Despite the ongoing advancements in Earth system simulation, the results from Global Climate Models (GCMs) are still not refined enough to be directly applied to numerous climate impact issues. There are many techniques available to downscale GCM outputs to finer resolutions, from basic statistical adjustments to more complex methods like dynamical downscaling and machine learning. However, these methods often yield different results, making it difficult to assess their relative reliability, particularly when comparing statistical versus dynamical downscaling methods.

We consider downscaled results to be credible when the phenomena and processes producing it are consistent; for instance, if it’s raining, the necessary conditions for rain (such as lift and atmospheric moisture) should be present. To assess various downscaling techniques, and demonstrate this technique, we examine the occurrence of rainfall at a location the Southern Great Plains, specifically near the DOE ARRM site in Oklahoma during May, the rainiest month. In this scenario, we are looking for an atmospheric setup that produces uplift at this location and corresponds with the northward movement of moisture from the Gulf of Mexico.

By comparing the composite synoptic-scale meteorological conditions on days with and without rain from the GCM being downscaled or from the downscaling method, as appropriate, we can verify if the outcomes of downscaling GCM precipitation align with the processes that drive them. This method offers a process-based added-value analysis strategy for all kinds of downscaling techniques, which extends beyond basic measures of statistical resemblance.

We’ve used two regional climate models (RegCM4 & WRF), a machine learning technique (U-Net CNN), and four statistical methods of different complexities to downscale precipitation from three distinct GCMs. By using this method to compare them with each other and the raw GCM results, we’ve discovered that all downscaling methods can yield plausible outcomes when the GCM performs well, as they inherit its credibility. However, when the GCM’s performance is subpar, only dynamical methods can rectify regional circulation errors, unlike the other methods. Interestingly, we also found that simpler statistical methods outperform more complex non-dynamical methods when dealing with poor GCM inputs.

How to cite: Bukovsky, M., McGinnis, S., McCrary, R., and Mearns, L.: A Process-Informed Determination of Credibility Across Different Downscaling Methods, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-13235, https://doi.org/10.5194/egusphere-egu24-13235, 2024.

Statistical Downscaling Methods (SDMs) play a pivotal role in climate change assessments at local and regional scales, as they can efficiently reproduce historical climate observations, overcoming the limitation of Global Climate Models (GCMs) in capturing fine-scale climatic features. However, the evaluation of GCMs and SDMs often focuses on historical climatology, neglecting extreme events representation and climate change signal preservation. In response, this paper proposes a methodological guideline for GCMs and SDMs selection, incorporating three key criteria: representation of historical climatology (Past Performance Index - PPI), representation of extreme wet climate indices (Climate Integrated Impact Index - CI³), and preservation of climate signal change (Climate Signal Performance Criteria - SCPI). Satisfactory GCM and SDM performance during the historical period is defined by meeting conditions such as PPI ≥ 0.5 for each climatic variable (precipitation, minimum and maximum temperature) and CI³ ≥ 0.4. For future projections, SCPI guides the selection process, considering short (2015 – 2040), medium (2041 – 2070), and long-term (2071 – 2100) projections across different Shared Socioeconomic Pathways (SSPs) (see step d) in Figure 1).

The study evaluates 18 GCMs from Sixth Model Intercomparison Phase (CMIP6), interpolated to the gridded meteorological product CR2METv2.0 (0.05° x 0.05°) for the northern region of Chile (17ºS – 32º). Ten SDMs are applied to short, medium, and long-term periods under SSP2-4.5 and SSP5-8.5 scenarios. Results indicate that no single SDM corrects all criteria for all GCMs. Climate projection groups are established based on the number of criteria met, distinguishing models that satisfy two or three criteria. The historical evaluation shows that interannual variability is the most influential in the PPI results, both for precipitation and temperatures (min and max). Better historical performance is also observed for multivariate methods family over quantile mapping family or hybrid methods family (combination of analogs, resampling, climate fingerprinting and quantile mapping). In the case of CI³, all SDMs for all the GCMs show a similar bias for maximum precipitation magnitude and their mean temperature, meanwhile the consecutive wet days, days with precipitation over 50 mm and snow process indices present a bias of less than 10%. For this metric, no SDM family has a better performance over another SDM family. Finally, the preserving of climate signal change (for each SSP scenario and projection period) is not observed with the hybrid method. For quantile methods, we observed a tendency of modification of the signal climate change, and the multivariate methods has the best performance in these criteria. This proposed methodology facilitates the selection of GCM subsets based on study objectives (climatology, extreme events, or climate change signals). Future work should focus on advancing additional statistical downscaling methods capable of representing diverse criteria, including natural variability and climate change signals.

Figure 1. Methodological scheme for the selection of suitable GCMs and SDMs.

How to cite: Jerez, C., Lagos-Zuñiga, M., and Montserrat, S.: Evaluating CMIP6 models under different statistical downscaling methods for climate assessments in the north of Chile, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12446, https://doi.org/10.5194/egusphere-egu24-12446, 2024.

    In this study, we aim to assess the impacts of GCM bias correction on dynamical downscaling simulation over the Asia-western North Pacific region. Three simulations were conducted with a 25-km grid spacing for the period 1980–2014. The first simulation (WRF_ERA5) was driven by the European Centre for Medium-Range Weather Forecasts Reanalysis 5 (ERA5) dataset and served as the validation dataset. The original GCM dataset (MPI-ESM1-2-HR model) was used to drive the second simulation (WRF_GCM), while the third simulation (WRF_GCMbc) was driven by the bias-corrected GCM dataset. The bias-corrected GCM data has an ERA5-based mean and interannual variance but the long-term trends are derived from the ensemble mean of 18 CMIP6 models. Results demonstrate that the WRF_GCMbc significantly reduced the root-mean-square errors (RMSEs) of the climatological mean of downscaled variables, including temperature, precipitation, snow, wind, relative humidity, and planetary boundary layer height by 50%–90% compared to the WRF_GCM. Similarly, the RMSEs of interannual-to-interdecadal variances of downscaled variables were reduced by 30%–60%. Furthermore, the WRF_GCMbc better captured the annual cycle of the monsoon circulation and intraseasonal and day-to-day variabilities. The leading empirical orthogonal function (EOF) shows a monopole precipitation mode in the WRF_GCM. In contrast, the WRF_GCMbc successfully reproduced the observed tri-pole mode of summer precipitation over eastern China. This improvement could be attributed to a better-simulated location of the western North Pacific subtropical high in the WRF_GCMbc after GCM bias correction.

How to cite: Xu, Z., Han, Y., Zhang, M.-Z., Tam, C.-Y., Yang, Z.-L., EL Kenawy, A., and Fu, C.: Dynamical Downscaling Simulation of Asian Climate with a Bias-Corrected CMIP6 Dataset: Evaluation , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-4810, https://doi.org/10.5194/egusphere-egu24-4810, 2024.

We present an interactive climate atlas providing visualisations of future regional climate projections of temperature and precipitation in northern Europe from multiple sources. It is based on results of both empirical-statistical and dynamical downscaling of multi-model ensembles from CMIP5 and CMIP6 including several emission scenarios. Displayed alongside each other, the projected climate change estimated from different model ensembles can be compared and contrasted. The comparison can be useful to evaluate the robustness of the climate change information and the influence of methodological choices such as the downscaling method and the selection of global climate models, and to explore how the level of greenhouse gas emissions may affect the future climate. The application is developed by researchers at the Norwegian Meteorological Institute and is freely available at the website futureclimate.met.no/dse4KSS.

How to cite: Parding, K., Dobler, A., Benestad, R., Lutz, J., Mezghani, A., and Dybdal, A. V.: An interactive climate atlas for northern Europe, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-17936, https://doi.org/10.5194/egusphere-egu24-17936, 2024.

Projections generated by global climate models (GCMs) are increasingly utilized to inform climate adaptation policies. It is known that climate models simplify the real climate system, leading to biases between simulated and observed climates. The spatial and temporal resolution of GCMs is ever increasing to provide a better representation of the Earth system and in turn, also provide higher quality information for users. To effectively handle the substantial climate data produced by these models, which can reach Terabytes to Petabytes, the Destination Earth (DestinE) initiative is exploring data streaming—a new approach that enables user applications to run Earth system models in an end-to-end workflow directly downstream of the climate simulations, eliminating the need to store entire time-series of variables to disk.

Traditional methods for quantile or percentile calculation typically involve sorting the data and directly computing the specific value corresponding to the desired quantile. These methods can be computationally intensive, especially for large datasets, as it necessitates storing and processing the entire dataset. While traditional bias-adjustment (BA) algorithms rely on data being fully available, a further challenge lies in developing bias-adjustment procedures capable of accommodating streamed data on-the-fly. In the DestinE Climate Digital Twin (CDT), we extend the quantile-mapping technique used in the ISI-MIP project (isimip.org) because it is a well-established method and preserves the trend of the original data. The technique involves aligning the CDFs of the model data with those of the observed data by adjusting the model's cumulative distribution to match that of the observed data. The enhancements of the BA method in DestinE-CDT is making use of the T-Digest algorithm, a sophisticated strategy that dynamically clusters data points into small groups, which is used to generate a summarized representation of the data distribution from streamed data and accurately calculate percentiles. This clustering technique offers an accurate estimate of percentiles while efficiently managing large and unbounded data streams where new data points are continuously added.

We apply the developed quantile-mapping BA for different variables on a global scale and compare it with the parametric distribution functions used in quantile-mapping BA from the ISI-MIP project.

How to cite: Sharifi, E., Grayson, K., Müller, S., and Thober, S.: A Novel Bias-Adjustment Methodology for Streaming Global Climate Models, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-9083, https://doi.org/10.5194/egusphere-egu24-9083, 2024.

The NASA Earth Exchange Global Daily Downscaled Projections CMIP6 archive (NEX-GDDP-CMIP6) contains daily climate projections of nine variables derived from thirty-five CMIP6 GCMs and four SSP scenarios (SSP2-4.5, SSP5-8.5, SSP1-2.6 and SSP3-7.0) for the period 2015-2100. Each of these climate projections was downscaled to a spatial resolution of 0.25 degrees x 0.25 degrees using the daily version of the BCSD statistical downscaling method. The purpose of this archive is to provide a set of global, high-resolution, bias-corrected climate change projections that can be used to evaluate climate change impacts on processes that are sensitive to finer-scale climate gradients and the effects of local topography on climate conditions. In this session, we will provide an overview of the methodology, as well as the details of its execution on the NASA Advanced Supercomputing (NAS) facility. In addition, we will discuss the various considerations, assumptions, and limitations of the downscaled data. Lastly, we will illustrate the various modes of accessing the archive, including examples using the NASA Regional Climate Model Evaluation System (RCMES) and cloud computing resources.

How to cite: Thrasher, B., Wang, W., Michaelis, A., Brosnan, I., and Khajehei, S.: NASA Earth Exchange Global Daily Downscaled Projections (NEX-GDDP-CMIP6), EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19303, https://doi.org/10.5194/egusphere-egu24-19303, 2024.

Statistical downscaling is a technique that allows to obtain high-resolution climate information from the coarse-resolution Global Climate Model (GCM) outputs through the long-term relationship between the GCM output and a reference dataset such as in-situ observations. The key benefit of employing statistical downscaling (SD) methods over the dynamical approaches is their significantly less computational costs. The cost-effectiveness of these methods enables the processing of large hindcasts, including multi-model systems with numerous ensemble members, which is highly relevant for the users. Thus, a comprehensive tool that allows users to apply state-of-the-art statistical downscaling methods on climate variables is crucial. CSDownscale is a new generation R package that has been  developed to statistically downscale subseasonal to seasonal to decadal climate predictions in the context of climate services, including its use in operational applications. The tool produces a downscaled field or time series using several bias correction, regression (i.e., linear and logistic) and analogs methods. Additionally, the package contains various interpolation methods such as nearest neighbor or bilinear approaches, which are used for regridding purposes. Users can easily combine these with bias correction and regression methods to perform downscaling. When applying these combined operations, the GCM data is initially interpolated to the resolution of the reference dataset, then the selected bias correction or regression method is implemented on the regridded data. However, the package also incorporates a method that infers the high-resolution values using a multi-linear regression with the four closest coarse-scale grid points, which skips the step of interpolation. Furthermore, the CSDownscale package includes an analogs based method, which looks for fields with similar conditions to the one being predicted and returns the high-resolution outcome of past conditions that are most similar, a certain number of similar fields or a combination of them. Finally, the CSDownscale package allows for the GCM data to be downscaled to either a reference grid space or a specific point location. All the methods are conceived to be done in cross-validation, that is, by excluding data from the specific time step being post-processed to avoid overfitting and, consequently, the overestimation of the actual prediction skill.

How to cite: Duzenli, E., Ramon Gamon, J., Llabres, A., Torralba, V., Palma Garcia, L., Moreno Montes, S., Delgado-Torres, C., Perez-Zanon, N., Corvillo Guerra, J., and Marcos, R.: Climate Services Downscale (CSDownscale): A statistical downscaling tool for (sub)seasonal to decadal climate predictions, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19420, https://doi.org/10.5194/egusphere-egu24-19420, 2024.

Stochastic weather generators are one of the most frequently used methodologies for producing input weather series for various process-based models (especially agricultural crop growth models and hydrological rainfall-runoff models) used e.g. in assessing impacts of climate change/variability on weather-dependent processes.

SPAGETTA (Dubrovsky et al., 2020, Theor. Appl. Climatol.) is a parametric multi-variable spatial weather generator run commonly (but not only) with daily time step. It is based on applying the spatialisation approach developed by Wilks (1998, J. Hydrol.) to our older single-site weather generator M&Rfi. Similarly to M&Rfi, SPAGETTA is designed for agricultural and hydrological modeling. Until recently, the stress was put on finetuning and validating the generator. Now, when the generator performs reasonably well, it is being used in various experiments.

In the first part of the presentation, the main results of the validation tests will be shown, focusing on the ability of the generator to reproduce spatial-temporal variability of multi-site temperature and precipitation series. Concerning the temporal variability, both high-frequency (interdiurnal) and low-frequency (intermonthly and interannual) variability was considered. The performance of the generator was compared with the performance of 19 RCMs taken from the CORDEX database.

In the second-part, to demonstrate different applications of the generators, we show results obtained in four experiments: (1) Assessment of separate effects of changes in the WG parameters, which represent the means, variability and lag-0 & lag-1 spatial correlations of temperature and precipitation, on a set of temperature and precipitation indices. The generator parameters were calibrated using the observational E-OBS data from 8 European regions and then modified with the changes (2070-99 vs. 1971-2000) derived from 19 RCM climate simulations (this experiment was already presented in EGU 2023). (2) To show the generator's performance in hydrological modeling, we applied the rainfall–runoff model to the watershed of Dyje river. The model outputs obtained using the synthetic weather series were compared with outputs obtained with the observed weather series (we call this type of experiment “indirect validation of WG”. (3) The generator was used to develop a new test for examining the collective significance of local trends at multiple sites (Huth and Dubrovsky, 2021, J. Clim.). This was made by applying large ensembles of realizations of synthetic multi-site weather series for user-defined lag-0 and lag-1 spatial correlation matrices, (4) The generator was used to assess the statistical significance of climate change scenarios produced by Regional Climate Models. The significance of the RCM-based changes (future vs. baseline) in individual WG parameters is based on comparing their values with the spread of the changes of these parameters based on ensembles of synthetic weather series, i.e. the pairs of synthetic series representing future and present climate; the generator was calibrated by RCM simulations for the corresponding time slices.

How to cite: Dubrovsky, M., Huth, R., Stepanek, P., Lhotka, O., Miksovsky, J., Meitner, J., Balek, J., Vizina, A., and Trnka, M.: Stochastic Spatial Weather Generator SPAGETTA: Development and Applications, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19461, https://doi.org/10.5194/egusphere-egu24-19461, 2024.

Rainfall data are needed as input to drive hydrological and soil erosion models. Daily rainfall data are commonly used and widely accessible, whether sourced from meteorological observations or simulated by Global Climate Models (GCMs). However, daily data cannot capture intensity variations during a storm event, and may not be sufficient to capture the changes during extreme weather events under climate change scenarios. Weather generators (WGs) are statistical models that can generate random sequences of meteorological variables that exhibit statistical characteristics that are similar to observations. However, the low accuracy of generated sub-daily rainfall intensities motivated this study to stochastically disaggregate daily precipitation total at hourly intervals so that observed or GCM generated daily rainfall can be downscaled into hourly scale stochastically. To achieve this, we developed a model, HRGEN, based on long-term hourly precipitation data from 1971 to 2020 from 2405 meteorological stations across mainland China. The major improvement of this model over CLIGEN includes: (1) HRGEN significantly enhances the simulation accuracy of maximum peak intensities on an hourly basis (H_max). The average H_max over 2405 stations of hourly observations and HRGEN-generated are 4.0 mm h^-1 and 4.2 mm h^-1, respectively, while that generated by CLImate GENerator (CLIGEN) is 6.5 mm h^-1. The mean absolute relative error (MARE) over 2405 stations is 8.2%. This improvement is critical for accurately estimating daily EI₃₀ values, a key index in soil erosion models and soil loss prediction; (2) HRGEN preserves the relationship between total daily precipitation and storm duration and peak intensity; (3) The model has only six parameters, markedly simplifying the calibration and simulation processes. The HRGEN-simulated hourly rainfall data can be used to estimate rainfall erosivity for erosion prediction. The R-factor estimated using HRGEN-generated hourly data agrees well with the observed R-factor values, with a high Nash-Sutcliffe efficiency coefficient (NSE) of 0.92. The average R-factor estimated from hourly observations and HRGEN-generated hourly observations over 2405 stations are 3699.2 and 3720.7 MJ mm ha^-1 h^-1 a^-1, respectively. In comparison, R-factor estimated by CLIGEN-generated rainfall is 9100.7 MJ mm ha^-1 h^-1 a^-1. This study highlights HRGEN’s potential as a robust tool for stochastic generation of sub-daily rainfall as input to hydrologic and soil erosion models.

How to cite: Wang, W., Yin, S., and Yu, B.: HRGEN: A stochastic generator of hourly rainfall, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-2693, https://doi.org/10.5194/egusphere-egu24-2693, 2024.

Downscaling techniques are one of the most prominent applications of Deep Learning (DL) in Earth System Modeling. A robust DL downscaling model can generate high-resolution fields from coarse-scale numerical model simulations, saving the timely and resourceful applications of regional/local models. Moreover, specific DL models can generate uncertainty information and provide ensemble-like pool scenarios, hardly achievable using traditional numerical simulations due to their high computational requirements. In this work, we present the application of deep generative models, namely a Generative Adversarial Network (GAN) and a Latent Diffusion model (LDCast, Leinonen et al., 2023), to perform the downscaling of ERA5 (Hersbach et al., 2018) data over Italy up to a resolution of 2km. The target high-resolution data used for training consists in the Italian high-resolution dynamical reanalyses obtained with COSMO-CLM (Raffa et al., 2021). The goal of the study is to show that recent advancements in generative modeling can learn to provide comparable results with numerical dynamical downscaling models, such as the COSMO-CLM model, given the same input data (i.e., ERA5 data), preserving the realism of fine-scale features and flow characteristics. The training and testing database is composed of hourly data from 2000 to 2020 (~184000 timestamps), and the target variables of the study are 2-m temperature and horizontal wind components. A selection of predictand variables from ERA5 is used as input to the DL models (e.g., 850hPa temperature, specific humidity, and wind). The generative models are compared with reference baselines, both DL-based (e.g., UNET) and statistical methods. Preliminary results are presented, highlighting the improvements introduced with this architecture with respect to the baselines. The results are evaluated by different quantitative verification scores: RMSE, predicted spectra, frequency distributions, and spatial distribution of errors. 

How to cite: Tomasi, E., Franch, G., and Cristoforetti, M.: Can AI be enabled to dynamical downscaling? Training Deep Generative Models to downscale ERA5 to high-resolution COSMO-CLM dynamical reanalyses over Italy , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10091, https://doi.org/10.5194/egusphere-egu24-10091, 2024.

To complement computationally expensive regional climate model (RCM) simulations, machine learning methods can predict the high-resolution RCM data from low-resolution global climate model (GCM) input. Instead of merely targeting the conditional mean of the RCM field given the GCM data, more recent works are based on generative adversarial networks or diffusion models and aim to learn the full conditional distribution. In this spirit, we present a novel generative model that relies on statistical tools from forecast evaluation. The model can sample several plausible RCM realisations and enables assessing their variability. To achieve this goal, we use a simple neural network architecture that predicts Fourier coefficients of the high-resolution fields for multiple variables jointly (temperature, precipitation, solar radiation and wind). The loss function of our model is a proper scoring rule that measures the discrepancy between the model’s predictive distribution and the RCM’s true distribution. The score is minimised if both distributions agree. Our generative model is trained on multiple GCM-RCM combinations from the Euro-Cordex project. Furthermore, we show how the framework can be augmented to perform a bias-correction task: With a modified loss function, it is possible to generate data from the observational distribution, for example resembling gridded E-OBS data. To summarise, our work presents a machine learning method that allows us to generate multivariate high-resolution climate data, and can be extended flexibly to include further variables or downscale and bias-correct future projections.

How to cite: Schillinger, M., Shen, X., Samarin, M., and Meinshausen, N.: Machine Learning for Multivariate Downscaling: A Generative Model Inspired by Forecast Evaluation, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-8464, https://doi.org/10.5194/egusphere-egu24-8464, 2024.

Land Surface Temperature (LST) is crucial in many areas; but seamless LST data are difficult to obtain due to limitations in thermal infrared sensor technologies. Numerical modeling, which is based on physics-driven process, can simulate continuous spatial and temporal data. Simultaneously, machine learning, a typical data-driven approach, has been effective in remotely-sensed data reconstruction. In this study, we designed a fusion framework that combines the strengths of numerical modeling and machine learning. The framework includes the following steps: 1) Optimization of the numerical model: We use the urbanized High-Resolution Land Data Assimilation System (u-HRLDAS) model. Various spatio-temporal data sources are used to refine and optimize the model's simulations. 2) Database creation for LST reconstruction: This database incorporates forcing variables like 2-meter temperature, relative humidity, air pressure, wind speed, downward longwave and shortwave radiation for the u-HRLDAS model, along with the model's simulated LST outputs. Additional remotely-sensed data such as the Digital Elevation Model (DEM), Normalized Difference Vegetation Index (NDVI), latitude, longitude, land use and cover, and slope are also included. The datasets span the summer months (June to August) from 2011 to 2014. Daily LST data from MOD11A1 and MYD11A1 are used as label data. 3) Optimal model identification via automatic machine learning framework: The MODIS LST data in the database serves as training labels, with a 70/30 split for training and validation. Evaluation metrics like RMSE, MAE, and R² guide the selection. We chose the AutoGluon-Tabular framework, developed by Amazon, for its superior performance, which is achieved through bagging and stacking methods.  Finally, the 1-km seamless LST is reconstructed based on the model with the highest accuracy in validation.

Taking Xi’an city in China as the study region, nine models (Weightensemble_L2, LightGBMLarge, XGBoost, LightGBM, CatBoost, LightGBM, ExtraTree, NeuralNetTorch, and NeuralNetFastAI) were trained within the Autogluon-Tabular framework. These models displayed RMSE values ranging from 0.737 to 1.417 K, MAE spanning 0.522 to 1.031 K, and R² from 0.967 to 0.991. Notably, the Weightedensemble_L2 model excelled, with the lowest RMSE (0.737) and MAE (0.522), and the highest R² (0.991), closely followed by the LightBGMlarge model. with RMSE, MAE, and R² values of 0.739, 0.526, and 0.991, respectively. Furthermore, we conducted supplementary testing using four reserved MODIS LST images. Employing the previously trained WeightedEnsemble_L2 model, seamless predictions of MODIS LST were generated at four overpass timestamps: 02:30, 05:30, 14:30, and 17:30. The resulting spatial distributions is similar with the observed LST, validating our method's capability to capture LST's spatial characteristics and ensure spatial continuity compared to the original MODIS LST data.

In conclusion, the proposed fusion framework which integrates numerical modeling and automatic machine learning, successfully reconstructed LST with high accuracy and strong spatial similarities. There are still shortcomings of this method, such as the predicted images losing some spatial details compared to the observations, which need to be improved in the future.

How to cite: Yumin, L., Meiling, G., and Zhenhong, L.: Retrieving gapless 1-km land surface temperature based on numerical model and auto machine learning approach, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-18741, https://doi.org/10.5194/egusphere-egu24-18741, 2024.

Recent advancements in the empirical downscaling of climate fields using Machine Learning have predominantly leveraged computer vision approaches. These methods treat a climate field as an image channel, applying image processing techniques to automatically extract features for the downscaling model from its latent space embeddings. In contrast, this work aims to revisit and validate the potential of tabular and sequential models in the context of grid-by-grid downscaling, where each grid cell in a map is individually downscaled and input features for the downscaling model are selected manually by a climate expert. We present downscaling results for precipitation and evapotranspiration using three distinct models: Long Short-Term Memory (LSTM), Multi-layer Perceptron (MLP), and a hybrid approach that combines Linear Regression with Random Forest. Our discussion includes the setup and optimization strategies for these models to enhance their ability to capture extremes. The merits of this grid-by-grid approach are highlighted, focusing not only on performance and effectiveness in preserving spatial features but also on its flexibility, spatial transferability, ease of model fine-tuning, and training efficiency.

How to cite: Hobeichi, S., Shao, Y., Rampal, N., Bittner, M., and Abramowitz, G.: Revisiting Tabular Machine Learning and Sequential Models to Advance Climate Downscaling, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-7111, https://doi.org/10.5194/egusphere-egu24-7111, 2024.

Accurate downscaling of daily precipitation is crucial for hydrological and climate modeling, especially in regions with complex terrain and a lack of observational data. In such regions, climate reanalysis are not reliable and thus accurate downscaling is usually limited to those locations captured by a (discrete) network of in-situ measurements instead. For this reason, learning to downscale in ungauged locations, whilst maintaining the spatial structure of precipitation, is crucial to effectively downscale (gridded) climate simulations. 

This study introduces a Gaussian Process - Multi-Layer Perceptron (GP-MLP) latent variable model tailored for the probabilistic downscaling of daily precipitation in ungauged locations. By generating spatially coherent precipitation fields, this model addresses key challenges in regional climate impact assessments and water resource management.

The GP-MLP model consists of an MLP that performs non-linear regression, mapping a set of inputs to distributional parameters of a given probability distribution for each spatio-temporal locations, and we induce spatial correlation between locations with a latent variable modelled by a GP  We jointly learn the GP and MLP parameters using variational inference, which critically allows us to model non-Gaussian probability distributions. 

We test our approach in two geographically and climatologically diverse regions: the Swiss Alps and the Langtang Valley in Nepal. The Swiss Alps, with their complex terrain and relatively dense observational network, serve as an ideal region for the initial training of our model. In the Langtang Valley, a high-mountain region with limited ground-based observations, we employ a transfer learning strategy on the model pre-trained in the Swiss Alps. This process involves fine-tuning the model parameters to the unique climatic and topographical features of the Himalayas, thereby enhancing its performance in predicting daily precipitation in this data-sparse region.

Our preliminary findings demonstrate the model's strong capability in producing accurate and spatially coherent predictions of daily precipitation for ungauged locations. The probabilistic nature of the model's outputs is particularly valuable, providing not only predictions of daily precipitation but also quantifying the associated uncertainties - a crucial aspect for risk management in hydrology and agriculture in areas where the paucity of data has traditionally limited detailed climate impact analysis.

How to cite: Girona-Mata, M., Orr, A., and Turner, R. E.: Spatially-Coherent Probabilistic Downscaling of Daily Precipitation in Ungauged Mountain Locations: a Transfer Learning Study in the Swiss Alps and the Langtang Valley, Nepal., EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-15911, https://doi.org/10.5194/egusphere-egu24-15911, 2024.

Climate change is a pressing global challenge, notably impacting sensitive regions like the Alpine area. Its diverse terrain and ecology make it vulnerable to heightened climate risks, including intensified weather extremes due to global warming. Precise local climate predictions are vital for managing risks in vulnerable areas like the Alpine region, necessitating reliable high-resolution climate data and forecasts. Global products often fall short in providing the fine-grained information needed for accurate localized assessments. This work aims to address the critical need for refined, high-resolution seasonal climate forecasts in the Alpine region as a tool to increase the ability to manage and anticipate climate variability and hazardous conditions. The study endeavors to utilize Perfect Prognosis (PP) within Statistical Downscaling (SD), leveraging regression-based Machine Learning (ML) algorithms to enhance the spatial resolution of daily temperature and total precipitation of ECMWF (European Centre for Medium Range Weather Forecasts) SEAS5 (Seasonal Forecast System 5) seasonal forecasts. Four ensemble learning methods — random forest, light gradient-boosting machine (LGBM), Adaptive Boosting (AdaBoost) and Extreme Gradient Boosting (XGBoost) are considered, while CERRA (Copernicus European Regional Reanalysis) reanalysis (5.5 km) is used as reference target. In order to define the optimal ML model and configuration, a preparatory phase is performed in which ML methods are implemented, optimized and inter-compared by considering ERA5 reanalysis predictor fields (~ 31 km) for the training period 1985-2015 and validation period 2016-2020. Initial findings show that LGBM reports better performance in training and validation, demonstrating superior computational speed and efficiency with respect to the others. LGBM reconstructs daily variability with R2 scores of 0.95 for mean temperature and 0.67 for precipitation. Remaining bias as yearly average is -0.05°C fo daily mean temperature and 5.34% for daily precipitation. Other error metrics, e.g., mean absolute error (MAE) and root mean squared error (RMSE) suggest room for improvements, especially in extreme value predictions and annual precipitation averages. LGBM is thus applied and further optimized on SEAS5. The contribution will further elaborate the inter-comparison of ML models and their downscaling skills for seasonal forecasts will be presented and discussed. The expected outcomes of this study in particular will serve as a crucial input of a drought prediction module in 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: Suriyah, D., Alice, C., Alexander, J., and Edzer, P.: Application of Machine Learning Statistical Downscaling to Seasonal Climate Forecasts for the Alpine Region, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-19640, https://doi.org/10.5194/egusphere-egu24-19640, 2024.