The large computational cost of Regional Climate Models (RCMs) means that only one ensemble member per climate model is typically downscaled; subsequently, internal variability uncertainty is generally not explicitly accounted for in coordinated regional climate downscaling efforts (e.g., CORDEX). Surrogate Artificial Intelligence-based emulators are several orders of magnitude faster than RCMs and have been well-tested in their ability to generate reliable regional climate projections. This study employs a Generative AI-based approach using Generative Adversarial Networks (GANs) to downscale daily precipitation from a large ensemble of climate projections from CanESM5 (n=20) and ACCESS-ESM-1-5 (n=40) at a 12km resolution for New Zealand. We show that this AI-based approach can reproduce key features including rainfall extremes and their increases in future climates with useful accuracy. Similar to previous studies using low-resolution climate models, our results show robust future changes in winter precipitation across the ensemble members, but significant uncertainty during summer. The large ensemble of downscaled climate projections better samples extremely rare localized extreme events, which are not adequately sampled using a single ensemble member. Using this ensemble, we can calculate the relative contributions of internal variability and model structural uncertainty (both GCM and downscaling) in climate projections of local-scale impact-relevant weather events. Overall, our study highlights the significant potential of AI to complete dynamical downscaling and allow quantification of internal variability uncertainty at regional scales.
How to cite: Sherwood, S., Rampal, N., and Gibson, P.: Quantifying Internal Variability Uncertainty in Regional Climate Projections using Artificial Intelligence, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-14590, https://doi.org/10.5194/egusphere-egu25-14590, 2025.
Statistical bias adjustment of climate models has become widespread practice to bridge the usability gap of climate information for impact studies and other societal applications. However, the application of bias adjustment offers potential for misuse and comes with several fundamental issues which have been highlighted in the literature. In this tension between widespread use and fundamental issues, different strategies for the application of statistical bias adjustment have developed, ranging from selecting a consistent bias adjustment method across applications to ensure comparability, to applying an ensemble of available methods in a given case study. In this contribution, we examine the specific methodological assumptions of different approaches to bias adjustment, such as the relevance and potential for trend preservation, and propose an evaluative framework based on recent literature in philosophy of science to assess the understanding of usability underlying different approaches to bias adjustment. We find that both methodological assumptions about bias adjustment, as well as the understanding of usability in the context of climate information determine the choice of bias adjustment strategy in current practice. For example, global application of a bias adjustment method generates information that is salient and credible and thus usable mostly for the purpose of model intercomparison, whilst local adaptation improves credibility, but compromises on the ease-of-use. With neither the methodological assumptions nor the understanding of what usable climate information is and who it is generated for often explicitly stated in practice, we hope to contribute to enhanced methodological practice and reflection through this discussion.
How to cite: Wessel, J. B., Spuler, F. R., Jebeile, J., and Shepherd, T. G.: Statistical bias adjustment and the usability of climate information: a perspective on strategies and underlying assumptions, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19329, https://doi.org/10.5194/egusphere-egu25-19329, 2025.
Besides regional climate models (RCMs), there exist two main approaches to tackle the insufficient resolution of global climate models: emulators and statistical downscaling. While both approaches are similar in the techniques they use (statistical and machine learning, ML, methods), they differ in their objectives and underlying assumptions. Emulators are intended to provide a cost-effective alternative to RCMs by emulating their downscaling functions. Alternatively, statistical downscaling (SD) models learn the empirical (observed) relationships that link a set of key large-scale predictors, to the local high-resolution predictand of interest. There is a key tradeoff between these two approaches: emulators are unconstrained by observed climate records, yet they also inherit RCM biases; conversely, SD methods are able to produce potentially bias-free simulations (at least when driven by reanalyses), but with extrapolation constrained by observed relationships.
This tradeoff between extrapolation and bias is a key research perspective, especially when compounded with the usual additional challenges ML methods face, like representation of extremes or the temporal/spatial consistency of the predictions. Within this context, the added value of generative/stochastic methods is highly relevant and timely. Indeed, recent studies using deterministic ML methods (such as Wang et al. 2023; Doury et al. 2024) have highlighted that emulating high-resolution fields does require generative/stochastic approaches, specially when it comes to representing extreme weather events for complex variables like precipitation (Watson, 2022, 2023). However, while generative methods such as diffusion models may offer an advantage when it comes to simulating extremes (Addison et al. 2024; Aich et al. 2024), they are also subject to more potential instability (e.g., diffusion models are known to have hallucinations, Aithal et al. 2024), hence also increasing the biases.
In this study we aim to address the added value and potential downsides generative/stochastic ML methods can bring to the field of statistical downscaling and emulation, by targeting the tradeoff between extrapolation and bias. Therefore, we will address both already well-established generative deep learning techniques and the latest generation diffusion models, and focus on how well they fare when capturing aspects beyond mean statistics, including extremes, which are of particular interest in terms of climate impacts.
References:
Addison, H. et al. (2024). Machine learning emulation of precipitation from km-scale regional climate simulations using a diffusion model. Preprint. DOI: https://doi.org/10.48550/arXiv.2407.14158
Aich, M. et al. (2024). Conditional diffusion models for downscaling & bias correction of Earth system model precipitation. Preprint. DOI: https://doi.org/10.48550/arXiv.2404.14416
Aithal, S. K. et al. (2024). Understanding Hallucinations in Diffusion Models through Mode Interpolation. Preprint. DOI: https://doi.org/10.48550/arXiv.2406.09358
Doury, A. et al. (2024). On the suitability of a convolutional neural network based RCM-emulator for fine spatio-temporal precipitation. Climate Dynamics, 62(9), 8587-8613. DOI: https://doi.org/10.1007/s00382-024-07350-8
Watson P. A. G. (2022). Machine learning applications for weather and climate need greater focus on extremes. Environmental Research Letters 17(11). DOI: https://doi.org/10.1088/1748-9326/ac9d4e
Watson, P. (2023). Machine learning applications for weather and climate predictions need greater focus on extremes: 2023 update. NeurIPS 2023 Workshop on Tackling Climate Change with Machine Learning.
Acknowledgement:
This work is funded by the National Research Agency under France 2030 bearing the references ANR-22-EXTR-0005 (TRACCS-PC4-EXTENDING project) and ANR-22-EXTR-0011 (TRACCS-PC10-LOCALISING project).
How to cite: Legasa, M. N., Lguensat, R., and Vrac, M.: Statistical Downscaling and Emulators: Can Generative Machine Learning add Value to Extrapolation and Bias?, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11385, https://doi.org/10.5194/egusphere-egu25-11385, 2025.
Generative deep learning models have shown remarkable skill in the probabilistic downscaling of climate and weather forecasts, with generative adversarial networks (GANs) as a particularly effective approach for precipitation downscaling. However, most existing methods are trained for specific regions, and their performance on unseen geographic areas remains largely unexplored. In our work, we evaluate the transferability of generative models to new locations outside their training domain. Using a global experimental setup, we employ ERA5 as the predictor dataset and IMERG as the high-resolution target dataset at 0.1° resolution. To systematically assess the performance across diverse regions, we design a hierarchical location split with 16 regions. We then train networks independently on the 16 regions and evaluate each of them on all others. Our findings provide insights on the robustness and limitations of generative models for global-scale precipitation downscaling, revealing challenges such as poor generalization to unseen orography and decreased performance in tropical regions, both for models applied in these areas and for those trained in the tropics and transferred elsewhere.
How to cite: Harder, P., Lessig, C., Chantry, M., Pelletier, F., and Rolnick, D.: Global Location Transferability of Generative Deep Learning Models for Precipitation Downscaling, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-21537, https://doi.org/10.5194/egusphere-egu25-21537, 2025.
Over the past decade, advancements in high-performance computing have led Machine Learning (ML) to play a key role in enhancing Earth System Models (ESMs), enabling progress beyond the current state-of-the-art. Downscaling techniques to generate high-resolution data starting from the results of large-scale models are one of the most promising Deep Learning (DL) applications for ESMs. This approach offers a computationally efficient alternative to numerical dynamical downscaling, particularly for climate projections.
In this study, we present the application of a state-of-the-art DL model to emulate the dynamical downscaling of 6-hourly climate data, focusing on precipitation and minimum and maximum temperatures. The model is trained to reconstruct fields at a 4 km resolution, starting from dynamical predictors at ~100 km resolution. Training data consists of coarsened ERA5 reanalysis data (Hersbach et al., 2018) as predictors and high-resolution target data from the COSMO-CLM dynamical reanalysis for Italy (Raffa et al., 2021). We utilize 40 years of 6-hourly data (1981–2020) for training.
This training setup is designed to prepare the model for inference on low-resolution outputs from a selection of diverse climate projections and decadal predictions. The ultimate goal is to generate an ensemble of high-resolution projections that deliver additional insights, particularly into extreme values, at a fraction of the computational cost of regional climate models.
The DL architecture employed is a recently developed Latent Diffusion Model applied with a residual approach (Tomasi et al. 2024), which has demonstrated exceptional performance in downscaling continuous variables, such as 2-m temperature and 10-m wind speed components. Results are compared against other ML models (e.g., UNET) and available numerical regional climate models for benchmarking. Preliminary results are presented, highlighting (i) the enhancements introduced by the LDM architecture compared to baseline models, (ii) its ability to reconstruct coherent structures and extreme values, and (iii) the added value of the high-resolution data obtained by the application of the LDM to low-resolution climate projections.
How to cite: Tomasi, E., Franch, G., Calmanti, S., and Cristoforetti, M.: AI for high-resolution climate data: downscaling climate projections and decadal predictions with a deep learning Latent Diffusion Model , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16520, https://doi.org/10.5194/egusphere-egu25-16520, 2025.
The National Partnership for Climate Projections (NPCP) aims to develop a consistent approach to deliver comparable, robust, fit-for-purpose future climate information to assess climate risks and inform adaptation planning. The NPCP climate projections roadmap identifies a number of priority areas of collaboration, including the delivery of national and regional downscaled climate projections. This involves selecting global climate models (GCMs), downscaling using regional climate models (RCMs), bias-adjusting model outputs, and conducting secondary and next-level analysis (e.g., impact modelling).
The focus on bias-adjustment is an acknowledgement of the fact that GCM and RCM outputs often show significant discrepancies when compared to observations. These systematic errors, or biases, can render raw outputs unsuitable for direct use in downstream impact models such as those for hydrology and agriculture, as well as in climate risk assessments. For the NPCP bias-adjustment intercomparison project, we evaluated various bias-adjustment techniques currently in use in the Australian climate research community. These include Equi-distant/ratio Cumulative Density Function matching (ECDFm), Quantile Matching for Extremes (QME), N-Dimensional Multivariate Bias Correction (MBCn), and Multivariate Recursive Nesting Bias Correction (MRNBC).
While previous studies have assessed some of these techniques for specific metrics and applications in Australia, our evaluation aimed to be broad and comprehensive. The participating techniques were applied to daily RCM data from the CORDEX-CMIP6 project for a baseline task, where bias-adjusted data were produced for the 1980-2019 period using 1980-2019 as a training period, as well as a cross-validation task, where data were produced for 1990-2019 using 1960-1989 for training. These bias-adjusted data were then compared to observations across Australia on various metrics relating to temperature and precipitation climatology, variability, statistical distribution and extremes. The impact of bias-adjustment on simulated trends was also assessed by producing bias-adjusted data for the 2060-2099 period. Additionally, we compared the bias-adjustment techniques with a simple quantile delta change approach and investigated scenarios where it may be sufficient to directly bias-adjust GCM data without the need for computationally expensive downscaling.
Based on the results of the intercomparison, the best-performing techniques were subsequently used by the Australian Climate Service (ACS) to bias-adjust outputs from the CORDEX-CMIP6 archive. This ensures the availability of a consistent set of high-resolution, bias-adjusted products for the Australian community to evaluate climate hazards and risks, and support adaptation planning.
How to cite: Takbash, A., Irving, D., Peter, J., Dao, T. L., Kapoor, A., Gammon, A., Dowdy, A., Black, M., Bende-Michl, U., Jakob, D., and Grose, M.: A Comprehensive Assessment of Climate Data Bias-Adjustment Techniques Over Australia, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-252, https://doi.org/10.5194/egusphere-egu25-252, 2025.
As part of its responsibilities within the Spanish National Climate Change Adaptation Plan (PNACC) 2021-2030, AEMET generates and makes available to the public, through its website, climate change scenario information for Spain using statistical methods. These methods require a robust and sufficiently long observational database to enable proper training and validation, which has traditionally constrained their application to temperature and precipitation. However, the adaptation community requires information on a broader set of essential climate variables to adequately characterise the impacts of climate change on each sector. Recent studies using Artificial Intelligence show potential to generate downscaled information for a broader set of variables. However, long records from other ECV are scarce, relying on reanalysis information for training the methods.
Advances in modelling, on the other hand, have made available regional reanalysis products sich as COSMO reanalysis (Bollmeyer et al., 2015), CERRA (Schimanke et al., 2021), and ERA5-LAND (Muñoz-Sabater et al., 2024). These types of products provide historical information on a wide range of Essential Climate Variables (ECVs), offering extensive spatial coverage and physical consistency.
This study evaluates the performance of various available reanalysis products as a preliminary step towards selecting the most suitable dataset for generating high-resolution scenario information for a comprehensive set of Essential Climate Variables. Despite the focus on a complete set of ECVs, the study will focus on precipitation, as it is the variable for which AEMET has the most comprehensive data network. Different domains across the Iberian Peninsula will be analysed, with particular
How to cite: Rodríguez-Guisado, E., Gutiérrez-Fernández, J., Ortega, M., Rodríguez-Muñoz, I., Hernanz, A., and Correa-Guinea, C.: Comparative Analysis of Daily Precipitation Using High-Resolution Reanalysis Data, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-17918, https://doi.org/10.5194/egusphere-egu25-17918, 2025.
Empirical-statistical Downscaling (SD) techniques are valuable tools able to generate high-resolution climate information needed to carry out impact studies. In this regard, Convolutional Neural Networks (CNNs) are promising SD techniques capable of handling large amounts of data and extracting relevant predictor information for each particular site. These characteristics of the CNN represent a major advantage over traditional SD methods, which typically rely on human-guided predictor selection. Notwithstanding, an adequate tuning of the CNN is key for optimising their potential.
In southern South America (SSA), CNNs has proven to be skilful in representing daily extreme temperatures and extrapolating into future scenarios. Although the selection of the activation function introduces a source of uncertainty in the future projections.
In this context, this study aims to explore the use of CNNs as a statistical downscaling tool to simulate the wet bulb temperature (Tw) over SSA, a multivariate heat-stress index estimated from temperature and humidity. Tw has been widely used as a heat-stress proxy in different parts of the world, however, its characterisation and modelling in SSA remain as a pending task. To this end, four different CNN architectures regarding the activation function (ReLU or linear), domain size and configuration of the CNN layers were tested. All CNN models were trained during summer days using a cross-validation (CV) scheme in the period 1991-2020 and then evaluated in four unseen summers between 2021 and 2024. For comparison purposes, CNN models were similarly trained and validated to simulate maximum temperature (Tx).
Overall, CNN models well represented all the features evaluated, including the heat-waves that took place in the summers evaluated independently. In particular, CNN models presents a better performance in simulating Tw than Tx with smaller errors in terms of mean and extremes aspects. Regarding the domain size, for both temperatures, the configuration with the smaller domain yields the best results. Also in this latter case, the reduction of the number of filter size in the last layer slightly improves the representation of Tx. When considering the large domain, the differences between the CNNs based on different activation functions increase, and CNN models with linear configuration outperform the ones with ReLu.
The findings of this work reinforces the potential of CNNs for climate downscaling in SSA, especially for its use to simulate multivariate impact indices.
How to cite: Glatstein, C. S., Balmaceda-Huarte, R., and Bettolli, M. L.: Downscaling a heat stress index in southern South America using deep-learning , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-13504, https://doi.org/10.5194/egusphere-egu25-13504, 2025.
The Global Ozone Monitoring Experiment 2 (GOME-2) and the TROPOspheric Monitoring Instrument (TROPOMI) are two significant satellite-based instruments dedicated to monitoring Earth’s atmosphere. GOME-2, part of the MetOp platform, has been operational since 2006, and was originally developed to monitor the ozone layer in the atmosphere. However, its onboard spectrometer can also detect pollutant gases, including NO2, which we will use as an initial example in this study.
GOME-2 spatial resolution is very coarse: a single data point is representative of an area of approximately 40 km x 80 km, which provides a broad view of atmospheric composition at global scale but limits its effectiveness in capturing fine-scale variations over cities and other human activity areas.
This study investigates whether TROPOMI high-resolution data can be utilized to downscale GOME-2 observations, potentially yielding insights into atmospheric changes dating back to 2006. We explore the implications of this process on spatial and radiometric accuracy and consider its broader significance for the future of satellite observations.
Given the abundance of available training data, we propose a novel approach involving deep learning. In particular, we used a combination of Residual Dense Blocks (RDBs) which state-of-the-art studies have shown to outperform similar Convolutional Neural Networks (CNNs) and Generative Neural Networks (GNNs) but still relies on the convolution operation, unlike transformers architectures (e.g., Vision Transformers ViTs). Then, to effectively train our model, we addressed challenges such as the resolution disparity between GOME-2 and TROPOMI (approximately a factor of 10), which requires working with a large pixel space, significantly increasing the memory needed for training. And the significant issue of missing data in atmospheric acquisition, e.g., due cloud cover.
Aside from the technical challenges of developing such model, the output validation plays a crucial role in ensuring the reliability and scientific utility of our results. We therefore evaluated our model performance on an independent dataset to verify the consistency of absolute reported NO2 values.
The approach involved training the model on one year of data (2023) over 10 selected locations and evaluate its performance using the ground-based Pandonia Global Network (PGN), a network of well-calibrated instruments designed to provide high-quality measurements of atmospheric trace gases at specific locations.
Results show an improvement not only limited to the reconstruction of fine details but also on the agreement of the absolute reported NO2 value between PGN data and the output from our model. We are currently working on expanding the dataset to further test the limits of our approach at global scale. Another active research area is the extension of the proposed approach to other common trace gases common between the two instruments. We hope to enhance the utility of this approach for broader applications in atmospheric science and to highlight the potential of leveraging deep learning downscaling for atmospheric data.
How to cite: Ratta, R., Mantovani, S., Houël, M., Beccarini, S., Schifano, S. F., and Fierli, F.: Leveraging Deep Learning for Downscaling GOME-2 Atmospheric Data Using TROPOMI Observations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4123, https://doi.org/10.5194/egusphere-egu25-4123, 2025.
The Italian Peninsula's climate is highly influenced by its complex topography and diverse regional weather systems, making high-resolution (HiRes) seasonal forecasting crucial for agriculture, water management, and energy sectors. Traditional seasonal prediction models, such as the CMCC Seasonal Prediction System (SPS3.5), provide valuable insights but lack the spatial resolution necessary to capture local-scale climatic details. Recent advances in Statistical Downscaling (SD) promise enhancing these coarse-resolution forecasts by generating more localised and accurate predictions. Thus, this study aims to provide a HiRes seasonal forecast for the Italian Peninsula by enhancing the SPS3.5 model through SD techniques tailored to the region's demand for finer-scale climate information.
The downscaling method involves a three-step process that utilises historical observational datasets and machine-learning techniques to refine SPS3.5 forecasts. The first regards the ground truth, composed of HiRes observational data from ERA5 reanalysis for 2m temperature (T2m), sea surface temperature, and 10m wind components, and from CHIRPS for precipitation. Then, SPS3.5 daily forecasts are spatially interpolated from 1º to 1/4° to match the observation data's grid. Finally, both data are combined through a machine-learning method based on the k-Nearest Neighbours (kNN) technique, which translates SPS3.5 into HiRes fields by matching forecasted conditions to observed patterns. The kNN algorithm utilises a set of k days of similar weather conditions (five predictors mentioned before) determined by the Euclidean distance to capture seasonally relevant weather analogues. Once the analogue days are defined, the kNN can forecast any meteorological field within the observational dataset. Finally, the SD method was accessed over the Italian Peninsula domain through cross-validation along the 24-year hindcast period available for SPS3.5 (1993-2016).
Preliminary results indicate that SD significantly enhances seasonal forecasts for the Italian Peninsula, achieving biases about 5-6 times smaller than the original SPS3.5 for all evaluated predictands. The main component of this improvement is the spatial accuracy promoted by downscaling, allowing the identification of domain characteristics unnoticed in SPS3.5. Even though the statistical indices show appreciable values for the domain as a whole when we evaluate smaller portions of this same domain, the original seasonal forecasts are still far from the desired. As expected, forecast bias increases with lead time also for kNN, with accuracy declining progressively from lead month 1 onward. For example, T2m bias increased from -0.14/-0.85°C in lead month 1 to -0.68/-1.41°C in month 6 (kNN/SPS3.5). This trend highlights the ongoing challenge of maintaining forecast skills over extended periods and the importance of adaptive correction strategies to extend lead-time reliability.
Integrating SD techniques with SPS3.5 outputs provides a promising solution for generating HiRes seasonal forecasts, offering valuable support for climate-sensitive applications by reducing forecast bias and enhancing spatial accuracy. This work demonstrates the potential of SD as an effective tool for bridging the gap between coarse seasonal forecasts and the localised weather information necessary for effective decision-making.
How to cite: Aragão, L. and Gualdi, S.: Statistical downscaling applied to the CMCC Seasonal Prediction System 3.5, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7043, https://doi.org/10.5194/egusphere-egu25-7043, 2025.
Weather generators (WGs) produce synthetic weather series, which are statistically similar to the real world weather series. The generators are used in assessing responses of weather-dependent processes on climate change (CC) or variability. Individual types of generators may differ in various parameters: (a) they may be parametric or non-parametric, (b) single-site or multi-site, (c) they differ in number of weather variables being generated and (d) the time step. Choice of these parameters depends on the purpose of their use. For example, in agrometeorology, single site (4-6)-variate daily generators are used to assess CC impacts on crop yields, which may include assessment of the sensitivity of the yields to changes in various climate characteristics.
In this contribution, we present our approach to using the generator in crop yield forecasting. Specifically, the crop yields are simulated by crop models, while the input weather series consisting of observational data till day D0 (when the forecast is made) are seamlessly followed by the synthetic series produced by the parametric single-site daily weather generator M&Rfi. Two approaches were implemented in M&Rfi to produce such series: (1) In the first, “operational” mode, the synthetic series are “forced” to exactly fit the available weather forecast, which accounts for the possible uncertainties and spans for the rest of the growing season; to make a probabilistic crop yield forecast, large number of possible weather series realisations is produced. (2) In the second, “research” mode, we do not assume to have a specific weather forecast, but we rather assume to have a knowledge on the accuracy of the available weather forecasts, which may be expressed as a function of the weather forecast error on the lead time. Having this function, we may produce a large ensemble of possible weather forecasts and corresponding ensemble of synthetic weather series.
Our methodology of producing synthetic weather series, which fit available weather forecasts, may be applied also for other weather dependent processes, for example in hydrological applications.
Acknowledgements: The experiment was made within the frame of projects PERUN (supported by TACR, no. SS0203004000) and YiPeeO (supported by ESA, no. 4000141154/23/I-EF).
How to cite: Dubrovský, M., Trnka, M., Bartošová, L., and Štěpánek, P.: Adjusting the Weather Generator for Use in Operational Forecasting Weather-Dependent Processes, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8914, https://doi.org/10.5194/egusphere-egu25-8914, 2025.
Regional climate projections are essential for guiding local governments in developing effective mitigation strategies. A common technique for downscaling General Circulation Model (GCM) outputs is dynamical downscaling, but its high computational demands have motivated the search for alternative approaches, including statistical downscaling. This study presents a two-phase statistical downscaling framework to improve the spatial resolution and accuracy of precipitation and temperature projections. In the first phase, a Convolutional Neural Network (CNN), trained to learn spatial patterns from ERA5 reanalysis data, is employed to refine the coarse grid of GCMs. In the second phase, bias correction is performed using a quantile delta mapping technique, with ERA5 still serving as the reference dataset. The resulting downscaling framework is applied to outputs from five GCMs participating in the Coupled Model Intercomparison Project Phase 6 (CMIP6) under two Shared Socioeconomic Pathways (SSPs): SSP1-2.6 and SSP3-7.0. This work is part of the OurMED PRIMA project, which focuses on the Mediterranean region, a recognized climate change hotspot. Results indicate substantial improvements in the accuracy of temperature and precipitation projections compared to other downscaling methods. The proposed approach effectively captures fine-scale spatial variability, a crucial aspect for regional climate studies in complex regions like the Mediterranean region. The downscaled climate data are used to assess climate extremes by computing the indices defined by the Expert Team on Climate Change Detection and Indices (ETCCDI). These indices can offer valuable insights into evolving climate trends and extremes throughout the 21st century. The proposed methodology demonstrates significant potential for broader applications in regions requiring high-resolution climate data to support adaptation strategies and policy development.
This work was supported by OurMED PRIMA Program project funded by the European Union’s Horizon 2020 research and innovation under grant agreement No. 2222. Valerio Todaro acknowledges financial support from the PNRR MUR project ECS_00000033_ECOSISTER.
How to cite: D'Oria, M., Todaro, V., Secci, D., and Tanda, M. G.: Statistical downscaling of climate models for the Mediterranean region combining convolutional neural network and quantile delta mapping, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9167, https://doi.org/10.5194/egusphere-egu25-9167, 2025.
This study presents a detailed assessment of very high-resolution reanalysis data covering the entire Italian territory and the broader Alpine domain for the three-decade period 1990-2020. The dataset was generated using a dynamical downscaling of ERA5 reanalysis with the convection-permitting model MOLOCH, implemented at a fine grid spacing of 1.8 km.
Validation against high-resolution observational datasets (GRIPHO, ARCIS, and the ISAC-CNR precipitation and temperature dataset) and comparisons with similar downscaled reanalysis products (ERA5-LAND, CERRA, MERIDA-HRES, and SPHERA) confirm the dataset’s reliability in reproducing key meteorological variables, such as temperature and precipitation. Importantly, the dataset leads in capturing higher-order statistics, including intensity and extremes.
The dataset’s versatility is illustrated through multi-disciplinary applications. In hydrology, it enables high-resolution drought characterization; in meteorology, it supports the analysis of extreme weather events and orographic effects. In climate research, it provides valuable insights into long-term trends and variability.
This work underscores the importance of very high-resolution datasets in advancing our understanding of the complex interactions between natural processes and human activities, especially in regions with challenging topography like the Alps. It establishes a strong foundation for future research and practical applications, including disaster risk management, water resource planning, and climate adaptation strategies.
How to cite: Tariq, M., Cavalleri, F., Davolio, S., Brunetti, M., Camici, S., Mastrangelo, D., and Stocchi, P.: Three Decades of high-Resolution ERA5 Downscaling over the Italian domain: Validation and Applications in Hydrology, Meteorology, and Climate Analysis, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9367, https://doi.org/10.5194/egusphere-egu25-9367, 2025.
The Alpine region faces heightened risks from climate change due to its complex terrain and ecosystems, highlighting the significant global challenge posed by a warming climate. The region is particularly susceptible to the effects of global warming, which not only intensifies weather extremes but also significantly impacts hydrological processes. These changes increase the frequency and severity of extreme events like droughts and floods, further heightening the region's vulnerability. Accurate local climate predictions are essential for effectively managing these risks, as they provide the spatial and temporal precision necessary for hydrological simulations. Such high-resolution data enable detailed modelling of water availability, runoff patterns, and flood risks, facilitating improved planning and adaptation strategies. However, existing global datasets often lack the resolution needed for these assessments. To address this gap, this research aims to generate high-resolution seasonal climate forecasts specifically designed for the Alpine region, providing an essential tool for understanding climate variability, managing hazards, and supporting hydrological analyses. The study proposes a novel two-stage downscaling approach within the perfect prognosis framework to enhance the spatial resolution of ECMWF (European Centre for Medium Range Weather Forecasts) SEAS5 (Seasonal Forecast System 5) seasonal forecasts from native 0.25°x0.25° to 1 km for the Alpine region. Key variables include daily temperature, precipitation, and downward surface solar radiation. In the first stage, pixel-by-pixel downscaling is performed though LGBM (Light Gradient Boosting Machine) regression applied to ERA5 reanalysis predictor fields matched against CHELSA-W5E5 (v1.1) fields, conservatively interpolated to 6-km resolution. Predictors are selected through feature importance analysis via cluster-based regression and is optimized for the 2005–2016 training period. The trained model is then applied to the 51 ensemble members of SEAS5 predictors, generating target variables at a 6 km resolution. In the second stage, the 6-km downscaled outputs, along with additional static predictors such as elevation, aspect, and cyclically encoded day of the year, are passed to a sliding-window Enhanced Super-Resolution Generative Adversarial Network (ESRGAN). This image super-resolution technique trained and optimized using CHELSA-W5E5 at its native 1-km resolution, further refines the forecasts to produce high-resolution seasonal predictions with 51 ensemble members at 1 km resolution. The two-stage scheme was found to improve the downscaling performance with respect to the application of one-step method. The contribution will present the overall methodology and the results of the model evaluation. The outcomes of this study are expected to play a key role as critical inputs for a drought prediction module within the framework of the EU-funded interTwin project. This research has been funded by the European Union through the interTwin project (101058386).
How to cite: Dhinakaran, S., Crespi, A., Castelli, M., Ferrario, I., and Jacob, A.: A Two-Stage Downscaling Approach using Machine Learning and image super-resolution techniques for high-resolution seasonal climate forecasts in the Alpine region , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11322, https://doi.org/10.5194/egusphere-egu25-11322, 2025.
Unprecedented climatic events have been occurring more frequently, highlighting the role of anthropogenic climate change and the need for accurate regional climate projections for adaptation planning. Such projections require high spatial resolution, typically achieved by downscaling global climate model outputs. However, evaluating climate model outputs remains challenging, as they represent statistical features of climate change and do not evolve consistently with observations.
In this study, we conduct an empirical statistical downscaling of a large number of historical (1951-2014) CMIP6 global climate projections using different configurations of the statistically downscaling method EPISODES. The domain covers Hydrological Germany, including Germany and its main rivers' basins.
We provide a comprehensive assessment of the performance of each downscaled projection. We evaluate the statistical characteristics of each model run against observational data from four key perspectives. Specifically, we assess the performance of each projection for six key climate variables based on annual and seasonal climate means, as well as internal variability across various timescales. To estimate the ability of each run to capture the persistence of weather regimes, we also compare the lagged autocorrelation function across the entire domain for daily mean variables. Additionally, we divide our domain into nine zones and compute the histograms of daily mean variables. We use various widely adopted statistical metrics and have developed new indices. This approach enables a comprehensive evaluation of the performance of each realization from multiple perspectives, facilitating the identification of the optimal configuration of EPISODES, which can serve as a key tool for climate model evaluation.
How to cite: Bordbar, M. H., Lorenz, P., Kreienkamp, F., and Schellander-Gorga, T.: A Comprehensive Approach for Evaluating Downscaled Climate Model Projections from Multiple Perspectives: A Case Study of Hydrological Germany, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11589, https://doi.org/10.5194/egusphere-egu25-11589, 2025.
Global gridded land use projection data is essential for the investigation of various research topics in global environmental change. Such information is commonly provided by various integrated assessment models (IAMs) at a relatively coarse resolution. For example, the Land Use Harmonization 2 (LUH2) provided future global land use data at 0.25-degree for CMIP 6. However, the demand for higher resolution land use projection data has been increasing in recent years for more granular analysis of various topics. The Asia-pacific Integrated Model (AIM), which is a widely known IAMs for climate policy study, could so far provide global gridded land use data at 0.5-degree resolution under the SSP-RCP scenario framework. In this study, I constructed a downscaling framework for the AIM land use model system, that combines an empirical land use change model and a cross-entropy minimization method and aimed to downscale land use projection from half-degree to 5 arcminutes or even higher resolution. The empirical land use change model is estimated by multinominal logit regression method with historical data from 1995 to 2015, which allows the land use change driven by various biophysical and socio-economic factors and provides prior land use distribution information for the cross-entropy minimization process. Validation for the period of 2015 to 2020 showed the effectiveness of the downscaling model. This newly developed downscaling model could provide high-resolution gridded land use projection information for global environmental change research community.
How to cite: Wu, W.: The development of a high-resolution global land use projection downscaling model, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15262, https://doi.org/10.5194/egusphere-egu25-15262, 2025.
Due to the excellent temporal continuity, reanalysis datasets are often used as input data for downscaling models. However, because of the relatively coarse spatial resolution, reanalysis datasets often exhibit significant value differences between adjacent pixels, making it challenging to accurately capture the distribution of meteorological parameters in heterogeneous urban areas. Although many downscaling studies have utilized reanalysis data, none have explored how to preprocess these datasets to achieve smoother patterns in the distribution of meteorological parameters at the urban level, making them closer to real distribution patterns. To address this limitation, this study proposes a novel iterative Gaussian filtering method. This method applies iterative Gaussian filtering while keeping the mean values unchanged within the coarse-resolution pixels to generate fine-resolution data with smoother distribution patterns. In this study, the 1-km land surface temperatures obtained from MODIS and its reprojected 0.1˚ resolution data are assumed to represent the true fine-resolution values and coarse-resolution values, respectively, to validate the effectiveness of the proposed method. The results indicate that, compared to the coarse-resolution data, the fine-resolution data processed through iterative Gaussian filtering achieves higher accuracy, with RMSE and MAE improvements of 11.06% and 11.89%, respectively. The distribution patterns of the fine-resolution data are also closer to real distribution patterns than those of the coarse-resolution data. These findings suggest that our proposed method could serve as a valuable tool for enhancing the accuracy of downscaling models in future studies.
How to cite: Wen, Z., Zhuo, L., Yu, J., and Han, D.: How to Make Downscaling Model Inputs Closer to Real Distribution Patterns?, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15411, https://doi.org/10.5194/egusphere-egu25-15411, 2025.
Global warming is increasing the frequency and intensity of extreme temperature events, posing significant risks to human health during major outdoor events such as the Summer Olympics. Providing decision-makers with robust, high-resolution extreme temperature forecasts well in advance is crucial to anticipate risks on the health of both athletes and spectators. Global subseasonal forecasts can play a key role in addressing this challenge because they offer data with relatively high temporal resolution (i.e., weekly) several weeks ahead. However, the coarse spatial resolution of these forecasts limits their utility for the types of localized decision-making required for major events, necessitating the use of downscaling methods to improve resolution.
Although numerous downscaling approaches exist, their ability to skillfully downscale subseasonal data has not been systematically evaluated. To address this gap, this study assesses the performance of 27 statistical downscaling methods – including bias correction, linear regression, logistic regression, and analogs – in enhancing the spatial resolution of subseasonal temperature hindcasts. We use Climate Prediction System version 2 (CFSv2) data at 100 km resolution as the raw hindcast product and downscale these hindcasts to a 5 km resolution. The process is conducted separately for temperature hindcasts from models initiated 1, 2, 3, and 4 weeks prior to the three target weeks of the Paris 2024 Olympics (starting from 22 July, 29 July and 5 August). In addition to using CFSv2 temperature outputs as predictors, we explore the added value of incorporating atmospheric patterns into the downscaling process. Models are constructed using both daily and weekly data, enabling a comparative analysis of performance across two temporal scales.
The results show that downscaling methods can successfully transfer the predictive skill of CFSv2 to the 5 km resolution. However, the choice of downscaling method is crucial to the performance, as some methods degrade the predictive skill of CFSv2, while others enhance it. Notably, methods that incorporate atmospheric patterns show promise in improving forecasts with longer lead times. Additionally, daily data models using analogs outperform their weekly counterparts, while regression-based methods perform better with weekly data.
In summary, this study demonstrates the potential of statistical downscaling to enhance coarse-resolution subseasonal temperature forecasts. However, it also highlights the significant variability in forecast skill depending on the choice of predictors and methods, which can either improve or degrade performance.
How to cite: Duzenli, E., Ramon, J., Torralba, V., Pickard, S., Bojovic, D., Trascasa-Castro, P., and Muñoz, Á. G.: Assessing the added value of statistical downscaling to the predictive skill of global subseasonal temperature forecasts during the Paris 2024 Olympics, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-18397, https://doi.org/10.5194/egusphere-egu25-18397, 2025.
This study evaluates the application of a deep learning approach employing a multi-head attention mechanism within a deep neural network (DNN) framework to enhance bias correction and downscaling of the fifth-generation European Centre for Medium-Range Weather Forecasts (ECMWF) reanalysis rainfall datasets. The proposed Encoder-Decoder with multi-head Attention (EDA) model leverages gridded 5-km daily rainfall observations and auxiliary inputs, such as surface wind data and high-resolution topography, to generate local-scale daily rainfall estimates across Taiwan—a mountainous subtropical island with complex terrain.
The model's performance is assessed using mean rainfall patterns, rainfall statistics, extreme climate indices, and interannual variations during Taiwan's rainy seasons. Results demonstrate that the EDA model effectively corrects biases in low-intensity rainfall and resolves inaccuracies in orographic rainfall placement present in reanalysis datasets, outperforming conventional quantile-mapping methods. Additionally, the integration of auxiliary surface wind information significantly improves the model's downscaling accuracy across various metrics.
This study highlights the potential of deep learning architectures, particularly those incorporating attention mechanisms and auxiliary data, for statistical bias correction and downscaling in regions characterized by intricate interactions between synoptic and local circulations modulated by topography.
How to cite: Wang, Y.-C., Chiang, C.-H., Tseng, W.-L., and Wang, K.-C.: Enhancing Bias Correction and Downscaling of Rainfall Pattern Over Taiwan with a Deep Learning Neural Network Over Complex Terrain, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19866, https://doi.org/10.5194/egusphere-egu25-19866, 2025.
