Working Group I: The Physical Science Basis. https://www.ipcc.ch/report/ar6/wg1/

IPCC (2022). AR6 Working Group II: Impacts, Adaptation and Vulnerability. https://www.ipcc.ch/report/ar6/wg2/

Lee, J. D. and See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors.

Doshi-Velez, F. and Kim, B. (2017). Towards a rigorous science of interpretable machine learning.

Maraun, D. and Widmann, M. (2018). Statistical Downscaling and Bias Correction for Climate Research. Cambridge University Press.