Machine Learning for Climate Prediction Training Course

Participants can prepare climate datasets with documented structure, metadata and quality controls.

• NetCDF, Zarr and CF conventions for multidimensional climate datasets
• xarray indexing, grouping and resampling of gridded time series
• Coordinate reference systems and grid-cell area weighting methods
• Missing observations, sensor changes and station metadata checks
• Quality-control flags, physical range tests and duplicate detection
• Reproducible data dictionaries, versioning and provenance registers

Participants can diagnose climate patterns, data limitations and candidate predictive relationships.

• Climate normals, anomalies and baseline periods selected deliberately
• Trend, seasonality and autocorrelation diagnosed before model fitting
• Extreme indices using ETCCDI definitions for heat and rainfall
• Spatial dependence mapped through variograms and Moran's I statistics
• Teleconnection indicators including ENSO, NAO and Indian Ocean Dipole
• Feature screening using correlation, mutual information and collinearity checks

Participants can construct physically credible predictors without introducing information leakage.

• Rolling rainfall totals aligned with cropping and reservoir decisions
• Lagged predictors constructed without introducing future-information leakage risks
• Calendar, elevation and land-cover features encoded for local prediction
• Anomaly features referenced to fixed climatological baseline periods
• Distributed lag features representing delayed hydrological responses
• Feature stores separating raw observations from derived predictors

Participants can establish meaningful baseline models before applying machine learning.

• Persistence, climatology and linear regression established as benchmark models
• ARIMA and seasonal autoregression fitted to station observations
• Generalised additive models capturing nonlinear weather-response relationships smoothly
• Skill scores calculated against relevant operational reference forecasts
• Baseline error patterns segmented by season, location and threshold
• Benchmark selection justified by decision cost and forecast horizon

Participants can train, tune and interpret tree-based climate prediction models.

• Random forests and gradient boosting trained for tabular climate predictors
• Regularisation, learning rates and tree depth tuned systematically
• Classification thresholds selected from false-alarm and miss costs
• Quantile regression forests producing conditional prediction intervals directly
• Feature importance assessed using permutation and SHAP values
• Model cards documenting intended use, limitations and data coverage

Participants can select and assess neural-network approaches for climate datasets.

• Convolutional neural networks processing spatial rainfall and temperature fields
• LSTM and temporal convolution models handling sequential climate signals
• Transfer learning adapted from data-rich regions to sparse networks
• Early stopping, dropout and batch normalisation controlling overfitting
• GPU training choices balanced against reproducibility and energy use
• Architecture selection matched to data volume and forecast value

Participants can test climate prediction models under realistic deployment conditions.

• Blocked temporal cross-validation preserving realistic forecast chronology
• Spatial cross-validation testing transferability beyond observed station locations
• Nested validation separating hyperparameter selection from final performance estimates
• Data leakage checks covering predictors, labels and preprocessing stages
• Train-test splits reflecting deployment lead times and data availability
• Backtesting protocols reproducing historical forecast issue dates faithfully

Participants can interpret climate-model ensembles and scenarios for adaptation planning.

• CMIP6 ensembles, scenarios and model families interpreted correctly
• Reanalysis products distinguished from observations and climate-model output
• Scenario selection linked to planning horizons and decision reversibility
• Ensemble spread separated from internal variability and model disagreement
• Global model resolution limits explained for local adaptation decisions
• Climate model metadata captured using ESGF catalogue fields

Participants can select and assess downscaling and bias-correction approaches.

• Delta change, quantile mapping and regression downscaling compared
• Bias correction fitted separately for precipitation and temperature distributions
• Empirical quantile mapping tested for trend preservation
• Statistical downscaling predictors chosen from physically plausible drivers
• Downscaled outputs checked against station climatology and extremes
• Nonstationarity risks documented when applying historical correction functions

Participants can model extreme climate events and communicate associated tail risks.

• ETCCDI indices calculated for heatwaves, drought and intense rainfall
• Generalised extreme value distributions fitted to annual maxima
• Peaks-over-threshold analysis using generalised Pareto distributions systematically
• Return periods translated into design thresholds and decision triggers
• Compound heat-drought events represented through joint probability measures
• Tail performance evaluated with threshold-specific precision and recall

Participants can evaluate and calibrate probabilistic climate prediction outputs.

• Prediction intervals, quantiles and exceedance probabilities communicated clearly
• Continuous ranked probability score comparing probabilistic forecast quality
• Brier scores and reliability diagrams assessing event probabilities
• Calibration methods including isotonic regression and Platt scaling
• Prediction uncertainty decomposed into data, model and scenario sources
• Decision thresholds set using expected loss and tolerance limits

Participants can interpret model outputs while identifying explanation and equity limits.

• SHAP dependence plots identifying influential predictors and interactions
• Partial dependence checked against known physical climate relationships
• Counterfactual scenarios testing sensitivity to controllable adaptation measures
• Explainability limits stated for correlated predictors and spatial features
• Fairness checks identifying uneven prediction error across communities
• Stakeholder briefing notes separating confidence, uncertainty and recommended action

Participants can connect prediction outputs to adaptation choices and action triggers.

• Impact chains connecting hazards, exposed assets and service disruption
• Forecast-based financing triggers designed from probability and consequence
• Multi-criteria decision analysis comparing adaptation response options
• Cost-loss models quantifying value of earlier protective action
• Threshold matrices assigning owners, actions and escalation routes
• Adaptation pathways updated when observed indicators cross signposts

Participants can specify operational controls for climate prediction model deployment.

• Operational pipelines scheduled for data ingestion, scoring and alerts
• Data drift and concept drift monitored after model release
• Model performance dashboards segmented by geography, season and hazard
• Retraining criteria defined through drift thresholds and performance decay
• Reproducible environments managed with containers and dependency files
• Incident logs recording failed runs, overrides and user feedback

Participants can establish governance controls for consequential climate prediction models.

• Model risk registers covering misuse, uncertainty and decision consequences
• Data licensing, consent and indigenous data sovereignty considered
• Documentation aligned with model cards and reproducible research principles
• Human review points specified for high-consequence adaptation decisions
• Energy and compute use tracked for model development
• Communications avoid false precision and unsupported causal claims

Participants can present prediction evidence clearly to technical and operational stakeholders.

• Forecast narratives tailored for water managers, planners and emergency teams
• Maps designed with accessible legends, uncertainty bands and thresholds
• Technical methods translated into decision-ready briefing materials clearly
• Scenario workshops testing responses under plausible forecast outcomes
• Assumption logs distinguishing evidence, judgement and unresolved questions
• Stakeholder feedback incorporated into forecast product requirements formally

Participants can build, review and present a complete climate-prediction project supporting an adaptation decision.

• Problem framing defines users, adaptation decisions, targets and lead times
• Curated climate dataset documented with provenance, quality checks and features
• Validated machine-learning model benchmarked against climatology and persistence
• Downscaling or extremes analysis selected to suit decision context
• Uncertainty assessment linked to thresholds, actions and residual risks
• Deployment plan specifies monitoring, governance and communication controls
• Peer review tests technical validity and decision usefulness
• Presentation defends recommendations, limitations and next implementation steps