Updated national climate projections for New Zealand

Downscaling involves taking relatively coarse resolution global climate model output (typically 100-150km resolution) and enhancing the resolution over New Zealand. ‘Dynamical downscaling’, as performed here, uses a physics-based atmospheric computer model to perform the downscaling. Downscaling is especially needed over New Zealand since coarse resolution global models have several limitations for representing our unique regional climate. Statistical downscaling approaches, that often simply perform grid cell-based corrections, have important limitations. Although much cheaper to run, these simple statistical approaches generally cannot capture important smaller scale feedbacks in the climate system, such as how long-term changes in soil moisture or snow cover impact temperature trends.

After the dynamical downscaling was run at 12km resolution over New Zealand, several key variables were bias corrected to NIWA’s Virtual Climate Station Network (VCSN) gridded observational data at 5km resolution (Campbell et al., 2024a). As such, these bias corrected fields will have similar observational accuracy to VCSN, when assessed in a climatological sense (i.e. long-term averages). The largest uncertainty in VCSN is in locations where the underlying station density is relatively low (i.e. in high elevation regions, away from main population centres). Prior to bias correction, the 12km dynamical downscaled runs were also comprehensively evaluated in terms of mean fields and extremes (Gibson et al., 2024a; Campbell et al., 2024b).

It is important to note that model errors in the historical period do not necessarily translate to expected errors in the future projections. Instead, a better way to assess future projection uncertainty is to look at the range of model outcomes (i.e. across the six downscaled models) for a given variable or season. For example, Figure 1 below shows the spread of projected changes in precipitation for Auckland – note that some models predict increasing precipitation (i.e. change greater than 0) and some models predicting decreasing precipitation (i.e. change less than 0). These changes and associated uncertainty can also vary across seasons.

For a given scenario (i.e. SSP) the largest uncertainty stems from the differences across the global climate models. While we targeted the top performing global climate models for the region, certain models can warm much faster or slower in response to the same amount of external forcing (i.e. ‘greenhouse’ gas emissions), and this is carried forward into the regional model projections. To account for this, we selected a range of different warming rates from the global models, aligned with the ‘likely range’ of warming from the latest IPCC sixth assessment report. The range across the downscaled projections (i.e. range from the six models) is therefore a practical and useful way to assess uncertainty in the projections, where each can be considered a plausible future outcome. Uncertainties also stem from the regional climate model uncertainty. For example, the assumptions that go into how rainfall is represented.

The downscaled model resolution (i.e. from CCAM) is approximately 12km over New Zealand, this is then bias corrected and produced on a final 5km national for key variables. The main limit to going to higher resolution (i.e. beyond 12km) in dynamical downscaling is computational resources. It costs roughly 8-10 times the computational resources for a doubling of model resolution. The downscaled model resolution is considerably improved compared to the earlier CMIP5 downscaling (approximately 35km resolution) produced by NIWA.

Stakeholders and partners should be aware that various high-resolution climate products are now becoming available for purchase. However, creating scientifically reliable climate projections is complex and requires rigorous methods. We recommend verifying that any climate product you use is supported by peer-reviewed research and based on repeatable scientific methods. Higher resolution does not always mean better quality, especially when produced by statistical approaches.

Base periods are used so that future changes can be assessed relative to a standard historical period that aligns with the current climatology or recent past. A commonly used historical baseline for CMIP6 is 1995–2014 (2014 being the last year of CMIP6 historical simulations). An additional baseline of 1986-2005 is provided for consistency with the pervious historical base period in CMIP5.

Climate models are initialised from the distant past and so are not designed to provide relevant information on any individual day, just as we would not expect a weather forecast to be reliable several months or years out. Instead, the appropriate way to evaluate climate models is to look at long-term statistics from the output of the models and compare this to observations. For example, comparing 30-year averaged temperatures from the model, or annual trends across a 30-year (and longer) time-period.

If you require climate data to represent a different period and location (i.e., a time series) please reach out the NIWA for assistance with this. Additionally, NIWA can help develop bespoke climate indicators for a particular application.

The CMIP6 downscaling has various general improvements and advantages over the previous CMIP5 downscaling. The primary regional climate model is run at substantially higher resolution both over New Zealand (12km) and the wider South Pacific domain, enabled by the stretched grid configuration of the model. The stretched grid configuration may additionally help to more seamlessly resolve certain types of storms including ex-Tropical Cyclones, as they approach New Zealand. The biases from the updated CMIP6 downscaling were comprehensively evaluated and benchmarked against other candidate regional climate models, with these results published in the peer-reviewed literature (Gibson et al., 2024a; Campbell et al., 2024b). The global climate models from CMIP6 also have generally higher resolution relative to CMIP5, with some improvements in large-scale circulation features that are highly relevant to New Zealand (Simpson et al., 2020; Curtis et al., 2020). The CMIP6 downscaling is also driven by more up to date forcings, as included in the SSPs that replaced the RCPs from CMIP5. Lastly, through collaboration between NIWA and the Ministry for the Environment (MfE), the updated downscaled data has an improved pathway for data access and visualisation, available here.

This highly depends on the region, variable and season being considered. A comprehensive comparison is provided in Broadbent et al. (2025). While the main spatial patterns of change are similar (e.g. wetter in the west, and drier in the east and north of the country), regional differences can be important. The updated projections show slightly more warming by mid-century and end-of-century, and the drying signal over the North Island is more pronounced.

The limiting factor is from the global climate models in CMIP6 used for downscaling, where most end in year 2100 (with a final date of 31 December 2099). Care is required when extrapolating trends beyond this through statistical approaches, since projected changes can be non-linear.

We have chosen six of the top performing GCMs to downscale. Further details of the GCM selection criteria are provided in Gibson et al. (2024). The main limiting factor is computational resources that go into performing the downscaling. Complementary research is underway at NIWA to use artificial intelligence (AI) and machine learning (ML) to downscale a larger ensemble to complement this effort (Rampal et al., 2024, 2025). While still under development, these more modern empirical approaches can have a number of benefits over traditional statistical downscaling (Rampal et al., 2022).

NIWA can help with additional interpretation and analysis of the data, including translating this climate information to respond to adaptation, planning and reporting needs. You can find additional information on NIWA’s services related to this data here.

There is generally no single best performing GCM. It is best practice to consider the spread across all downscaled GCMs.

NIWA has prioritised downscaling SSP126, SSP245, and SSP370 scenarios. This is consistent with regional climate downscaling efforts in other regions and recommended by CORDEX. Broadly speaking, the SSP245 scenario can be interpreted as a ‘middle-of-the-road scenario’, while SSP370 can be interpreted as a relatively high emissions scenario.

Additionally, NIWA has conducted additional downscaling for MfE using SSP585. We suggest that the main value in using SSP585 lies in sampling very low probability and high impact future weather/climate events. However, this scenario should not be treated as ‘business-as-usual’ or the most likely scenario (Hausfather and Peters, 2020).

Where models disagree on the sign of changes (e.g. whether rainfall in summer is projected to increase or decrease for a given region) this generally means that we have less confidence in the particular projection. This stems from differences in the projections from the global climate models, such as how dominant weather patterns are projected to change.

These climate projections focus on changes over land. Sea-level rise and costal hazards projections for New Zealand have been recently produced under the NZ SeaRise Endeavour Research Programme funded by MBIE.

These climate projections focus on changes over land. River flow and flood risk changes are being investigated under a NIWA-led Endeavour Research Programme funded by MBIE.

The climate change component of HIRDS focuses on very extreme rainfall events based on return periods. The climate projections here are much broader than this (e.g. spanning different meteorological variables, regions, seasons, and indices). In the future, it is anticipated that these updated CMIP6 projections will be used as inputs into future releases of HIRDS.

Please cite the following:

Gibson, P. B., Stuart, S., Sood, A., Stone, D., Rampal, N., Lewis, H., Broadbent, A., Thatcher, M., & Morgenstern, O. (2024). Dynamical downscaling CMIP6 models over New Zealand: added value of climatology and extremes. Climate Dynamics, 27p.

Gibson, P. B., Campbell, I., Lewis, H., Rampal, N., Fedaeff, N., Woolley, J-M. (2024). User Guidance of CMIP6 Downscaled Data for Aotearoa New Zealand. NIWA Report prepared for MfE, 20p.

Gibson, P. B., Campbell, I., Lewis, H., Rampal, N., Fedaeff, N., Woolley, J-M. (2024b). User Guidance of CMIP6 Downscaled Data for Aotearoa New Zealand. NIWA Report prepared for MfE, 20p.

Campbell, I., Gibson, P.B., Rampal, N., (2024a). Bias Correction of Downscaled CMIP6 Output. NIWA Report prepared for MfE, 40p.

Gibson, P. B., Stuart, S., Sood, A., Stone, D., Rampal, N., Lewis, H., Broadbent, A., Thatcher, M., & Morgenstern, O. (2024a). Dynamical downscaling CMIP6 models over New Zealand: added value of climatology and extremes. Climate Dynamics, 27p.

Campbell, I., Gibson, P. B., Stuart, S., Broadbent, A. M., Sood, A., Pirooz, A. A., & Rampal, N. (2024b). Comparison of three reanalysis‐driven regional climate models over New Zealand: Climatology and extreme events. International Journal of Climatology, 44(12), 4219-4244.

Gibson, P.B., Rampal, N., Dean, S.M., & Morgenstern, O. (2024) Storylines for Future Projections of Precipitation Over New Zealand in CMIP6 Models. JGR Atmospheres, 129(5), e2023JD039664

Rampal, N. et al. (2024). Enhancing Regional Climate Downscaling through Advances in Machine Learning. Artificial Intelligence for the Earth Systems 3(2)

Gibson, P.B. et al. (2023) High-Resolution CCAM Simulations Over New Zealand and the South Pacific for the Detection and Attribution of Weather Extremes. JGR Atmospheres 128(14) e2023JD038530

Simpson, I. R., Bacmeister, J., Neale, R. B., Hannay, C., Gettelman, A., Garcia, R. R., ... & Richter, J. H. (2020). An evaluation of the large‐scale atmospheric circulation and its variability in CESM2 and other CMIP models. Journal of Geophysical Research: Atmospheres, 125(13), e2020JD032835.

Curtis, P. E., Ceppi, P., & Zappa, G. (2020). Role of the mean state for the Southern Hemispheric jet stream response to CO2 forcing in CMIP6 models. Environmental Research Letters, 15(6), 064011.

Rampal, N., Gibson, P. B., Sood, A., Stuart, S., Fauchereau, N. C., Brandolino, C., ... & Meyers, T. (2022). High-resolution downscaling with interpretable deep learning: Rainfall extremes over New Zealand. Weather and Climate Extremes, 38, 100525.

Rampal, N., Gibson, P. B., Sherwood, S., & Abramowitz, G. (2024). On the extrapolation of generative adversarial networks for downscaling precipitation extremes in warmer climates. Geophysical Research Letters, 51(23), e2024GL112492.

Rampal, N., Gibson, P. B., Sherwood, S., Abramowitz, G., & Hobeichi, S. (2025). A reliable generative adversarial network approach for climate downscaling and weather generation. Journal of Advances in Modeling Earth Systems, 17(1), e2024MS004668.

Hausfather, Z., & Peters, G. P. (2020). RCP8. 5 is a problematic scenario for near-term emissions. Proceedings of the National Academy of Sciences, 117(45), 27791-27792.

Broadbent et al. (2025). How do New Zealand’s downscaled CMIP6 climate projections compare with CMIP5? Weather and Climate (accepted for publication)