Multi-site Statistical Downscaling Research Articles

Convolutional conditional neural processes for local climate downscaling

Abstract. A new model is presented for multisite statistical downscaling of temperature and precipitation using convolutional conditional neural processes (convCNPs). ConvCNPs are a recently developed class of models that allow deep-learning techniques to be applied to off-the-grid spatio-temporal data. In contrast to existing methods that map from low-resolution model output to high-resolution predictions at a discrete set of locations, this model outputs a stochastic process that can be queried at an arbitrary latitude–longitude coordinate. The convCNP model is shown to outperform an ensemble of existing downscaling techniques over Europe for both temperature and precipitation taken from the VALUE intercomparison project. The model also outperforms an approach that uses Gaussian processes to interpolate single-site downscaling models at unseen locations. Importantly, substantial improvement is seen in the representation of extreme precipitation events. These results indicate that the convCNP is a robust downscaling model suitable for generating localised projections for use in climate impact studies.

Comparison of CMIP5 and CMIP6 Multi-Model Ensemble for Precipitation Downscaling Results and Observational Data: The Case of Hanjiang River Basin

Evaluating global climate model (GCM) outputs is essential for accurately simulating future hydrological cycles using hydrological models. The GCM multi-model ensemble (MME) precipitation simulations of the Climate Model Intercomparison Project Phases 5 and 6 (CMIP5 and CMIP6, respectively) were spatially and temporally downscaled according to a multi-site statistical downscaling method for the Hanjiang River Basin (HRB), China. Downscaled precipitation accuracy was assessed using data collected from 14 meteorological stations in the HRB. The spatial performances, temporal performances, and seasonal variations of the downscaled CMIP5-MME and CMIP6-MME were evaluated and compared with observed data from 1970–2005. We found that the multi-site downscaling method accurately downscaled the CMIP5-MME and CMIP6-MME precipitation simulations. The downscaled precipitation of CMIP5-MME and CMIP6-MME captured the spatial pattern, temporal pattern, and seasonal variations; however, precipitation was slightly overestimated in the western and central HRB and precipitation was underestimated in the eastern HRB. The precipitation simulation ability of the downscaled CMIP6-MME relative to the downscaled CMIP5-MME improved because of reduced biases. The downscaled CMIP6-MME better simulated precipitation for most stations compared to the downscaled CMIP5-MME in all seasons except for summer. Both the downscaled CMIP5-MME and CMIP6-MME exhibit poor performance in simulating rainy days in the HRB.

Study of climate change effects on hydrological processes in Siminehroud and Zarrinehroud watersheds northwest of Iran

This study sets out to simulate various hydrological responses to climate change. The semi-distributed hydrologic model of the Soil and Water Assessment Tool, SWAT, was used to simulate different parts of the hydrological cycles. The comprehensive assessment of the effect of climate change on runoff, crop yield, and water balance in Siminehroud and Zarrinehroud watersheds is the novelty of this research. The simulation period was from 1988 to 2014. The runoff model was calibrated using the plant parameters, and the initial crop yield was used to calibrate the model. During the calibration and validation periods, the statistical measures, namely NS and R2, were obtained as 0.69 and 0.82, respectively. Using the multisite statistical downscaling of the LARS-WG climate model, we introduced the future climatic conditions as inputs to the model based on two optimistic (RCP2.6) and pessimistic (RCP8.5) scenarios. The most significant changes in the runoff in the upcoming period, was in June, reduced by 4.2 m3s−1, and in May, increased by 4.8 m3s−1 in the optimistic scenario. In the RCP8.5 pessimistic scenario, the most significant runoff change was observed in June and October. The results indicated that these changes would lead to a decrease in the main crops across the region.

Multi-site statistical downscaling of precipitation using generalized hierarchical linear models: a case study of the imperilled Lake Urmia basin

ABSTRACT A downscaling model capable of explaining the temporal and spatial variability of regional hydroclimatic variables is essential for reliable climate change studies and impact assessments. This study proposes a novel statistical approach based on generalized hierarchical linear model (GHLM) to downscale precipitation from the outputs of general circulation models (GCMs) at multiple sites. GHLM partitions the total variance of precipitation into within- and between-site variability allowing for transferring information between sites to develop a regional downscaling model. The methodology is demonstrated by downscaling precipitation using the outputs of eight GCMs in Lake Urmia basin in northwestern Iran. Multi-model ensemble simulations are merged and bias-corrected using Bayesian model averaging and equidistant quantile mapping, respectively. The results of this study show projected declining trends in precipitation resulting in approximately 11.2% and 15.3% decrease during 2060–2080 compared to the historical period of 1985–2005 considering representative concentration pathways (RCPs) 4.5 and 8.5, respectively.

A perfect prognosis approach for daily precipitation series in consideration of space–time correlation structure

Downscaling techniques are the required tools to link the global climate model outputs provided at a coarse grid resolution to finer scale surface variables appropriate for climate change impact studies. Besides the at-site temporal persistence, the downscaled variables have to satisfy the spatial dependence naturally observed between the climate variables at different locations. Furthermore, the precipitation spatial intermittency should be fulfilled. Because of the complexity in describing these properties, they are often ignored, which can affect the effectiveness of the hydrologic process modeling. This study is a continuation of the work by Khalili and Nguyen (Clim Dyn 49(7–8):2261–2278. https://doi.org/10.1007/s00382-016-3443-6 , 2017) regarding the multi-site statistical downscaling of daily precipitation series. Different approach of multi-site statistical downscaling based on the concept of the spatial autocorrelation is presented in this paper. This approach has proven to give effective results for multi-site multivariate statistical downscaling of daily extreme temperature time series (Khalili et al. in Int J Climatol 33:15–32. https://doi.org/10.1002/joc.3402 , 2013). However, more challenges are presented by the precipitation variable because of the high spatio-temporal variability and intermittency. The proposed approach consists of logistic and multiple regression models, linking the global climate predictors to the precipitation occurrences and amounts respectively, and using the spatial autocorrelation concept to reproduce the spatial dependence observed between the precipitation series at different sites. An empirical technique has also been involved in this approach in order to fulfill the precipitation intermittency property. The proposed approach was performed using observed daily precipitation data from ten weather stations located in the southwest region of Quebec and southeast region of Ontario in Canada, and climate predictors from the NCEP/NCAR (National Centers for Environmental Prediction/National Centre for Atmospheric Research) reanalysis dataset. The results have proven the ability of the proposed approach to adequately reproduce the observed precipitation occurrence and amount characteristics, temporal and spatial dependence, spatial intermittency and temporal variability.

Use of beta regression for statistical downscaling of precipitation in the Campbell River basin, British Columbia, Canada

Summary Impacts of global climate change on water resources systems are assessed by downscaling coarse scale climate variables into regional scale hydro-climate variables. In this study, a new multisite statistical downscaling method based on beta regression (BR) is developed for generating synthetic precipitation series, which can preserve temporal and spatial dependence along with other historical statistics. The beta regression based downscaling method includes two main steps: (1) prediction of precipitation states for the study area using classification and regression trees, and (2) generation of precipitation at different stations in the study area conditioned on the precipitation states. Daily precipitation data for 53 years from the ANUSPLIN data set is used to predict precipitation states of the study area where predictor variables are extracted from the NCEP/NCAR reanalysis data set for the same interval. The proposed model is applied to downscaling daily precipitation at ten different stations in the Campbell River basin, British Columbia, Canada. Results show that the proposed downscaling model can capture spatial and temporal variability of local precipitation very well at various locations. The performance of the model is compared with a recently developed non-parametric kernel regression based downscaling model. The BR model performs better regarding extrapolation compared to the non-parametric kernel regression model. Future precipitation changes under different GHG (greenhouse gas) emission scenarios also projected with the developed downscaling model that reveals a significant amount of changes in future seasonal precipitation and number of wet days in the river basin.

Multisite multivariate modeling of daily precipitation and temperature in the Canadian Prairie Provinces using generalized linear models

Based on the Generalized Linear Model (GLM) framework, a multisite stochastic modelling approach is developed using daily observations of precipitation and minimum and maximum temperatures from 120 sites located across the Canadian Prairie Provinces: Alberta, Saskatchewan and Manitoba. Temperature is modeled using a two-stage normal-heteroscedastic model by fitting mean and variance components separately. Likewise, precipitation occurrence and conditional precipitation intensity processes are modeled separately. The relationship between precipitation and temperature is accounted for by using transformations of precipitation as covariates to predict temperature fields. Large scale atmospheric covariates from the National Center for Environmental Prediction Reanalysis-I, teleconnection indices, geographical site attributes, and observed precipitation and temperature records are used to calibrate these models for the 1971–2000 period. Validation of the developed models is performed on both pre- and post-calibration period data. Results of the study indicate that the developed models are able to capture spatiotemporal characteristics of observed precipitation and temperature fields, such as inter-site and inter-variable correlation structure, and systematic regional variations present in observed sequences. A number of simulated weather statistics ranging from seasonal means to characteristics of temperature and precipitation extremes and some of the commonly used climate indices are also found to be in close agreement with those derived from observed data. This GLM-based modelling approach will be developed further for multisite statistical downscaling of Global Climate Model outputs to explore climate variability and change in this region of Canada.

Probabilistic Multisite Statistical Downscaling for Daily Precipitation Using a Bernoulli–Generalized Pareto Multivariate Autoregressive Model

Abstract A Bernoulli–generalized Pareto multivariate autoregressive (BMAR) model is proposed in this paper for multisite statistical downscaling of daily precipitation. The proposed model relies on a probabilistic framework to describe the conditional probability density function of precipitation at each station for a given day and handles multivariate dependence in both time and space using a multivariate autoregressive model. Within a probabilistic framework, BMAR employs a regression model whose outputs are parameters of the mixed Bernoulli–generalized Pareto distribution. As a stochastic component, the BMAR employs a latent multivariate autoregressive Gaussian field to preserve lag-0 and lag-1 cross correlations of precipitation at multiple sites. The proposed model is applied for the downscaling of AOGCM data to daily precipitation in the southern part of Québec, Canada. Reanalysis products are used in this study to assess the potential of the proposed method. Based on the mean errors (MEs), the root-mean-square errors (RMSEs), precipitation indices, and the ability to preserve lag-0 and lag-1 cross correlation, results of the study indicate the superiority of the proposed model over a multivariate multiple linear regression (MMLR) model and a multisite hybrid statistical downscaling procedure that combines MMLR and stochastic generator schemes.

Probabilistic Gaussian Copula Regression Model for Multisite and Multivariable Downscaling

Abstract Atmosphere–ocean general circulation models (AOGCMs) are useful to simulate large-scale climate evolutions. However, AOGCM data resolution is too coarse for regional and local climate studies. Downscaling techniques have been developed to refine AOGCM data and provide information at more relevant scales. Among a wide range of available approaches, regression-based methods are commonly used for downscaling AOGCM data. When several variables are considered at multiple sites, regression models are employed to reproduce the observed climate characteristics at small scale, such as the variability and the relationship between sites and variables. This study introduces a probabilistic Gaussian copula regression (PGCR) model for simultaneously downscaling multiple variables at several sites. The proposed PGCR model relies on a probabilistic framework to specify the marginal distribution for each downscaled variable at a given day through AOGCM predictors, and handles multivariate dependence between sites and variables using a Gaussian copula. The proposed model is applied for the downscaling of AOGCM data to daily precipitation and minimum and maximum temperatures in the southern part of Quebec, Canada. Reanalysis products are used in this study to assess the potential of the proposed method. Results of the study indicate the superiority of the proposed model over classical regression-based methods and a multivariate multisite statistical downscaling model.

Evaluation of downscaled daily rainfall hindcasts over Sydney, Australia using statistical and dynamical downscaling approaches

The assessment of local and regional impacts of climate change often requires downscaling of general circulation model (GCM) projections from coarser GCM-scale to finer local- or catchment-scale spatial resolution. This paper provides an assessment of two downscaling approaches for simulation of daily rainfall over Sydney, Australia. The two downscaling alternatives compared include a multivariate multisite statistical downscaling model based on semi-parametric conditional simulation and a dynamical downscaling approach that uses the National Center for Atmospheric Research (NCAR) weather research and forecasting (WRF) model. The two approaches are evaluated for their ability to reproduce important at-site rainfall statistics at a network of 45 raingauge stations and regional statistics over the catchment area of the Warragamba Dam (9,050 km2). The results indicate that the simulations from these approaches capture many regionally observed climate features, including the simulated seasonal and annual means and daily extreme rainfall values. Further analyses suggest that the statistical downscaling approach provides improved simulations of attributes related to point rainfall, spell lengths and amounts, whereas the dynamical approach is well-suited for applications where regionally averaged rainfall is of primary concern.

Projection of multi-site daily temperatures over the Montréal area, Canada

This study presents a post-adjustment procedure for a multivariate multi-site statistical downscaling model (MMSDM) which can simultaneously downscale multiple predictands at multiple observation sites by combining multivariate multiple linear regression and the stochastic randomization procedure. In the post-adjustment procedure, bias and determinant adjustment factors correct the systematic bias on the downscaled series using atmosphere-ocean coupled global climate model (AOGCM) predictors, and prevent the propagation of systematic error to the projected future predictands. The MMSDM with the post-adjustment procedure is applied to project a realistic series of 2 predictands (daily Tmax and Tmin) for 10 observation sites in the region of Montreal (southern Quebec, Canada). The Canadian CGCM3 reference outputs (1961-2000) and future outputs under the A1B and A2 SRES scenarios (2061-2100) were employed as AOGCM predictors. On average over the 10 observation sites, the monthly means of the daily Tmax and Tmin were increased by 2.0-4.7 and 2.7-5.4°C while seasonal 90th percentile of daily Tmax and 10th percentile of daily Tmin (Tmax90 and Tmin10) were increased by 2.1-4.5 and 2.7-5.8°C for the A1B and A2 scenarios with the MMSDM, respectively. Future Tmax and Tmin series showed higher increases in winter than in the other seasons, as anticipated from AOGCMs or regional climate models over the same area. The average diurnal temperature ranges of future series suggest small increases in spring and autumn. Finally, the projected series yielded frost seasons that are 23 and 28 d shorter, whereas 23 and 27 more days are projected for the length of the growing season than in the present-day climate series.

A nonparametric kernel regression model for downscaling multisite daily precipitation in the Mahanadi basin

Key PointsMulti‐site statistical downscaling to preserve spatio‐temporal variabilitySimulations with kernel regressionClimate Change Impacts on rainfall in Mahanadi Basin, India

A multivariate multi-site statistical downscaling model for daily maximum and minimum temperatures

A multivariate multi-site statistical downscaling model (MMSDM) was developed for simultaneous downscaling of climate variables including daily maximum and minimum tempera- tures (Tmax and Tmin) for multiple observation sites. The MMSDM employs multivariate multiple linear regression (MMLR) to simulate deterministic series from large-scale reanalysis data and adds spatially correlated random series to the deterministic series of the MMLR to complement the underestimated variance and to reproduce a spatial correlation of Tmax and Tmin from multiple sites and an at-site temporal correlation between Tmax and Tmin. The MMSDM model is called MMLRc. The downscaled results of the MMLRc were compared to those of MMLR without random noise (MMLRn) and MMLR with uncorrelated random noise (MMLRi) over the southern Quebec area of Canada. The MMLRc almost exactly reproduced the cross-site correlation of Tmax and Tmin among multiple observation sites, and it accurately reproduced the at-site temporal correlation between Tmax and Tmin at each observation site. The MMLRi and MMLRc reproduced monthly standard deviations of daily Tmax and Tmin, the 90th percentile of Tmax (Tmax90), the 10th percentile of Tmin (Tmin10), and the frost and thaw cycle (Fr-Th) more accurately than the MMLRn model. However, both MMLRc and MMLRi yielded a larger standard error for the monthly mean of daily Tmax and daily Tmin, frost season length (FSL), and growing season length (GSL). For the Fr-Th and diurnal temperature range, the MMLRc performed better than the MMLRn and MMLRi. We conclude that the MMLRn may serve as an alternative to downscaling deterministic signals of a predictand, con- sistent with global climate model predictors, and it may serve to project the averaged central ten- dency of a predictand. The MMLRc, however, is recommended for reproduction of variance, extreme events, and the inter-annual variability of the predictands.

Projection of future daily precipitation series and extreme events by using a multi-site statistical downscaling model over the great Montréal area, Québec, Canada

This study suggested strategies to project future precipitation series based on a multi-site hybrid SDM (statistical downscaling model), which can downscale precipitation series at multiple observation sites simultaneously by combining the multivariate multiple linear regression (MMLR) model and the stochastic randomization procedure. The hybrid SDM and future projection methodologies applied to 10 observation sites located in the great area of Montréal, Québec, Canada. Six future independent precipitation series were projected from six sets of future atmospheric predictors using three AOGCMs (Atmosphere-Ocean Global Climate Models, i.e. CGCM2, CGCM3, HadCM3) and three IPCC SRES emission scenarios (B2, A1B and A2). Downscaled climate change signals on wet/dry sequences and extreme indices of precipitation time series were evaluated over the future period from 2060 to 2099 with respect to the historical period from 1961 to 2000. The future scenarios of all three AOGCMs showed a consistent increase of 7.9–44.6% in winter while only those of HadCM3 and CGCM3 showed a decrease of 2.3–23.0% in summer compared to their historical values. Precipitation series of CGCM2 A2 and CGCM3 A2 scenarios yielded the largest increase in winter, while those of HadCM3 B2 and A2 scenarios yielded the largest decrease in summer for all statistics indices.

A comparison of three multi-site statistical downscaling models for daily rainfall in the North China Plain

Three statistical downscaling methods (conditional resampling statistical downscaling model: CR-SDSM, the generalised linear model for daily climate time series: GLIMCLIM, and the non-homogeneous hidden Markov model: NHMM) for multi-site daily rainfall were evaluated and compared in the North China Plain (NCP). The comparison focused on a range of statistics important for hydrological studies including rainfall amount, extreme rainfall, intra-annual variability, and spatial coherency. The results showed that no single model performed well over all statistics/timescales, suggesting that the user should chose appropriate methods after assessing their advantages and limitations when applying downscaling methods for particular purposes. Specifically, the CR-SDSM provided relatively robust results for annual/monthly statistics and extreme characteristics, but exhibited weakness for some daily statistics, such as daily rainfall amount, dry-spell length, and annual wet/dry days. GLIMCLIM performed well for annual dry/wet days, dry/wet spell length, and spatial coherency, but slightly overestimated the daily rainfall. Additionally, NHMM performed better for daily rainfall and annual wet/dry days, but slightly underestimated dry/wet spell length and overestimated the daily extremes. The results of this study could be applied when investigating climate change impact on hydrology and water availability for the NCP, which suffers from intense water shortages due to climate change and human activities in recent years.

Multisite statistical downscaling model for daily precipitation combined by multivariate multiple linear regression and stochastic weather generator

This study provides a multi-site hybrid statistical downscaling procedure combining regression-based and stochastic weather generation approaches for multisite simulation of daily precipitation. In the hybrid model, the multivariate multiple linear regression (MMLR) is employed for simultaneous downscaling of deterministic series of daily precipitation occurrence and amount using large-scale reanalysis predictors over nine different observed stations in southern Quebec (Canada). The multivariate normal distribution, the first-order Markov chain model, and the probability distribution mapping technique are employed for reproducing temporal variability and spatial dependency on the multisite observations of precipitation series. The regression-based MMLR model explained 16 % ~ 22 % of total variance in daily precipitation occurrence series and 13 % ~ 25 % of total variance in daily precipitation amount series of the nine observation sites. Moreover, it constantly over-represented the spatial dependency of daily precipitation occurrence and amount. In generating daily precipitation, the hybrid model showed good temporal reproduction ability for number of wet days, cross-site correlation, and probabilities of consecutive wet days, and maximum 3-days precipitation total amount for all observation sites. However, the reproducing ability of the hybrid model for spatio-temporal variations can be improved, i.e. to further increase the explained variance of the observed precipitation series, as for example by using regional-scale predictors in the MMLR model. However, in all downscaling precipitation results, the hybrid model benefits from the stochastic weather generator procedure with respect to the single use of deterministic component in the MMLR model.