News & Updates

Tairāwhiti Hui
June 2023

DeepWeather project members Greg Bodeker, Mark Schwarz, Tui Warmenhoven and Emily O’Riordan visited several communities in Tairāwhiti at the end of June 2023, a trip that has been rescheduled several times due to extreme weather events in the district. Tairāwhiti has been severely impacted by heavy rainfall events and cyclones over the past few years, and faces a future with more frequent extreme weather exacerbated by climate change.

The purpose of the trip was not only to introduce the DeepWeather project to its end-users, but to seek input as to how this project can be developed to best meet their needs. The team met with the Ruatoria Civil Defence Emergency Management (CDEM) group, of which Tui is a member, to discuss how recent weather events have affected them, both individually and as a community. The hui was the start of a working relationship with the CDEM group, whereby the DeepWeather team will seek to regularly engage with the community to share ideas and knowledge.

The team met with communities that have been particularly impacted by recent weather events, including the residents of Makarika and Anaura Bay. Both communities commented on a lack of communication from decision-makers during extreme weather events, as well as the need for funding to mitigate the impacts of such events. The need for reliable energy sources to prevent power outages was discussed, and the possibility of investigating funding for solar panels and gas was raised. Such conversations will be explored by the DeepWeather team and fed back to government officials in our MBIE reporting.

The trip also included outreach at local schools. The team first attended Tolaga Bay Area School, where students gave their kōrero about recent weather events, and asked questions about the weather, artificial intelligence and science in general. At Ngata Memorial College, an interactive session explained how a warm front moves through the atmosphere, with students acting as grid cells within a weather model. At both schools, the idea to equip volunteering students with their own rain gauges to track local rainfall was raised, allowing students to track rainfall at their house or school and compare with others in the area. This could make for an engaging science project, provide useful data for the DeepWeather project, and possibly kickstart an interest in becoming a weather scientist! 

The DeepWeather team are extremely grateful to the individuals and communities who gave their time to help us understand the current situation and concerns in Tairāwhiti. Special thanks go to Tui Warmenhoven for her mahi in coordinating this hui. 

DeepWeather Project Update
June 2023

The DeepWeather project team have been collecting and curating appropriate data sources to initialise and train the artificial neural network model. Thanks to MetService and University of Canterbury, we have access to a number of forecast realisations from different models and at different scales. MetService also have access to observation data, such as gauge-corrected rain radar data.

An initial basic model has been developed using XGBoost, taking Weather Research and Forecasting (WRF) model forecasts from MetService as input. Initial tests using this basic model have shown promising results using only a limited source of data; see the figure below for examples of using this model to downscale coarse-resolution precipitation and temperature fields. The downscaled images on the right show better resolved fields at 4km resolution over the whole of New Zealand. The DeepWeather project aims to downscale to higher resolutions than this, but this initial XGBoost model allows us to build a pipeline between incoming data and model outputs. Using this model and domain expertise of the team, we have also been able to select key variables which, when using only a data-driven approach, best predict our selected target variables.

Initial outputs from a simple XGBoost model. A 20km resolution precipitation field (top-left) has been downscaled to a 4km resolution precipitation field (top-right) using the XGBoost model. A 20km resolution temperature field (bottom-left) has been downscaled to a 4km resolution temperature field (bottom-right). 

The immediate goals for the near future are to start building a more complex and comprehensive model, including considerations such as physics constraints as well as physical features such as topography and land use. We will investigate the outcomes of different architectures, as well as the effectiveness of transfer learning to allow for computationally efficient training of the model.

12th International Conference on Climate Informatics
April 2023

DeepWeather's partner organisation, British Antarctic Survey, co-hosted the 12th International Conference on Climate Informatics  at the University of Cambridge in April.  DeepWeather team members Scott Hosking, Tom Andersson, and Emily O'Riordan attended what was a fantastic conference, which brought together international researchers and users in climate science, data science, and computer science, to share state-of-the-art developments in climate data and informatics. The aim of the conference series is to accelerate the rate of discovery in climate science and adaptation of climate applications. 

To learn more about Climate Informatics, please visit:

This was also a wonderful opportunity for members of our team in the UK and NZ to meet in-person for the first time!

DeepWeather stakeholder meeting 14 Dec 2022

Stakeholder Meeting
14 December 2022

Thank you to everyone who attended our first stakeholder meeting. All the feedback and suggestions provided were very useful in shaping the programme to generate hyperlocal weather forecasts to suit your needs. The slides from the meeting are available on the left, in case you missed anything.

If you'd like to share further thoughts, please email