AI-Driven Spatio-Temporal Soil Moisture Forecasting for Agriculture
Soil moisture is the critical intermediary between atmospheric forcing and agricultural outcomes: it determines root-zone water availability, crop stress onset, irrigation and sowing decisions, pasture recovery after drought, and the antecedent conditions that govern flood generation. Despite its importance, operational soil moisture outlooks — including the Australian Water Resources Assessment Landscape model (AWRA-L) and the Bureau of Meteorology’s Soil Moisture Seasonal Outlook — are limited in spatial resolution, lack principled uncertainty quantification, or rely on process-model chains to ingest the relatively low-skill daily forecasts produced by seasonal climate models.

Project Outline

This project proposes a unified spatio-temporal generative model for daily soil moisture forecasting at 1–5 km resolution over Australia, with lead times spanning 1–12 days driven by numerical weather prediction and extending to sub-seasonal to seasonal scales (up to three months) driven by ACCESS-S2 ensemble forecasts. The core generative engine is a Conditional Flow Matching (CFM) architecture trained within a compressed latent space, enabling rapid, high-fidelity ensemble generation that remains tractable at national scale. The model will be driven by two complementary input streams: historical weather observations (gridded rainfall, temperature, solar radiation, and antecedent soil moisture) and daily ensemble forecasts from numerical weather prediction (1–12 days) and seasonal climate models such as ACCESS-S2 (2 weeks to 3 months). The project will build on recent advances at CSIRO in (i) high-resolution realistic precipitation downscaling ensemble generation using probabilistic GANs or flow matching, and (ii) the state-of-the-art long-term time series forecasting with future values of exogenous variables. Further architectural gains are possible through principled co-design of how historical observations and forward-looking exogenous forecasts are coupled across space and time within the generative model.

The resulting AI-driven system is positioned as complementary to AWRA-L and the Soil Moisture Seasonal Outlook rather than a replacement: process-model outputs can serve as physically consistent prior states, while the generative model contributes fine-scale spatial structure, sharp probabilistic envelopes, and direct compatibility with AI-derived precipitation ensembles. Evaluation will prioritise metrics of direct agricultural relevance — threshold-weighted Continuous Ranked Probability Score (CRPS), Critical Success Index (CSI), and Brier Skill Score for deficit-threshold exceedance — with particular attention to tail calibration at the onset and duration of moisture deficits that drive real farm decisions.

You will learn:

  • Deep learning at scale: develop and train latent flow matching models on national-scale gridded climate and soil data using high-performance computing infrastructure.
  • Soil and hydrological science: understand key soil moisture dynamics, the distinction between surface and root-zone reservoirs, and their connection to agricultural outcomes.
  • Probabilistic forecasting methods: master flow matching, ensemble calibration, and agricultural-decision-oriented verification metrics (twCRPS, False Alarm Ratio (FAR), Critical Success Index (CSI), and Heidke Skill Score (HSS), Brier Score, and/or reliability diagrams).
  • Sub-seasonal climate forecasting: gain working knowledge of ACCESS-S2 ensemble forecasts, their biases and predictability windows, and how to prepare them as model drivers.
  • Interdisciplinary collaboration: engage with agronomists, hydrologists, and climate services alongside AI researchers in a live applied research environment.
  • Scientific communication: write and present research for peer-reviewed journals and operational stakeholder audiences including farm advisors, irrigation authorities, and emergency managers.

Supervisors:

  • Warren Jin, Principal Research Scientist, Statistical Machine Learning Group, Analytics & Decision Science, CSIRO
  • Enli Wang, Chief Research Scientist and Crop Modelling Team Lead, CSIRO
  • Miaomiao Liu, Assoc. Prof School of Computing, ANU

Questions about this project can be directed to: Warren.jin@csiro.au

To register an expression of interest: click here. You will need to outline why you have selected the research project and how your skills, experience and/or knowledge meet the project requirements.