Zero‐Shot Transferability of Deep Learning Models for Precipitation Super‐Resolution Across Diverse Climate Regimes
- 1Department of Earth and Environmental Sciences Indian Institute of Science Education and Research Bhopal Bhopal India
Abstract
High‐resolution precipitation fields are essential for characterizing climate extremes, assessing sectoral impacts, and improving early warning systems. However, the resolution of global Earth system models (ESMs) and numerical weather prediction (NWP) systems remains limited by computational constraints. This study assesses the zero‐shot transferability of deep learning (DL)‐based super‐resolution (SR) models for precipitation across diverse geographic domains, data sources, and temporal scales. A suite of DL‐based SR architectures, ranging from convolutional baselines to spatial‐attention‐based generative models, is trained on a limited regional observational record of satellite‐derived precipitation from three monsoon‐dominated regions of the Indian subcontinent during 2001–2014. DL model generalization is assessed through out‐of‐sample transfer across climatically diverse domains using observations (2015–2023), NWP forecasts (2018–2023), and a 150‐year ESM simulation spanning 1950–2100, all without retraining or fine‐tuning. Across tropical, subtropical, and mid‐latitude climate regimes, the models show strong zero‐shot transferability, achieving Kling‐Gupta efficiency (KGE) scores approximately 0.85 or higher. Skill gains are evident in fine‐scale spatial variability and extreme precipitation statistics that are poorly represented in the coarse‐resolution inputs. Cross‐source application to weather forecasts and long‐term climate simulations further demonstrates the applicability of SR models across differing forcing data and evolving climate conditions. While differences among architectures are generally modest, spatial attention‐based models show enhanced performance in spatial fidelity and extreme‐event detection. This framework does not explicitly represent the physical processes governing precipitation formation or enforce physical constraints; within these limitations, its performance under our validation strategy indicates robustness and utility as an efficient post‐processing tool for weather and climate models.
Related
- A machine learning model for early season wildfire risk in Ghana’s Guinea savannah assessing discrimination, operational prevalence, spatial transferability, and CMIP6 climate scenarios
- PrecipCorrDiff: A Deep Learning Framework for Generating 0.0625∘ High-Resolution IMD Gridded Precipitation Data
- Climate models lack decadal variability in precipitation because of a weak atmospheric response to sea surface temperature anomalies
- Spatio-temporal precipitation downscaling at hourly resolution based on km-scale regional simulations
- Evidence from 2024 Storm Boris indicates anthropogenic climate change amplifies floods more than extreme precipitation
BibTeX
@article{murukesh2026zeroshot,
title = {{Zero‐Shot Transferability of Deep Learning Models for Precipitation Super‐Resolution Across Diverse Climate Regimes}},
author = {Midhun Murukesh and Pankaj Kumar},
journal = {Earth and Space Science},
year = {2026},
month = oct,
doi = {10.1029/2026ea005059},
url = {https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EA005059},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Zero‐Shot Transferability of Deep Learning Models for Precipitation Super‐Resolution Across Diverse Climate Regimes
AU - Midhun Murukesh
AU - Pankaj Kumar
PY - 2026
DA - 2026/10/27
JO - Earth and Space Science
DO - 10.1029/2026ea005059
UR - https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2026EA005059
AB - High‐resolution precipitation fields are essential for characterizing climate extremes, assessing sectoral impacts, and improving early warning systems. However, the resolution of global Earth system models (ESMs) and numerical weather prediction (NWP) systems remains limited by computational constraints. This study assesses the zero‐shot transferability of deep learning (DL)‐based super‐resolution (SR) models for precipitation across diverse geographic domains, data sources, and temporal scales. A suite of DL‐based SR architectures, ranging from convolutional baselines to spatial‐attention‐based generative models, is trained on a limited regional observational record of satellite‐derived precipitation from three monsoon‐dominated regions of the Indian subcontinent during 2001–2014. DL model generalization is assessed through out‐of‐sample transfer across climatically diverse domains using observations (2015–2023), NWP forecasts (2018–2023), and a 150‐year ESM simulation spanning 1950–2100, all without retraining or fine‐tuning. Across tropical, subtropical, and mid‐latitude climate regimes, the models show strong zero‐shot transferability, achieving Kling‐Gupta efficiency (KGE) scores approximately 0.85 or higher. Skill gains are evident in fine‐scale spatial variability and extreme precipitation statistics that are poorly represented in the coarse‐resolution inputs. Cross‐source application to weather forecasts and long‐term climate simulations further demonstrates the applicability of SR models across differing forcing data and evolving climate conditions. While differences among architectures are generally modest, spatial attention‐based models show enhanced performance in spatial fidelity and extreme‐event detection. This framework does not explicitly represent the physical processes governing precipitation formation or enforce physical constraints; within these limitations, its performance under our validation strategy indicates robustness and utility as an efficient post‐processing tool for weather and climate models.
SN - 2333-5084
SN - 2333-5084
ER -