Raw Model Output Research Articles

Confidence in Covid-19 models

Epidemiological models of the transmission of SARS-CoV-2 played an important role in guiding the decisions of policy-makers during the pandemic. Such models provide output projections, in the form of time -series of infections, hospitalisations, and deaths, under various different parameter and scenario assumptions. In this paper I caution against handling these outputs uncritically: raw model-outputs should not be presented as direct projections in contexts where modelling results are required to support policy -decisions. I argue that model uncertainty should be handled and communicated transparently. Drawing on methods used by climate scientists in the fifth IPCC report I suggest that this can be done by: attaching confidence judgements to projections based on model results; being transparent about how multi-model ensembles are supposed to deal with such uncertainty; and using expert judgement to ‘translate’ model-outputs into projections about the actual world. In a slogan: tell me what you think (and why), not (just) what your models say. I then diffuse the worry that this approach infects model-based policy advice with some undesirably subjective elements, and explore how my discussion fares if one thinks the role of a scientific advisor is to prompt action, rather than communicate information.

Melt pool level flaw detection in laser hot wire directed energy deposition using a convolutional long short-term memory autoencoder

While additive manufacturing has seen rapid proliferation in recent years, process monitoring and quality assurance methods capable of detecting common flaws have seen little improvement and remain largely expensive and time-consuming. This work focuses on process monitoring using video data for a large-scale directed energy deposition process, laser hot wire additive manufacturing. This data was collected using a video camera mounted on the robot arm pointed at the melt pool. A deep learning enabled video-based process monitoring method was applied to the data, which leveraged a combined convolutional long short-term memory (ConvLSTM) artificial neural network. The trained architecture was used to predict future frames of footage. The model’s application in flaw detection is based on its ability to faithfully construct future frames in the video. A metric of process stability was quantified by computing a regularity score between the raw frames and model outputs, with low regularity scores being indicative of an anomaly. Results have demonstrated that the flaw formation mechanisms of interest to this study, including wire dripping, arcing, melt pool oscillation, and wire stubbing are detectable using this method. This implementation of a ConvLSTM-based deep learning enabled melt pool video process monitoring technique is therefore potentially a viable quality assurance method as well as a tool to guide traditional quality assurance methods.

Producing reliable hydrologic scenarios from raw climate model outputs without resorting to meteorological observations

Abstract. A simplified hydroclimatic modelling workflow is proposed to quantify the impact of climate change on water discharge without resorting to meteorological observations. This alternative approach is designed by combining asynchronous hydroclimatic modelling and quantile perturbation applied to streamflow observations. Calibration is run by forcing hydrologic models with raw climate model outputs using an objective function that excludes the day-to-day temporal correlation between simulated and observed hydrographs. The resulting hydrologic scenarios provide useful and reliable information considering that they (1) preserve trends and physical consistency between simulated climate variables, (2) are implemented from a modelling cascade despite observation scarcity, and (3) support the participation of end-users in producing and interpreting climate change impacts on water resources. The proposed modelling workflow is implemented over four sub-catchments of the Chaudière River, Canada, using nine North American Coordinated Regional Climate Downscaling Experiment (NA-CORDEX) simulations and a pool of lumped conceptual hydrologic models. Results confirm that the proposed workflow produces equivalent projections of the seasonal mean flows in comparison to a conventional hydroclimatic modelling approach. They also highlight the sensibility of the proposed workflow to strong biases affecting raw climate model outputs, frequently causing outlying projections of the hydrologic regime. Inappropriate forcing climate simulations were however successfully identified (and excluded) using the performance of the simulated hydrologic response as a ranking criterion. Results finally suggest that further works should be conducted to confirm the reliability of the proposed workflow to assess the impact of climate change on high- and low-flow events.

Additive dynamic models for correcting numerical model outputs

Numerical air quality models are pivotal for the prediction and assessment of air pollution, but numerical model outputs may be systematically biased. An additive dynamic model is proposed to correct large-scale raw model outputs using data from other sources, including readings collected at ground monitoring networks and weather outputs from other numerical models. An additive partially linear model specification is employed for the nonlinear relationships between air pollutants and covariates. In addition, a multi-resolution basis function approximate is proposed to capture the different small-scale variations of biases, and a discretized stochastic integro-differential equation is constructed to characterize the dynamic evolution of the random coefficients at each spatial resolution. An expectation-maximization algorithm is developed for parameter estimation and a multi-resolution ensemble-based scheme is embedded to accelerate the computation. For statistical inference, a conditional simulation technique is applied to quantify the uncertainty of parameter estimates and bias correction results. The proposed approach is used to correct the biased raw outputs of PM2.5 from the Community Multiscale Air Quality (CMAQ) system for China's Beijing-Tianjin-Hebei region. Our method improves the root mean squared error and continuous rank probability score by 43.70% and 34.76%, respectively. Compared to other statistical methods under different metrics, our model has advantages in both correction accuracy and computational efficiency.

Robust bias-correction of precipitation extremes using a novel hybrid empirical quantile-mapping method

High-resolution, daily precipitation climate products that realistically represent extremes are critical for evaluating local-scale climate impacts. A popular bias-correction method, empirical quantile mapping (EQM), can generally correct distributional discrepancies between simulated climate variables and observed data but can be highly sensitive to the choice of calibration period and is prone to overfitting. In this study, we propose a hybrid bias-correction method for precipitation, EQM-LIN, which combines the efficacy of EQM for correcting lower quantiles, with a robust linear correction for upper quantiles. We apply both EQM and EQM-LIN to historical daily precipitation data simulated by a regional climate model over a region in the northeastern USA. We validate our results using a five-fold cross-validation and quantify performance of EQM and EQM-LIN using skill score metrics and several climatological indices. As part of a high-resolution downscaling and bias-correction workflow, EQM-LIN significantly outperforms EQM in reducing mean, and especially extreme, daily distributional biases present in raw model output. EQM-LIN performed as good or better than EQM in terms of bias-correcting standard climatological indices (e.g., total annual rainfall, frequency of wet days, total annual extreme rainfall). In addition, our study shows that EQM-LIN is particularly resistant to overfitting at extreme tails and is much less sensitive to calibration data, both of which can reduce the uncertainty of bias-correction at extremes.

Exploring the Watch-to-Warning Space: Experimental Outlook Performance during the 2019 Spring Forecasting Experiment in NOAA’s Hazardous Weather Testbed

Abstract During the 2019 Spring Forecasting Experiment in NOAA’s Hazardous Weather Testbed, two NWS forecasters issued experimental probabilistic forecasts of hail, tornadoes, and severe convective wind using NSSL’s Warn-on-Forecast System (WoFS). The aim was to explore forecast skill in the time frame between severe convective watches and severe convective warnings during the peak of the spring convective season. Hourly forecasts issued during 2100–0000 UTC, valid from 0100 to 0200 UTC demonstrate how forecasts change with decreasing lead time. Across all 13 cases in this study, the descriptive outlook statistics (e.g., mean outlook area, number of contours) change slightly and the measures of outlook skill (e.g., fractions skill score, reliability) improve incrementally with decreasing lead time. WoFS updraft helicity (UH) probabilities also improve slightly and less consistently with decreasing lead time, though both the WoFS and the forecasters generated skillful forecasts throughout. Larger skill differences with lead time emerge on a case-by-case basis, illustrating cases where forecasters consistently improved upon WoFS guidance, cases where the guidance and the forecasters recognized small-scale features as lead time decreased, and cases where the forecasters issued small areas of high probabilities using guidance and observations. While forecasts generally “honed in” on the reports with slightly smaller contours and higher probabilities, increased confidence could include higher certainty that severe weather would not occur (e.g., lower probabilities). Long-range (1–5 h) WoFS UH probabilities were skillful, and where the guidance erred, forecasters could adjust for those errors and increase their forecasts’ skill as lead time decreased. Significance Statement Forecasts are often assumed to improve as an event approaches and uncertainties resolve. This work examines the evolution of experimental forecasts valid over one hour with decreasing lead time issued using the Warn-on-Forecast System (WoFS). Because of its rapidly updating ensemble data assimilation, WoFS can help forecasters understand how thunderstorm hazards may evolve in the next 0–6 h. We found slight improvements in forecast and WoFS performance as a function of lead time over the full experiment; the first forecasts issued and the initial WoFS guidance performed well at long lead times, and good performance continued as the event approached. However, individual cases varied and forecasters frequently combined raw model output with observed mesoscale features to provide skillful small-scale forecasts.

In silico assessment of household level closed water cycles: Towards extreme decentralization

Water management in most of the developed world is currently practiced in a highly centralized manner, leading to major infrastructure and energy costs to transport water. To decrease the impacts of water scarcity and climate change, the decentralization of water can increase local robustness. In extremis, decentralization can involve building or house level water supply and treatment. Here, we constructed a MATLAB/Simulink model for two decentralized water management configurations at the household level, assuming the socio-environmental setting of Flanders, Belgium. Independence from the potable water grid and sewer system was pursued through rainwater harvesting, reuse of wastewater streams fit-for-purpose, and discharge via infiltration. The mass balance for water was calculated over the system boundaries showing high potential for independence from the grid with a reasonable treatment train and storage options. Next, the risk of contaminant accumulation within the circular system was assessed, showing a key limitation on decentralized system performance necessitating a system purge. Up to 59% of system rainwater usage was due to the replacement of this purge. Employing treatment units with high (95%) contaminant rejection efficiencies eliminated contaminant accumulation issues. The raw model output was quantitatively assessed by constructing four newly proposed key performance indicators (KPIs), quantifying system independence, circularity, drought tolerance and local water body recharge, which allowed for facilitated system comparison and communication to stakeholders. A sensitivity analysis was performed in which the effect of input parameter variability and uncertainty on system performance was quantified. The sensitivity analysis showed the importance of water recovery and contaminant removal efficiencies of the applied treatment technologies on system performance when contaminant accumulation in the system forms an issue. In systems not severely affected by pollutant accumulation, parameters such as inhabitant number and roof surface had the largest effect. As a whole, this work shows the potential of extreme decentralization of water systems and addresses the obstacle towards implementation formed by the accumulation of contaminants due to system circularity. Additionally, this study provides a framework for operational and technological decision support of decentralized household-scale water systems and, by extension, for future water policy-making.

Future projections of precipitation, surface temperatures and drought events over the monsoon transitional zone in China from bias‐corrected CMIP6 models

AbstractBias‐corrected monthly precipitation and surface temperature datasets are constructed for the monsoon transitional zone (MTZ) in China during 1965–2100 based on 21 coupled climate models from phase 6 of the Coupled Model Intercomparison Project (CMIP6) using the equi‐ratio and equidistant cumulative distribution function quantile‐based mapping methods (equi‐ratio cumulative distribution function [ERCDF] and equidistant cumulative distribution function [EDCDF]), respectively. The efficiencies of the two methods are verified via cross‐validation by the jack‐knife method, and the biases are remarkably reduced compared to those of the raw model outputs. Then, the bias‐corrected model outputs are applied to future projections of precipitation, surface temperatures and drought events under the medium (SSP2‐4.5) and high (SSP5‐8.5) shared socioeconomic pathway scenarios. The obtained results present pronounced increases in both the projected annual mean precipitation and temperature fields for the model ensemble mean. In accompany, the MTZ is predicted to become drier in the three future periods; this result was attributed to the dramatic increases in semiarid events and to the decreases in humid events. The drought tension showed more rapid development under SSP585 than under the other scenario, with comparable occurrence frequencies of relatively wet and dry events in the long‐term period, posing a serious threat to regional sustainable development.

Characterizing non-stationary compound extreme events in a changing climate based on large-ensemble climate simulations

The dependence structure of temperature-precipitation compound events is analyzed across Canada using three datasets derived from Canadian Regional Climate Model Large Ensemble simulations, including raw model outputs (CanRCM4-LE) and two sets of multivariate bias-corrected model outputs (Canadian Large Ensembles Adjusted Datasets, CanLEAD-EWEMBI/S14FD). The performance of the ensembles to represent tail dependencies corresponding to warm-wet and warm-dry events is evaluated against NRCANmet observations for 1951–2000 using the copula goodness of fit test. The parameters of the copula model are estimated using a Bayesian framework to characterize the corresponding uncertainties. The non-stationarity of compound extreme climate events is analyzed for 1951–2100 using an ensemble pooling approach and the results are compared with the ones based on the independence assumption. Results show that multivariate bias-corrected climate simulations (i.e. CanLEAD) can better represent the correlated temperature-precipitation extremes compared to raw CanRCM4-LE outputs. The estimated joint return periods reduce significantly when the dependence structure is considered, compared to the independence assumption, for most regions especially in winter and summer. Therefore, analysis of extreme temperature and precipitation in isolation can result in dramatic underestimations of compound warm-wet and warm-dry events. Further, there is strong non-stationarity in the dependence structure of temperature and precipitation under climate change that can play a significant role in future compound extremes.

North Atlantic climate far more predictable than models imply.

Quantifying signals and uncertainties in climate models is essential for the detection, attribution, prediction and projection of climate change1-3. Although inter-model agreement is high for large-scale temperature signals, dynamical changes in atmospheric circulation are very uncertain4. This leads to low confidence in regional projections, especially for precipitation, over the coming decades5,6. The chaotic nature of the climate system7-9 may also mean that signal uncertainties are largely irreducible. However, climate projections are difficult to verify until further observations become available. Here we assess retrospective climate model predictions of the past six decades and show that decadal variations in North Atlantic winter climate are highly predictable, despite a lack of agreement between individual model simulations and the poor predictive ability of raw model outputs. Crucially, current models underestimate the predictable signal (the predictable fraction of the total variability) of the North Atlantic Oscillation (the leading mode of variability in North Atlantic atmospheric circulation) by an order of magnitude. Consequently, compared to perfect models, 100 times as many ensemble members are needed in current models to extract this signal, and its effects on the climate are underestimated relative to other factors. To address these limitations, we implement a two-stage post-processing technique. We first adjust the variance of the ensemble-mean North Atlantic Oscillation forecast to match the observed variance of the predictable signal. We then select and use only the ensemble members with a North Atlantic Oscillation sufficiently close to the variance-adjusted ensemble-mean forecast North Atlantic Oscillation. This approach greatly improves decadal predictions of winter climate for Europe and eastern North America. Predictions of Atlantic multidecadal variability are also improved, suggesting that the North Atlantic Oscillation is not driven solely by Atlantic multidecadal variability. Our results highlight the need to understand why the signal-to-noise ratio is too small in current climate models10, and the extent to which correcting this model error would reduce uncertainties in regional climate change projections on timescales beyond a decade.

Statistical adjustment, calibration and downscaling of seasonal forecasts: a case-study for Southeast Asia

The present paper is a follow-on of the work presented in Manzanas et al. (Clim Dyn 53(3–4):1287–1305, 2019) which provides a comprehensive intercomparison of alternatives for the post-processing (statistical adjustment, calibration and downscaling) of seasonal forecasts for a particularly interesting region, Southeast Asia. To answer the questions that were raised in the preceding work, apart from Bias Adjustment (BA) and ensemble Re-Calibration (RC) methods—which transform directly the variable of interest,—we include here more complex Perfect Prognosis (PP) and Model Outputs Statistics (MOS) downscaling techniques—which operate on a selection of large-scale model circulation variables linked to the local observed variable of interest. Moreover, we test the suitability of BA and PP methods for the post-processing of daily—not only seasonal—time-series, which are often needed in a variety of sectoral applications (crop, hydrology, etc.) or to compute specific climate indices (heat waves, fire weather index, etc.). In addition, we also undertake an assessment of the effect that observational uncertainty may have for statistical post-processing. Our results indicate that PP methods (and to a lesser extent MOS) are highly case-dependent and their application must be carefully analyzed for the region/season/application of interest, since they can either improve or degrade the raw model outputs. Therefore, for those cases for which the use of these methods cannot be carefully tested by experts, our overall recommendation would be the use of BA methods, which seem to be a safe, easy to implement alternative that provide competitive results in most situations. Nevertheless, all methods (including BA ones) seem to be sensitive to observational uncertainty, especially regarding the reproduction of extremes and spells. For MOS and PP methods, this issue can even lead to important regional differences in interannual skill. The lessons learnt from this work can substantially benefit a wide range of end-users in different socio-economic sectors, and can also have important implications for the development of high-quality climate services.

Bias Correction of High‐Resolution Regional Climate Model Precipitation Output Gives the Best Estimates of Precipitation in Himalayan Catchments

AbstractThe need to provide accurate estimates of precipitation over catchments in the Hindu Kush, Karakoram, and Himalaya mountain ranges for hydrological and water resource systems assessments is widely recognized, as is identifying precipitation extremes for assessing hydro‐meteorological hazards. Here, we investigate the ability of bias‐corrected Weather Research and Forecasting model output at 5‐km grid spacing to reproduce the spatiotemporal variability of precipitation for the Beas and Sutlej river basins in the Himalaya, measured by 44 stations spread over the period 1980 to 2012. For the Sutlej basin, we find that the raw (uncorrected) model output generally underestimated annual, monthly, and (particularly low‐intensity) daily precipitation amounts. For the Beas basin, the model performance was better, although biases still existed. It is speculated that the cause of the dry bias over the Sutlej basin is a failure of the model to represent an early‐morning maximum in precipitation during the monsoon period, which is related to excessive precipitation falling upwind. However, applying a nonlinear bias‐correction method to the model output resulted in much better results, which were superior to precipitation estimates from reanalysis and two gridded datasets. These findings highlight the difficulty in using current gridded datasets as input for hydrological modeling in Himalayan catchments, suggesting that bias‐corrected high‐resolution regional climate model output is in fact necessary. Moreover, precipitation extremes over the Beas and Sutlej basins were considerably underrepresented in the gridded datasets, suggesting that bias‐corrected regional climate model output is also necessary for hydro‐meteorological risk assessments in Himalayan catchments.

Near‐surface mean wind in Switzerland: Climatology, climate model evaluation and future scenarios

AbstractNear‐surface seasonal and annual mean wind speed in Switzerland is investigated using homogenized observations, Twentieth Century Reanalysis (20CRv2c) data and raw model output of a 75 member EURO‐COoRdinated Downscaling EXperiment regional climate model (RCM) ensemble for present day and future scenarios. The wind speed observations show a significant decrease in the Alps and on the southern Alpine slopes in the period 1981–2010. However, the 20CRv2c data reveal that the recent trends lie well within the decadal variability over longer time periods and no clear signs of a systematic wind stilling can be found for Switzerland. The ensemble of RCMs shows large biases in the annual mean wind speed over the Jura mountains, and some members also show large biases in the Alps compared to station observations. The spatial distribution of the model biases varies strongly between the RCMs, while the resolution and the driving global model have less impact on the pattern of the model bias. The RCMs are mostly able to represent the seasonality of wind speed on the Plateau but miss important details in complex terrain related to local wind systems. Most models show no significant changes in near‐surface mean wind speed until the end of the 21st century. The model ensemble changes range from a 7% decrease to a 6% increase with an ensemble mean decrease of 1 to 2%. Due to model biases, the scale mismatch between model grid and station observations and the missing representation of local winds in the simulations, the changes need to be interpreted with utmost care. Future assessments might lead to major revisions even for the sign of the projected changes, in particular over complex terrain.

Using statistical downscaling to assess skill of decadal predictions

We test a hypothesis that the signal-to-noise ratio of decadal forecasts improves with larger scales and can be utilised through model output statistics (MOS) involving empirical-statistical downscaling and dependencies to large-scale conditions. Here, we used MOS applied to an ensemble of decadal forecasts to predict local wet-day frequency, the wet-day mean precipitation, and the mean temperature for one to nine year long intervals of forecasts with a one year lead time. Our study involved a set of decadal forecasts over the 1961–2011 period, based on a global coupled ocean-atmosphere model that was downscaled and analysed for the North Atlantic region. The MOS for the decadal forecasts failed to identify aspects that were associated with higher skill for temperature and precipitation over Europe on time scales shorter than five years. A likely explanation for not enhancing skillful parts of the forecasts was that the raw model output had low skill for the general large-scale atmospheric circulation. This was particularly true for the wet-day frequency over Europe, which had a strong connection to the mean sea-level pressure (SLP) anomalies over the North Atlantic. There was a weak connection between large-scale maritime surface temperature anomalies and local precipitation and temperature variability over the European continent. The decadal forecasts for time scale of nine years and longer, on the other hand, exhibited moderate skill. The dependency between temporal and spatial scales was found to differ for the temperature and the mean SLP anomalies, but we found little indication that the decadal predictions for anomalies with large-scale regional extent were associated higher skill than for more local patterns in these forecasts.

Process-conditioned bias correction for seasonal forecasting: a case-study with ENSO in Peru

This work assesses the suitability of a first simple attempt for process-conditioned bias correction in the context of seasonal forecasting. To do this, we focus on the northwestern part of Peru and bias correct 1- and 4-month lead seasonal predictions of boreal winter (DJF) precipitation from the ECMWF System4 forecasting system for the period 1981–2010. In order to include information about the underlying large-scale circulation which may help to discriminate between precipitation affected by different processes, we introduce here an empirical quantile–quantile mapping method which runs conditioned on the state of the Southern Oscillation Index (SOI), which is accurately predicted by System4 and is known to affect the local climate. Beyond the reduction of model biases, our results show that the SOI-conditioned method yields better ROC skill scores and reliability than the raw model output over the entire region of study, whereas the standard unconditioned implementation provides no added value for any of these metrics. This suggests that conditioning the bias correction on simple but well-simulated large-scale processes relevant to the local climate may be a suitable approach for seasonal forecasting. Yet, further research on the suitability of the application of similar approaches to the one considered here for other regions, seasons and/or variables is needed.

Improved Seasonal Prediction of Rainfall over East Africa for Application in Agriculture: Statistical Downscaling of CFSv2 and GFDL-FLOR

AbstractStatistically downscaled forecasts of October–December (OND) rainfall are evaluated over East Africa from two general circulation model (GCM) seasonal prediction systems. The method uses canonical correlation analysis to relate variability in predicted large-scale rainfall (characterizing, e.g., predicted ENSO and Indian Ocean dipole variability) to observed local variability over Kenya and Tanzania. Evaluation is performed for the period 1982–2011 and for the real-time forecast for OND 2015, a season when a strong El Niño was active. The seasonal forecast systems used are the National Centers for Environmental Prediction Climate Forecast System, version 2 (CFSv2), and the Geophysical Fluid Dynamics Laboratory Forecast-Oriented Low Ocean Resolution (GFDL-FLOR) version of CM2.5. The Climate Hazards Group Infrared Precipitation with Station Data (CHIRPS) rainfall dataset—a blend of in situ station observations and satellite estimates—was used at 5 km × 5 km resolution over Kenya and Tanzania as benchmark data for the downscaling. Results for the case-study forecast for OND 2015 show that downscaled output from both models adds realistic spatial detail relative to the coarser raw model output—albeit with some overestimation of rainfall that may have been derived from the downscaling procedure introducing a wet response to El Niño more typical of historical cases. Assessment of the downscaled forecasts over the 1982–2011 period shows positive long-term skill better than that documented in previous studies of unprocessed GCM forecasts for the region. Climate forecast downscaling is thus a key undertaking worldwide in the generation of more reliable products for sector specific application including agricultural planning and decision-making.

Assessing the suitability of statistical downscaling approaches for seasonal forecasting in Senegal

AbstractThis work tests the suitability of statistical downscaling (SD) approaches to generate local seasonal forecasts of daily maximum temperature and precipitation for a set of selected stations in Senegal for the July–August–September season during the period 1979–2000. Two‐month lead raw daily maximum temperature and precipitation from the five models included in the ENSEMBLES seasonal hindcast are compared against the corresponding downscaled predictions, which are obtained by applying the analog technique based on two different types of predictors: the direct surface variables and a combination of appropriate upper‐air variables. Beyond correcting the large biases of the low‐resolution raw model outputs, SD is found to add noteworthy value in terms of forecast association (as measured by interannual correlation), providing thus suitable (i.e. calibrated) predictions at the local‐scale needed for practical applications, which means a clear advantage for the end‐users of seasonal forecasts over the area of study. Moreover, a recommendation on the adequacy of surface (large‐scale) predictors for SD of maximum temperature (precipitation) is also given.

CHASE-PL—Future Hydrology Data Set: Projections of Water Balance and Streamflow for the Vistula and Odra Basins, Poland

There is considerable concern that the water resources of Central and Eastern Europe region can be adversely affected by climate change. Projections of future water balance and streamflow conditions can be obtained by forcing hydrological models with the output from climate models. In this study, we employed the SWAT hydrological model driven with an ensemble of nine bias-corrected EURO-CORDEX climate simulations to generate future hydrological projections for the Vistula and Odra basins in two future horizons (2024–2050 and 2074–2100) under two Representative Concentration Pathways (RCPs). The data set consists of three parts: (1) model inputs; (2) raw model outputs; (3) aggregated model outputs. The first one allows the users to reproduce the outputs or to create the new ones. The second one contains the simulated time series of 10 variables simulated by SWAT: precipitation, snow melt, potential evapotranspiration, actual evapotranspiration, soil water content, percolation, surface runoff, baseflow, water yield and streamflow. The third one consists of the multi-model ensemble statistics of the relative changes in mean seasonal and annual variables developed in a GIS format. The data set should be of interest of climate impact scientists, water managers and water-sector policy makers. In any case, it should be noted that projections included in this data set are associated with high uncertainties explained in this data descriptor paper.

Impacts of weighting climate models for hydro-meteorological climate change studies

Weighting climate models is controversial in climate change impact studies using an ensemble of climate simulations from different climate models. In climate science, there is a general consensus that all climate models should be considered as having equal performance or in other words that all projections are equiprobable. On the other hand, in the impacts and adaptation community, many believe that climate models should be weighted based on their ability to better represent various metrics over a reference period. The debate appears to be partly philosophical in nature as few studies have investigated the impact of using weights in projecting future climate changes. The present study focuses on the impact of assigning weights to climate models for hydrological climate change studies. Five methods are used to determine weights on an ensemble of 28 global climate models (GCMs) adapted from the Coupled Model Intercomparison Project Phase 5 (CMIP5) database. Using a hydrological model, streamflows are computed over a reference (1961–1990) and future (2061–2090) periods, with and without post-processing climate model outputs. The impacts of using different weighting schemes for GCM simulations are then analyzed in terms of ensemble mean and uncertainty. The results show that weighting GCMs has a limited impact on both projected future climate in term of precipitation and temperature changes and hydrology in terms of nine different streamflow criteria. These results apply to both raw and post-processed GCM model outputs, thus supporting the view that climate models should be considered equiprobable.

Can bias correction and statistical downscaling methods improve the skill of seasonal precipitation forecasts?

Statistical downscaling methods are popular post-processing tools which are widely used in many sectors to adapt the coarse-resolution biased outputs from global climate simulations to the regional-to-local scale typically required by users. They range from simple and pragmatic Bias Correction (BC) methods, which directly adjust the model outputs of interest (e.g. precipitation) according to the available local observations, to more complex Perfect Prognosis (PP) ones, which indirectly derive local predictions (e.g. precipitation) from appropriate upper-air large-scale model variables (predictors). Statistical downscaling methods have been extensively used and critically assessed in climate change applications; however, their advantages and limitations in seasonal forecasting are not well understood yet. In particular, a key problem in this context is whether they serve to improve the forecast quality/skill of raw model outputs beyond the adjustment of their systematic biases. In this paper we analyze this issue by applying two state-of-the-art BC and two PP methods to downscale precipitation from a multimodel seasonal hindcast in a challenging tropical region, the Philippines. To properly assess the potential added value beyond the reduction of model biases, we consider two validation scores which are not sensitive to changes in the mean (correlation and reliability categories). Our results show that, whereas BC methods maintain or worsen the skill of the raw model forecasts, PP methods can yield significant skill improvement (worsening) in cases for which the large-scale predictor variables considered are better (worse) predicted by the model than precipitation. For instance, PP methods are found to increase (decrease) model reliability in nearly 40% of the stations considered in boreal summer (autumn). Therefore, the choice of a convenient downscaling approach (either BC or PP) depends on the region and the season.