To study climate change we rely on global climate models (GCMs) but their resolution is coarse to investigate impacts at the local scale. Hence, a downscaling task is required for the use of these coarse-resolution outputs. In this sense, statistical downscaling methods (SDMs) are commonly applied to analyse the local impacts. Furthermore, a quantification of the uncertainty share of the SDMs is advised to complement the results. However, many choices need to be done before their application and these decisions can bias the outcome of the analysis. This work examines the SDMs’ uncertainty share to evaluate to what extent the different adopted strategies can impact the climate change signal (CCS) associated with the study. For this, eleven research indicators (six representing precipitation extremes) are used with four future scenarios, 28 state-of-art GCMs, and 15 SDMs of two different types (change factor and quantile mapping methods). The uncertainty involved is quantified by the variance decomposition procedure. Three different decisions are tested:

(i) The selection of the Coupled Model Intercomparison Project (CMIP) era. The uncertainty shares in phases five and six (CMIP5 and CMIP6, respectively) are compared.  

(ii) The selection of the SDM ensemble based on the SDMs’ methodological construction. More specifically, based on an ensemble of five methods of change factor type (including an event-based change factor weather generator) and an ensemble of ten methods of quantile mapping.

(iii) The selection of the optimal SDM ensemble number. Different unique SDMs combinations are tested from k-ensemble members in [2,n] with n as the ensemble with the largest number of members (n=15).

To complement the analysis, the outcomes of the CCSs from all the combinations in (ii) and (iii) are analysed. The results showed that the uncertainty quantification of the SDMs is not sensitive to the selection of the CMIP era. However, this choice is important if the focus is on the GCMs and future scenarios. Hence, it is preferable (but not mandatory) to perform the analysis with the most recent era. The selection of the SDMs based on a methodological construction might bias the conclusions. Therefore, it is better to include methods from all possible types since the results showed that the more methods included in the downscaling, the more reliable the estimation of the SDMs’ uncertainty share. The CCS seems to strongly depend on the choice of the SDM ensemble, and it tends to converge from different k-ensemble members in [2,n] towards the largest ensemble (n). Hence, CCSs from large SDM ensembles will be more reliable. Future work must extend the analysis into different climatological regions and include more methods from all the possible types.

How to cite: Mendoza Paz, S. and Willems, P.: The statistical downscaling methods’ uncertainty share as a measure for adopted strategies in downscaling studies for climate change, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-83, https://doi.org/10.5194/egusphere-egu23-83, 2023.

The intensity and frequency of extreme rainfall events are likely to increase under projected climate change scenarios. Given the adverse socio-economic impacts of these extreme events, we need to model their risk to develop effective policies for adaptation and mitigation. Simulating local hydrometeorological processes at the resolutions essential for assessing impacts and planning is computationally expensive using global climate models. Thus, there is a demand for efficacious downscaling from the coarse-resolution climate model outputs to the finer local scales of interest. Here, we develop a dynamic data-driven model coupled with physics, to downscale coarse-resolution climate model outputs (0.25° × 0.25°) to high-resolution (0.01° × 0.01°) rainfall. The downscaled rainfall is initially estimated by actively searching data on a manifold to learn the downscaling function incrementally using an iterative Gaussian process (GP). Upon convergence, the “first-guess” downscaled rainfall field, along with a physics-based estimation of orographic rainfall are processed by an adversarial learning framework (GAN) to refine finer-scale details. A stochastic sampling model and optimal estimation are used to correct the biases and obtain the final rainfall super-resolution fields. We assess the skill of the proposed model, using ERA5 reanalysis data and Daymet observation data at different terrain conditions (plain and hilly), and show that the downscaled rainfall closely matches the ground truth spatial patterns and extreme rainfall risk. By comparing the performance of individual components of our model (GAN, GP, and Physics) we find that the combined model outperform the individual components, and the GAN accounts for the maximum performance gain of the downscaling model.

How to cite: Saha, A. and Ravela, S.: Downscaling Precipitation Extremes Using Physics-coupled Dynamic Data Driven Adversarial Learning, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-10931, https://doi.org/10.5194/egusphere-egu23-10931, 2023.

High-resolution precipitation time series are required for numerous applications in hydrology. For data-scarce regions, precipitation reanalysis products (PRP) are a promising data source. We validated two PRP for Slovenia and identified biases, which disable a direct usage of the PRP. However, the PRP were used for the parameter estimation of a cascade model to disaggregate daily time series, which exist for long periods for the whole country. The so assimilated data benefits from the advantages from both datasets: the daily rainfall amounts from the observations and the high-resolution temporal structure from the PRP. The disaggregated time series show a superior representation of the observed high-resolution point and areal precipitation time series in comparison to the PRP themselves, and their usage is recommended instead. The developed concept can be transferred to other data-scarce regions.

In more detail, from the latest PRP two are most promising due to their spatial and temporal resolution: ERA5-Land (raster width ~9 km width, temporal resolution of 1 h) and REA6 (6 km, 1 h). ERA5-Land and REA6 are evaluated in space and time by continuous and event-based characteristics as well as precipitation extreme values for five recording stations and 20 catchments in Slovenia. Both PRP show underestimations of dry spell duration, wet spell amount and average intensity, while wet spell duration is overestimated. For extreme values with 1 h duration both PRP lead to underestimations, whereby the bias increases with the return period. The identified biases are larger for ERA5-Land than for REA6. The PRP time series were used for the parameter estimation of a micro-canonical cascade model to disaggregate observed daily values to hourly values. The so estimated parameters differ from station-based estimations, e.g. probabilities for the generation of dry time steps (P(1/0), P(0/1)) are underestimated. Nevertheless, starting from observed daily rainfall amounts the disaggregated time series show a superior representation of the high-resolution precipitation characteristics in comparison to ERA5-Land and REA6. This conclusion is based on all studied precipitation characteristics.

How to cite: Müller-Thomy, H., Kellner, J., Nistahl, P., Bezak, N., Zabret, K., and Schröter, K.: The unintended usage of precipitation reanalysis products: Downscaling of daily precipitation time series to hourly values using reanalysis products for parameter estimation, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-624, https://doi.org/10.5194/egusphere-egu23-624, 2023.

Climate change impact on extreme precipitation is of great importance to society. Small-scale, short-term events can have massive social and socioeconomic consequences. The present study analyses a new sub-kilometre (750 m) HARMONIE-Climate¹ model simulation driven by ERA5 reanalysis data. The new sub-kilometre climate model data (750 m) is compared to NorCP data² from climate models in 3, 5, and 12 km grid spacing, rain gauge station data and reanalysis data in 31 and 79 km resolution. The study examines a case area covering Denmark for five cloudburst seasons (April – October). The study aims to analyse how convective events are represented in the climate model data across grid resolution, and if an added benefit can be identified moving to sub-kilometre resolution.

Extreme convective events are analysed across datasets with respect to diurnal cycle, intensity levels and spatial structure. This is done at both hourly and sub-hourly scales. The 750 m climate model performs better for most metrics. However, climate models with 3 and 5 km grid spacing also perform well. The added computational and storage cost of the sub-kilometre scale experiments, thus only results in limited added benefit for this specific model set-up. Analysing hourly and sub-hourly temporal scales shows that the model performance varies between different temporal scales. The convection-permitting models, in general, represent hourly extremes much better than sub-hourly extremes. The sub-hourly scale is, therefore, essential to analyse to assess the model performance of convective events.

¹ Belušić D, De Vries H, Dobler A, Landgren O, Lind P, Lindstedt D, Pedersen RA, Carlos Sánchez-Perrino J, Toivonen E, Van Ulft B, et al (2020) HCLIM38: A flexible regional climate model applicable for different climate zones from coarse to convection-permitting scales. Geosci Model Dev 13:1311–1333. https://doi.org/10.5194/gmd-13-1311-2020

² Lind P, Lindstedt D, Kjellström E, Jones C (2016) Spatial and Temporal Characteristics of Summer Precipitation over Central Europe in a Suite of High-Resolution Climate Models. J Clim 29:3501–3518. https://doi.org/10.1175/JCLI-D-15-0463.1

How to cite: Dybro Thomassen, E., Arnbjerg-Nielsen, K., J. D. Sørup, H., L. Langen, P., Olsson, J., A. Pedsersen, R., and B. Christensen, O.: Sub-kilometre resolution climate model data: Added benefits in the representation of extreme precipitation?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6866, https://doi.org/10.5194/egusphere-egu23-6866, 2023.

Recent increases in computational resources have led to the application of kilometre- and sub-kilometre-scale simulations in research, numerical weather prediction, and climate modelling alike. Despite anticipated improvements with resolution, there is still considerable work needed to evaluate how well such models improve the representation of intense convection. In this study we conduct ensemble simulations with kilometre- and sub-kilometre-scale horizontal grids to investigate intense convective events in the tropical island thunderstorm system Hector, which frequently occurs over the Tiwi Islands in North Australia. To avoid losing information through spatio-temporal averaging we apply a tracking algorithm to simulated and observed storms. When compared with observations, the model storms exhibit a lack of propagation across the study domain. In general, simulated storms are too intense but too small and too short-lived. This is especially true for the sub-kilometre simulations, where storms are more intense, smaller, and more numerous than in the kilometre-scale counterparts. We argue that size and duration errors compensate for storm number and intensity errors, which could lead to misleading interpretations when only comparing time and space averages of rainfall fields. Investigating some properties of the simulated storms suggests that storms with high rainfall intensities have stronger updrafts in the sub-kilometre model and are accompanied by an increase in cold pool intensity. The results and their resolution sensitivities highlight that the remaining parametrisations and their many tuning parameters in high-resolution set-ups influence the representation of convective storms in such models.

How to cite: Bergemann, M., Lane, T., Wales, S., Narsey, S., and Louf, V.: High-resolution simulations of tropical island thunderstorms: Does an increase in resolution improve the representation of extreme rainfall?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-6981, https://doi.org/10.5194/egusphere-egu23-6981, 2023.

Precipitation is one of the most important climate variables in many aspects due to its key impact on agriculture, water management, etc. Furthermore, extreme precipitation events can lead to excess surface water and floods and are becoming an amplifying societal cost as a result of urbanization and our warming climate. However, it remains a challenge for climate models to realistically simulate the regional patterns, temporal variations, and intensity of precipitation. Detailed knowledge about extreme precipitation events is important for advanced predictions on weather-to-climate time scales. The difficulty arises from the complexity of precipitation processes within the atmosphere stemming from cloud microphysics, cumulus convection, large-scale circulations, planetary boundary layer processes, and many others. This is especially true for heterogeneous surfaces with complex orography such as the Carpathian region.

In order to quantify the impact of the use of different parameterization schemes on regional climate model outputs, hindcast experiments were completed applying RegCM4.7 to the Carpathian region and its surroundings at 10-km horizontal resolution using ERA-Interim reanalysis data as initial and boundary conditions. In this study, 24 simulations were carried out by using various combinations of the physics schemes (2 land surface, 2 microphysics, 3 cumulus and 2 boundary layer schemes) for the year 2010, which was the wettest year in the region since the beginning of the regular measurements. Each parameterization combination leads to different simulated climates, so their spread is an estimate of the model uncertainty arising from the representation of the unresolved phenomena. The analysis of the RegCM ensemble indicates systematic precipitation biases, which are linked to different physical mechanisms in the summer and winter seasons.

Based on the results, RegCM is sensitive to the applied convection scheme, but the interactions with the other schemes (e.g., land surface or microphysics) affect not only the total precipitation, but also the convective and stratiform precipitation in some cases. Due to the different treatment of moisture in the schemes, there are differences not only between the representation of the precipitation cycle, but also in other climatological variables such as soil moisture, temperature and cloud cover.

How to cite: Kalmár, T., Pongrácz, R., and Pieczka, I.: Parameterization-based uncertainties in RegCM simulations over the Carpathian region in a wet year, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-736, https://doi.org/10.5194/egusphere-egu23-736, 2023.

Tuning is now recognised as a key step in climate modelling, and the rise of machine learning techniques is increasing the number of targets used to tune these models. There are several reasons to focus on continental surface tuning. Firstly, a significant part of the sources of uncertainty in regional climate projections lie in the interactions between the atmosphere and the land surface. Secondly, the quality of climate change impact studies highly depends on a good representation of the climate at the surface. Finally, tuning at the surface can benefit from observational sites that provide multivariate, in situ hourly data of many meteorological, radiative and turbulent flux variables.

The objective here is to constrain the water and energy balances at the atmosphere-continental surface interface in the IPSL GCM, using as reference the in-situ observations of the SIRTA instrumented site (Paris suburb). A configuration of the coupled atmosphere (LMDZ) and continental surface (ORCHIDEE) model is set up on a zoomed grid in order to have a 30 km side mesh on the SIRTA point while keeping a reasonable computational cost. In addition, the winds (and possibly the temperature and humidity) are nudged towards the ERA5 reanalyses in order to compare the weather sequences observed at SIRTA with those of the climate model. This nudging technique allows a significant part of the internal variability of the local meteorology simulated by the GCM to be removed and to compare observations and model on a day-to-day basis. An essential step in setting up the tuning of this configuration is to assess the different sources of uncertainty involved. In this presentation, the characterisation of the uncertainties associated with the choice of configuration and the internal variability will be addressed more specifically, with a focus on clouds and precipitation.

In order to characterise the uncertainty linked to the internal variability, we compare the precipitation variability of a simulation ensemble with perturbed initial conditions with that of a perturbed physical ensemble obtained by machine-assisted exploration of the free parameters of the models. The internal variability of the precipitation simulated at SIRTA is found to be of the same order of magnitude as the parametric sensitivity, especially during convective periods, which questions the possibility of a tuning against SIRTA observations. We use a rainfall product (combining radar and rain gauges) from Météo-France in order to evaluate both the representation of spatial and temporal variability in a wider area around SIRTA and the associated uncertainty for tuning. We also present results concerning the uncertainty due to the configuration based on sensitivity tests to the grid configuration and the nudging setup. Finally, we evaluate the part of the precipitation variability due to the soil response by imposing an evaporation factor on the study area. We show how this configuration can be used in the atmosphere model tuning strategy, as it allows to get rid of the rainfall evaporation feedback.

How to cite: Coulon Decorzens, M. and Hourdin, F.: Assessment of precipitation variability sources in a GCM for the implementation of a tuning methodology using in-situ observations, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-12080, https://doi.org/10.5194/egusphere-egu23-12080, 2023.

Around 22% of the global population depend on mountain runoff for their water supply. Due to its importance for future water resources, as well as flood and drought planning, an improved understanding of spatial precipitation patterns in mountain regions is needed. Precipitation gauge networks are sparse and traditional methods of interpolation yield inadequate precipitation fields for poorly gauged mountain catchments.

This research project builds on a new method, Random Mixing, to generate multiple random spatial daily precipitation fields, conditioned on gauge observations. The Random Mixing algorithm has so far been tested on larger, densely gauged catchments. This project adapts the approach for a sparsely gauged, small 9.1 km² mountain catchment, Marmot Creek Research Basin in Alberta, Canada, where elevations range between 1600 m and 2825 m above sea level (a.s.l.). Quality-controlled total precipitation (i.e., rainfall and snowfall) gauge observations, for an 11-year period, from three weather stations around the catchment have been used to condition the random spatial fields.

To optimise selection of the most plausible fields, ensemble hydrological simulations are run, initially using a Python-coded version of the HBV spatially-distributed conceptual model, on a 50 m² regular model grid. Optimisation involves the use of metrics, primarily Nash-Sutcliffe Efficiency (NSE) and bias, to identify the fields that result in the best match between observed and simulated streamflows. Sensitivity of these fields to seasonality, elevation and precipitation intensity is tested.

Results so far are promising. Even with very few gauges, improving the way that spatial covariance relationships between gauge locations are represented in the model has enhanced the quality of the spatial fields. The biggest improvement to date is from explicitly modelling the precipitation / elevation relationship, introducing gradients, and applying daily dry day and wet day parameters to each grid cell across the model domain.

Intended future work will aim to further refine the process using a physically-based spatially distributed model, the Cold Regions Hydrological Model (CRHM). Spatial fields generated using other random methods will be used to evaluate the performance of the new technique. Long time-period flood frequency curves generated using each approach will be compared. Different methods of phase partitioning will be evaluated to identify impacts on extreme flooding which is often controlled by snowpack melt. Climate change perturbations will be applied to generate potential future flood estimates.

How to cite: Shotton, K., Lewis, E., Pritchard, D., and Rutter, N.: Developing precipitation datasets for mountain regions in a changing climate, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1126, https://doi.org/10.5194/egusphere-egu23-1126, 2023.

In recent years, several major rainfall events have been observed in Italy, with amounts that have broken previous all-time records. Several questions concerning the adequacy of the statistical tools that we have at our disposal to determine the "real" rarity of these events emerge, especially considering the limited availability of long and complete rainfall records. In this work, we investigate the influence of "Super-Extremes" on the rainfall regional frequency analysis framework. More specifically, we consider the all-time Italian record events up to now, some of which were observed in 2021 (377.8 mm / 3h, 496 mm / 6h, 740.6 mm / 12h).

The approach is undertaken through a rainfall regional frequency analysis performed over the North-West of Italy based on the patched kriging (PK) technique. PK requires a year-by-year application of ordinary kriging, that overcomes the data inconsistency by considering all the time series, without the need to discard those shorter than a specific length. The morphology of the areas is quite complex, which implies that extremes are expected to be influenced by the elevation: the orographic gradient is computed and removed and, for each duration, the sample variogram is evaluated as the mean of the annual variograms weighted on the number of active rain gauges for any year.

The sequential application of the ordinary kriging allows to reconstruct both a "rainfall data cube" and a "variance data cube" in the (x, y, t) space. A complete series of measured and estimated values are obtained by coring the data cube along the time axis in each location. The cored series are then used to compute the L-moments, in a framework that assigns weights based on the kriging variance, to consider the different nature of the data (measured and estimated). To overcome possible inconsistencies of the L-moment, a bias-correction procedure is applied to preserve the coefficient of variation from the smoothing effect induced by the spatial interpolation.

The methodology is applied to short-duration (1 to 24 hours) annual maximum rainfall depths recorded by rain gauges coming from the Improved Italian – Rainfall Extreme Dataset (I²-RED). The effects in the local frequency curves when introducing new record-breaking data are examined and commented, in view of the role that these values assume in the surrounding region.

How to cite: Mazzoglio, P., Butera, I., and Claps, P.: Effects of Super-Extremes in the evaluation of the design rainfall: a case study in Northern Italy, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-668, https://doi.org/10.5194/egusphere-egu23-668, 2023.

The space-time behaviour of precipitation is very complex. The knowledge of the dependence structures in space and time is very important for the assessment of flood risks. There are many different models available for stochastic simulations of precipitation time series. Most of the models are constructed such that the simulated time series match the autocorrelation structure of the observations in time along with the reproduction of spatial correlations. However, both auto and spatial correlations are value dependent i.e., if it is the upper or the lower tail. High and low intensity values have different dependence structures which have a significant influence on simulated extremes in space. In this presentation, first indicator correlations are introduced to show the intensity dependence of precipitation both in space and time at various resolutions. Then, a stochastic simulator based on gradual change of the correlations for the values in different parts of the distributions is introduced. The idea is that value dependent correlation is changed in such a way that the overall values remains same as that of a reference but not when considering values in different sections of the distributions alone. The model is applied to a large number of German catchments with hourly temporal resolution. The results are carefully analysed and compared to classical approaches.

How to cite: Bárdossy, A. and Anwar, F.: A stochastic rainfall model with intensity dependent autocorrelations., EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7682, https://doi.org/10.5194/egusphere-egu23-7682, 2023.

We propose an approach for stochastic simulation of realistic continuous snow depth time series using a snow depth estimation model and a stochastic weather generation model. The snow depth estimation model consists of three steps: (1) determination of the precipitation type, (2) estimation of  the snow ratio, and (3) estimation of the decreased snow depth. In the first step, air temperature and relative humidity are used as indicators to determine the type of precipitation when precipitation occurs. In the second step, when the type is determined as snow, the snow ratio is estimated, converting the depth of precipitation into depth of fresh snow. Here, the air temperature is used as an indicator to estimate the snow ratio using sigomidal relationship with the snow ratio. In the last step, the amount of decreased snow depth was estimated using a novel temperature index snowmelt equation considering a trend of depth-dependent decreasing snow depth. The snow depth estimation model was applied to the four snowiest meteorological stations of Korea and yielded high Nash Sutcliffe efficiency values which ranged between 0.745 and 0.875 for calibration, and ranged between 0.432 and 0.753 for validation. This calibrated snow depth estimation model was then applied to the simulated weather time series (precipitation, temperature, and relative humidity) from the stochastic weather generation model to simulate continuous snow depth time series. The simulated snow depth data accurately reproduced standard and extreme value statistics of the observed data, the latter of which were consistent with the estimates provided in Korean Building Code. Then, the model was extended to investigate the influence of climate change on the future snow depth. For this, future weather statistics were obtained by applying factor of change to the current weather statistics and then were used to calibrate the weather generation model. Lastly, the future snow depth time series for three future time windows (2021-2040, 2041-2070, and 2071-2100) were simulated using future weather time series and snow depth estimation model.

This research was supported by a grant(2022-MOIS61-003) of Development Risk Prediction Technology of Storm and Flood for Climate Change based on Artificial Intelligence funded by Ministry of Interior and Safety(MOIS, Korea).

How to cite: Park, J. and Kim, D.: Stochastic Simulation of Realistic Continuous Snow Depth Time Series, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-4747, https://doi.org/10.5194/egusphere-egu23-4747, 2023.

West Africa is one of the most data-poor regions in the world. In-situ precipitation observations are not available for many sites or contain many data gaps, thus leading to uncertainties and biases in hydrological studies in this region. To address this fundamental problem, we present a straightforward stochastic approach based on turning bands to simulate daily precipitation fields. Our approach is based on meta-Gaussian frameworks that generate Gaussian random fields, which are transformed into "real-world" precipitation fields using transfer functions. The simulation approach is tested for multiple extremes (1991 – 2016) in the Ouémé river basin in West Africa using different model settings and the most comprehensive station-based precipitation dataset available for this region. The evaluation shows that our approach is a valuable tool for simulation of daily precipitation fields and clearly outperforms classical interpolation techniques (e.g., ordinary kriging). Moreover, the simulation method can be conditioned on observations, uses only a small set of parameters and is an efficient algorithm for ensemble generation of precipitation fields for ungauged areas and design events.  In our West African research projects FURIFLOOD, the precipitation simulations are used as input information for hydrological modeling to reconstruct observed flood events and to create improved hazard maps for this region. Overall, the application of this advanced technique contributes to a better understanding of precipitation uncertainties and to the provision of improved station-based precipitation products for this challenging region.  

How to cite: Bliefernicht, J., Rauch, M., Maranan, M., Fink, A., and Kunstmann, H.: Stochastic simulation of daily precipitation extremes in West Africa, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-9370, https://doi.org/10.5194/egusphere-egu23-9370, 2023.

Long continuous time series of meteorological variables such as temperature and precipitation are required for applications such as derived flood frequency analyses. Observed time series are however generally too short, too sparse in space, or incomplete, especially at the sub-daily timestep. Stochastic weather generators allow an alternative to using observations, being able to generate time series of arbitrary length which are then used as input to hydrological models.

A hybrid hourly space-time weather generator has been developed based on a stochastic alternating renewal rainfall model. Modelling of non-rainfall climate variables is achieved using a non-parametric k-nearest neighbour (k-NN) resampling approach, which is coupled to the space-time rainfall model via rainfall state.

Circulation pattern (CP) or weather pattern classifications can be useful as a conditioning variable for stochastic rainfall models and weather generators. One primary use is the downscaling of future climate scenarios. Furthermore, CP conditioned models may better simulate rainfall and other climate variables through a better partitioning of observations into distinct rainfall and weather types.

Previous research has shown that the point rainfall model performs better, particularly regarding extremes, if conditioned on an optimised fuzzy-rule based objective weather pattern classification. Appropriate model revisions have now been made to allow the full hybrid space-time weather generator to also be conditioned on this classification.

This study assesses the performance of the weather pattern conditioned hybrid weather generator compared to the previous seasonal (summer-winter) conditioned model. For testing, 400 meso-scale catchments across Germany were selected. 

How to cite: Pidoto, R. and Haberlandt, U.: A CP conditioned hybrid weather generator, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14531, https://doi.org/10.5194/egusphere-egu23-14531, 2023.

Rainfall Intensity-Duration-Frequency (IDF) curves are widely used in studies related to planning, design, and operation of various water control (e.g., barrages, dams, levees) and conveyance structures (e.g., culverts, spillways, storm sewers) for mitigating risk associated with floods attributable to extreme precipitation. In many parts of the globe, precipitation data are limited, and the network of gauges is sparsely distributed. Therefore, the use of only at-site data for the construction of IDF curves could have large uncertainties. To overcome this impediment, regional IDF relationships could be developed by regional frequency analyses (RFA) which uses information pooled from several meteorologically similar sites. Recently, there is growth in the use of fine spatial scale remote-sensing precipitation products to arrive at IDF relationships for ungauged locations, as the spatial coverage of these products is exhaustive. However, recent studies indicate that most of the remote sensing products underestimate the precipitation intensities corresponding to different durations and return periods and also perform worse at shorter time scales (e.g., daily and sub-daily). Although remote sensing products can be corrected for biases before use in developing the IDF relationships, there is ambiguity in the choice of bias correction methods. Furthermore, in sparsely gauged locations, the availability of only a limited number of ground observation stations for bias correction enhances uncertainty in the developed IDF relationships. In addition, relying on only one satellite product may not be meaningful, as the skill of different satellite products varies across the globe. Also, the conventional practice of developing IDF curves considering the stationary assumption may lead to large biases in estimates of precipitation extremes in a changing climate. To address these issues, this study proposes a novel methodology to develop non-stationary regional IDF relationships for use in climate change scenarios. The methodology involves nonstationary RFA utilizing fine grid-scale daily precipitation derived by merging multiple satellite-based precipitation products and ground-based precipitation products for homogenous extreme precipitation regions (EPRs). The merging of different products is achieved using a novel random forest-based regression method. Effectiveness of the proposed methodology is demonstrated through a case study on Karnataka state in India, which extends over approximately 0.2 million square kilometers. The homogenous EPRs are delineated in the study area using ensemble cluster analysis of the relevant predictor variables/covariates. Non-stationary regional IDF curves are developed using the proposed methodology corresponding to different CMIP6 climate change scenarios, considering an ensemble mean precipitation derived from eleven GCMs (General Circulation Models). The curves are compared with those obtained using conventional stationary methods considering block-maxima and partial duration series of extreme precipitation.

How to cite: Goel, A. and Srinivas, V. V.: Deriving Regional IDF Curves for Data-Sparse Areas in Climate Change Scenarios using Merged Satellite and Ground-based Precipitation and GCMs, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-304, https://doi.org/10.5194/egusphere-egu23-304, 2023.

Precipitation nowcasting, aiming to predict the rainfall intensity in the near future (usually 0-2h) [1], is crucial for urban planning, flood monitoring, agriculture management, and so on. Numerical weather modeling (NWP) takes a variety of data sources as the input of complex computer models that use mathematical equations to simulate the behavior of the atmosphere. Limited by the time needed for model spin-up, the performance in the short near future is not satisfactory. Deep learning (DL)-based method fills in the gap by treating nowcasting as a video prediction problem. The Convolutional LSTM [2] extracts spatial information when dealing with temporal series. The Generated Adversarial Network (GAN)-based [3] method shows potential in simulating the realisticness of the precipitation field. However, training such a model is very time-consuming and data-demanding [3] [4]. Different from natural images, the precipitation field to be estimated usually has a larger spatial size. Moreover, the convolutional layers tend to oversmooth the output and eliminate the small patterns that are important for the meteorologists to make the decision. Thus, we proposed a two-stage framework: the first one is to train an RNN-based model on the coarse field. The second is to downscale and style transfer from the coarse field to high-resolution precipitation maps based on GAN and Graph Convolutional Network (GCN). The coarse prediction will act as a constraint to the finer scale output and allows re-assignment of the spatial distribution of intensities. Such probabilistic output prevents the overestimation of the intensity. RNN is good at capturing long-range characteristics, and GCN [5] can extract local and neighborhood information, thus these two channels are naturally complementary to improve both local patterns and global accuracy scores. The GAN is used to make final output similar to real precipitation maps such as radar scans. To train the model, we downloaded the 2006-2016 ERA5 total precipitation at 0.25-degree spatial resolution and the DWD radar map [6] at 1km spatial resolution. We expect our model can capture the overall coverage of rainfall events and depict the spatial details. More importantly, this alleviates the data shortage problem, i.e., high-resolution precipitation nowcasting at places without ground-based radar stations can be acquired.

[1] Shi, Xingjian, et al. "Deep learning for precipitation nowcasting: A benchmark and a new model." Advances in neural information processing systems 30 (2017).

[2]Shi, Xingjian, et al. "Convolutional LSTM network: A machine learning approach for precipitation nowcasting." Advances in neural information processing systems 28 (2015).

[3] Ravuri, Suman, et al. "Skilful precipitation nowcasting using deep generative models of radar." Nature 597.7878 (2021): 672-677.

[4] Sønderby, Casper Kaae, et al. "Metnet: A neural weather model for precipitation forecasting." arXiv preprint arXiv:2003.12140 (2020).

[5] Shi, Yilei, Qingyu Li, and Xiao Xiang Zhu. "Building segmentation through a gated graph convolutional neural network with deep structured feature embedding." ISPRS Journal of Photogrammetry and Remote Sensing 159 (2020): 184-197.

[6] Ayzel, Georgy, Tobias Scheffer, and Maik Heistermann. "RainNet v1. 0: a convolutional neural network for radar-based precipitation nowcasting." Geoscientific Model Development 13.6 (2020): 2631-2644.

How to cite: Zhao, S., Xiong, Z., and Zhu, X. X.: A Coarse-to-Fine Deep Learning Framework for High-Resolution Future Precipitation Map Generation, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-7751, https://doi.org/10.5194/egusphere-egu23-7751, 2023.

Accurate precipitation representation at local and global scales will greatly improve our understanding of climate system changes. However, no precipitation estimate consistently has the lowest errors (systematic biases, random error, and rain/no-rain classification error) under varying environmental gradients, resulting in considerable uncertainty when investigating mechanisms and making predictions. Multiple Source Precipitation Ensemble (MSEP) is regarded as an indispensable approach to this challenge. Based on an automatic machine learning workflow, we propose an MSPE framework that uses machine learning classification and regression jointly. Six distinct precipitation products (e.g., satellite- and reanalysis-based estimates) and their ensembles based on different framework strategies were examined comprehensively at 818 gauges across China and 500 randomly selected sites (representing ungauged regions). The unique features of MSPE were investigated, including the necessity of assigning spatiotemporal dynamic weights and the usage of machine learning classification and regression jointly. Results demonstrated that MSPE could effectively reduce both random and classification errors associated with precipitation occurrences. In addition, the capacity to generalize and the interpretability of the ML models developed within the framework were compared and discussed in depth. We also summarized the current framework's limitations and potential expansions. The framework presented in this research is expected to be a robust and flexible framework for the global application of ensembles of precipitation estimates from numerous scales, data sources, and time periods.

How to cite: Chen, H., Wang, T., Montzka, C., Gao, H., Guo, N., Chen, X., and Vereecken, H.: On the improved ensemble of multi-source precipitation through joint automated machine learning-based classification and regression, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14002, https://doi.org/10.5194/egusphere-egu23-14002, 2023.

Climate models and their future projections, are normally provided in coarse spatial resolutions which makes them an imprecise source of information for certain hydrological purposes. Finding the proficient means of downscaling such data is one of the open questions of climate research. Previous research has shown that, the rainfall extremes show self-similarity in time and that a relatively similar behavior exists in regard to the spatial scale as well (Veneziano et al 2002). This study aims at determining the spatial scaling relationship of the rainfall extremes by using fine grids of radar datasets and upscaling them. In an empirical manner by aggregating the radar rainfall cells in space and for different cell sizes with a = 1, 2, 3, …12 km and for different durations of d = 5 min, 15 min 30 min, 1 hr, 2 hr, 4 hr, …, 24 hr the Annual Maximum Series are extracted. Using the AMS of different spatial and temporal scales and applying the Koutsoyiannis et at. 1998 method for rainfall extreme value analysis, the probability distribution function is fitted. Assessing the changes of the PDF parameters with the scale, with a logarithmic transformation on both variables; ln(parameter) vs. ln(scale), can show the sought relationship. The preliminary results of the study show definable non-linear relationships for location and scale parameters of the GEV distribution and the eta parameter of the Koutsoyiannis et al. 1998 parametrization.

Koutsoyiannis, D. Kozonis, and A. Manetas, A mathematical framework for studying rainfall intensity-duration-frequency relationships, Journal of Hydrology, 206 (1-2), 118–135, doi:10.1016/S0022-1694(98)00097-3, 1998.

Veneziano, Daniele; Furcolo, Pierluigi (2002): Multifractality of rainfall and scaling of intensity-duration-frequency curves. In Water Resour. Res. 38 (12), 42-1-42-12. DOI: 10.1029/2001WR000372.

How to cite: Goshtsasbpour, G., Haberlandt@iww.uni-hannover.de, U., Sharma, A., El Hachem, A., Seidel, J., and Bardossy, A.: Determination of the spatial scaling relationship of rainfall extremes using radar data, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-14968, https://doi.org/10.5194/egusphere-egu23-14968, 2023.

The Mediterranean basin is a complex environment characterized by both exceptional biodiversity and intense human presence. This environment is already highly influenced by the anthropogenic activities, but their importance is expected to rise up due to the forecasted increase of population (from 480 million people to 520-570 million by 2030) and ongoing climate change. These conditions will trigger an increase of human pressures, including urbanization, industrialization, the expansion of intensive agriculture activities (i.e., irrigation) and aquaculture, thus threatening the natural resources availability, and specifically the water availability. Droughts, floods and landslides events are already stepping up in this environment, making it urgent to develop models systems capable to predict the extreme complex and widespread climate variations of this area.

ESA recognized the crucial role of this region by funding the Digital Twin Earth (DTE) Hydrology Evolution and the 4dMED projects, specifically dedicated to reconstruct the Mediterranean terrestrial water cycle at 1km spatial and 1 day temporal resolution and to develop a prototype of Digital Twin for the entire Mediterranean basin, which can be used for the prediction of hydrological extremes, the management of the water resource cycle and the simulation of the changes that the system may undergo. To reach those objectives, the latest developments of Earth Observation (EO) data as those derived from the ESA-Copernicus missions will be exploited together with in situ observations, hydrological and hydraulic models, artificial intelligence tools and advanced digital platform functionalities.

Among the hydrologic variables datasets generated within these projects, precipitation holds a major role due to its influence on the natural hazards occurrence. Here, we show the procedure adopted to generate the high spatial and temporal resolution precipitation product over the Mediterranean region by downscaling and merging different satellite and in-situ precipitation products. The obtained dataset is evaluated against high resolution observed data in several region of the Mediterranean basin, in order to assess its performance with respect to others EO derived precipitation datasets.

How to cite: Filippucci, P., Mosaffa, H., Brocca, L., and Massari, C.: High spatial and temporal resolution precipitation over Mediterranean basin for Digital Twin Earth Hydrology and 4dMED projects, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-15320, https://doi.org/10.5194/egusphere-egu23-15320, 2023.

As a result of satellite observations, ground observations, and data assimilation, global precipitation datasets have been developed for regions like Iran, where ground observations are limited. This study presents a comprehensive evaluation of WaPOR precipitation dataset over Iran at daily time scale. We considered a period of three years from 2019 to the end of 2021 and 394 synoptic rain gauges are used for the assessment. Daily WaPOR precipitation data at 250m scale downloaded and compared pixel-to-point with in-situ data. In addition, the WaPOR data and stations data were compared based on time classification (seasonal), location in the main catchment basins of Iran, and elevation above sea level. Calculating MSE, R score, RMSE and MSLE between real data(stations) and predict data(Wapor) shows some important result: 1. From the time point of view, WaPOR has best performance in summer (MSE = 4.94 and MSLE = 0.16) 2. Location, the best performance is related to stations of the catchment areas of the eastern part of Iran (Qaraqom basin with MSE = 11.9 and eastern border basin with MSE = 6.26) and the worst performance is related to the catchment area of the Caspian Sea (Mazandaran Sea basin with MSE = 64.06). 3. For analyzing the effect of elevation on precipitation, we divided the stations into 5 groups with an interval of approximately 600 meters (according to the lowest and highest elevation, which is -25 meters and 2965 meters). The best performance is related to stations with an altitude between 572 and 1170 meters (MSE = 21.61) and the worst is related to stations with an altitude between 1768 and 2366 meters (MSE = 43.79). 4. Moreover, on average for each station, in the three years of study (1096 days), we have 166 days (with standard deviation 119 days) that station has recorded precipitation but WaPOR dataset didn’t represent any record, so it’s not appropriate for daily hydrological models. 5. The difference between the three-year precipitation total at the station and the WaPOR precipitation total is 449.6 mm on average (with standard deviation 724.5).

How to cite: Moghaddas, M. and Tajrishy, M.: Is WaPOR precipitation data reliable over Iran?, EGU General Assembly 2023, Vienna, Austria, 23–28 Apr 2023, EGU23-1357, https://doi.org/10.5194/egusphere-egu23-1357, 2023.