On the Extrapolation of Generative Adversarial Networks for Downscaling Precipitation Extremes in Warmer Climates DOI Creative Commons
Neelesh Rampal, Peter B. Gibson, Steven C. Sherwood

et al.

Geophysical Research Letters, Journal Year: 2024, Volume and Issue: 51(23)

Published: Dec. 5, 2024

Abstract While deep‐learning downscaling algorithms can generate fine‐scale climate projections cost‐effectively, it is unclear how effectively they extrapolate to unobserved climates. We assess the extrapolation capabilities of a deterministic Convolutional Neural Network baseline and Generative Adversarial (GAN) built with this baseline, trained predict daily precipitation simulated by Regional Climate Model (RCM) over New Zealand. Both approaches emulate future changes in annual mean well, when on historical data, though training improves performance. For extreme (99.5th percentile), RCM simulations robust end‐of‐century increase warming (∼5.8%/C average from five simulations). When climate, GANs capture 97% warming‐driven compared 65% baseline. Even historically 77% increase. Overall, offer better generalization for extremes, which important applications relying data.

Language: Английский

Dynamical downscaling CMIP6 models over New Zealand: added value of climatology and extremes DOI Creative Commons
Peter B. Gibson, Stephen Stuart, Abha Sood

et al.

Climate Dynamics, Journal Year: 2024, Volume and Issue: 62(8), P. 8255 - 8281

Published: July 17, 2024

Abstract Dynamical downscaling provides physics-based high-resolution climate change projections across regional and local scales. This is particularly important for island nations characterized by complex terrain, where the coarse resolution of global model (GCM) output often prohibits direct use. One main motivations dynamical to reduce biases relative host GCM at scale, which can be quantified through assessing ‘added value’. However, added value from not guaranteed; quantifying this help users make informed decisions about how best use available projection data. Here we describe experiment design updated national New Zealand based on downscaling. The non-hydrostatic Conformal Cubic Atmospheric Model (CCAM) primarily used downscaling, with a stretched grid targeting high over (12-km) wider South Pacific region (12–35-km). Focusing historical simulations, assess range metrics, climatological fields, extreme indices, tropical cyclones. strengths include generally large improvements temperature orographic precipitation. Inter-annual variability in well captured Zealand, several precipitation-based indices show improvements. representation cyclones reaching least category 2 intensity improved consistent under-representation GCMs. remaining are explored discussed forming basis ongoing bias-correction work.

Language: Английский

Citations

8
A Reliable Generative Adversarial Network Approach for Climate Downscaling and Weather Generation DOI Creative Commons
Neelesh Rampal, Peter B. Gibson, Steven C. Sherwood

et al.

Journal of Advances in Modeling Earth Systems, Journal Year: 2025, Volume and Issue: 17(1)

Published: Jan. 1, 2025

Abstract Anticipating climate impacts and risks in present or future climates requires predicting the statistics of high‐impact weather events at fine‐scales. Direct numerical simulations fine‐scale are computationally too expensive for many applications. While deterministic‐based (deep‐learning statistical) downscaling low‐resolution several orders magnitude faster than direct simulations, it suffers from limitations. These limitations include tendency to regress mean, which produces excessively smooth predictions underestimates extreme events. They also fail preserve statistical measures that key research. We use a conditional GAN (cGAN) architecture downscale daily precipitation as Regional Climate Model (RCM) emulator. The cGAN generates plausible residuals on top predictable expectation state produced by deterministic deep learning algorithm. skill cGANs is highly sensitive hyperparameter known weight adversarial loss (), where value required accurate results varies with season performance metric, casting doubt reliability usually implemented. However, applying simple intensity constraint function, possible obtain reliable across spanning two magnitude. CGANs considerably more skillful capturing climatological statistics, including distribution spatial characteristics With this modification, we expect be readily transferable other applications time periods, making them useful generator representing event climates.

Language: Английский

Citations

0
A Reliable Generative Adversarial Network Approach for Climate Downscaling and Weather Generation DOI
Neelesh Rampal, Peter B. Gibson, Steven C. Sherwood

et al.

Authorea (Authorea), Journal Year: 2024, Volume and Issue: unknown

Published: Aug. 28, 2024

Anticipating climate impacts and risks in present or future climates requires predicting the statistics of high-impact weather events at fine-scales. Direct numerical simulations fine-scale are computationally too expensive for many applications. While deterministic-based (deep-learning statistical) downscaling low-resolution several orders magnitude faster than direct simulations, it suffers from limitations. These limitations include tendency to regress mean, which produces excessively smooth predictions underestimates extreme events. They also fail preserve statistical measures that key research. We use a conditional GAN (cGAN) architecture downscale daily precipitation as Regional Climate Model (RCM) emulator. The cGAN generates plausible residuals on top predictable expectation state produced by deterministic deep learning algorithm. skill cGANs is highly sensitive hyperparameter known weight adversarial loss (\(\lambda_{adv}\)), where value \(\lambda_{adv}\) required accurate results varies with season performance metric, casting doubt reliability usually implemented. However, applying simple intensity constraint function, possible obtain reliable across spanning two magnitude. CGANs considerably more skillful capturing climatological statistics, including distribution spatial characteristics With this modification, we expect be readily transferable other applications time periods, making them useful generator representing event climates.

Language: Английский

Citations

2
On the Extrapolation of Generative Adversarial Networks for Downscaling Precipitation Extremes in Warmer Climates DOI Creative Commons
Neelesh Rampal, Peter B. Gibson, Steven C. Sherwood

et al.

Geophysical Research Letters, Journal Year: 2024, Volume and Issue: 51(23)

Published: Dec. 5, 2024

Abstract While deep‐learning downscaling algorithms can generate fine‐scale climate projections cost‐effectively, it is unclear how effectively they extrapolate to unobserved climates. We assess the extrapolation capabilities of a deterministic Convolutional Neural Network baseline and Generative Adversarial (GAN) built with this baseline, trained predict daily precipitation simulated by Regional Climate Model (RCM) over New Zealand. Both approaches emulate future changes in annual mean well, when on historical data, though training improves performance. For extreme (99.5th percentile), RCM simulations robust end‐of‐century increase warming (∼5.8%/C average from five simulations). When climate, GANs capture 97% warming‐driven compared 65% baseline. Even historically 77% increase. Overall, offer better generalization for extremes, which important applications relying data.

Language: Английский

Citations

1