Machine learning for seasonal climate forecasting
Machine learning for long range seasonal climate forecasting
IBM Research work on forecasting 2m temperature and precipitation weeks to months ahead with machine learning, published at NeurIPS 2020 CCAI and EGU 2022.
- Status
- Published
- Period
- 2020 to 2022, at IBM Research
- Institution
- IBM Research
- Role
- Co-author and machine learning research scientist
- Co-authors
- EE Vos, A Gritzman, S Makhanya, CD Watson, MA Zaytar, B Zadrozny, D Salles Civitarese, TM Mathonsi
- Topics
- Seasonal forecasting
- Climate risk
- Time series
- Recurrent neural networks
- Deep learning
- Geospatial machine learning
The question
A central challenge in seasonal climate prediction is whether a forecast can beat climatology, the long run average for that place and time of year. The work asked whether data driven models can do so for 2m temperature, and later for precipitation, at lead times of weeks to months.
How it was done
The NeurIPS 2020 workshop paper tested two classes of model on predicting 2m temperature out to 52 weeks for six geographically diverse locations: a convolutional network, to use information related to teleconnections, and a recurrent network, to use long term historical signals. The EGU 2022 work built a daily probabilistic forecast of 2m temperature and total precipitation, combining physics based ensembles, climate modes and recent climatology as features for gradient boosting, U-Net and natural gradient boosting models.
What was found
- The 2020 models improved the accuracy of long range temperature forecasts up to a lead time of 30 weeks by correlation and 52 weeks by RMSE skill score, but only for select locations.
- The 2020 paper states that further work is needed for the models to add value beyond regions where climatology already has reduced correlation skill, namely the tropics.
- The 2022 abstract reports that the probabilistic model outperforms ECMWF 46 day forecasts and climatology.
Limitations
Skill in the 2020 work was location dependent. The 2022 result is a conference abstract rather than a peer reviewed paper.
What it might mean
Possible uses named in this line of work include climate risk assessment and agricultural and energy planning. These are applications, not results the papers demonstrate.
Publications
- Long-Range Seasonal Forecasting of 2m-Temperature with Machine Learning
EE Vos, A Gritzman, S Makhanya, T Mashinini, CD Watson. NeurIPS 2020 Workshop on Tackling Climate Change with Machine Learning, 2020.
- ML-based Probabilistic Prediction of 2m Temperature and Total Precipitation
MA Zaytar, B Zadrozny, C Watson, D Salles Civitarese, EE Vos, TM Mathonsi, TL Mashinini. EGU General Assembly 2022, Vienna, Austria, EGU22-11063, 2022.
doi:10.5194/egusphere-egu22-11063
Where it was applied
Lineage
- Led to
- Climate and geospatial risk continued in insurance, as flood and natural catastrophe risk models across more than 230,000 insured properties, and in the proposed doctoral research on flood mapping.
Sources for this page
Related research
- Shares deep learning, recurrent neural networksEcho state networks for level set image segmentationMSc research at Wits: can recurrent models learn the iterative evolution of a level set segmentation? Echo state networks against trained RNN, LSTM and GRU.
- Shares geospatial machine learning, climate riskProposed: self supervised flood mapping from radarProposed doctoral research, not yet registered: physics informed self supervised learning for flood extent mapping from synthetic aperture radar.
- Shares recurrent neural networksMachine learning for hearing loss risk in mine workersTwo 2019 IFAC papers: a recurrent network estimating hearing threshold shift in mine workers, and a classifier based noise policy advising system.
Related work
- Directly connectedClimate forecasting and geospatial ML at IBM ResearchApplied ML research at IBM Research: climate forecasting models deployed into IBM PAIRS Geoscope, published work, and a provincial COVID-19 risk dashboard.
- Shares time seriesTFiltersPy: Kalman and particle filters for PythonTFiltersPy is an open source Python library of Bayesian filters: Kalman, extended, unscented, ensemble and particle, behind one scikit-learn style API.
- Shares geospatial machine learning, climate riskInsurance data science: telematics, flood risk and MLOpsLeading insurance data science at ABSA Insurance: telematics processing cut from months to under a day, flood risk across 230,000+ properties, MLOps.