More accurate and locally relevant seasonal climate forecasts

  • Hello everyone

    Just wanted to say that I am currently leading a research project aimed at trying to develop better (more accurate) and more locally relevant (‘downscaled’) seasonal climate forecasts using ML. In a nutshell the project is / will explore the potential of several ML approaches (including Deep Learning) to derive more local seasonal climate forecast products. We are using large scale fields from several GCMs (e.g. ECMWF, UKMO, Meteo-France, NCEP-CFSv2, etc …) as inputs (with some experimentations with different feature engineering methods) and ‘local’, observed, seasonally aggregated climate variables as target variables. The geographical focus of this project is New Zealand, where I live and work, but the approaches are generalizable.

    Even though this project is not strictly speaking aimed at the climate change problem, I think better climate services, including better forecasts at different time-scales, have an important role to play in adaptation strategies.

    Happy to hear if other people have similar projects or are interested in discussing more

    Also wanted to say thanks to everyone involved in building this forum …


  • @NicolasF Welcome and thanks for sharing your project with us. This is an extremely interesting piece of work and adaptation strategies are going to be quite relevant to climate discussions as we progress in the future.

    What drove your team towards deciding to use machine learning to solve this problem? Did you feel that established forecasting techniques could be substantially improved with ML? Or is that the hypothesis of your project? In any case kudos is deserved. There aren't enough people working on this type of problem.

  • @Eric-Vanular the main motivation for this project is that the current generation of General Circulation Models (GCMs) used to generate seasonal climate forecasts are still limited by a relatively coarse spatial resolution (typically 2.5 degrees, at most 1 degree in latitude / longitude), this means GCMs are missing important orographic and land-atmosphere interactions, which are highly relevant to New Zealand (or any 'small' country with complex topography ...). They are therefore not well-adapted for the provision of local-scale climate forecasts. The idea is that we use the information contained in the well-validated, large-scale fields from the GCMs (SSTs, geopotential, ...) to train ML models along with observed, seasonally integrated climate variables as target variables, therefore providing potentially more accurate, and also more locally relevant seasonal climate forecast products.

  • @NicolasF Makes total sense, thanks for explaining. Sounds like you're combining top down (GCMs) and bottom up (observed climate records) approaches to converge on a more accurate predictor. What level of locality are you hoping to achieve?

  • @NicolasF Awesome project. How do you think this could help with adaptation strategies?

  • @NicolasF I imagine that this will be very relevant in the locations most affected by climate change (i.e. coastal regions). What are the biggest challenges you're experiencing so far?

  • @Pascal-Ramsey Hi Pascal ... so far the challenges have been related to the 'harmonisation' of the different GCMs hindcasts use to train the models: they come in various formats (grib, netcdf) with various conventions, variable and dimension names etc., and sometimes the documentation available on how the data is organized in these datasets is a bit sparse :-). Also thinking hard about how to structure the various tasks involved in the project, from pre-processing to processing, to feature extraction, model building, evaluation, testing etc ... there are lots of moving parts, and we want to release the code on github, but need to ensure it will be structured in a way that other people could take it and adapt it to their needs.

  • @Eric-Vanular Hi Eric, the district level is our first target (see and we have also some application cases with similar spatial scales but across districts (e.g. fruit producing regions).

  • @Sam-Hughes Hi Sam, the thinking in a nutshell is that if people can anticipate climate extremes (e.g. droughts, floods, warm seasons) they can take actions to mitigate their impacts, more accurate and locally relevant seasonal climate forecasts would help with this ...

Log in to reply