N–PLS regression as empirical downscaling tool in climate change studies

Summary An N–PLS regression technique was tested as an empirical downscaling method. Average monthly near-ground air temperature (t), specific humidity (q), and sea-level pressure (p) fields across Central and Western Europe were used as predictors for average monthly air temperature (T), dew temperature (D), and precipitation amount (P) at 4 locations in Slovenia. The empirical downscaling models (EM) were developed by means of available predictand data from the ARSO archive and predictor data from the NCEP/NCAR reanalysis project, for the period 1951–2002, separately for single months. Using the combination of t and p as predictors, the EM for T explained from 73% to 95% of predictand variability, for D from 74% to 97% of predictand variability, and for P from 31% to 76% of predictand variability. The use of q as an additional predictor did not improve the quality of the EM considerably. Developed EM using p and t as predictors were applied to the results of 5 general circulation models (GCM): CSIRO/Mk2, CCC/CGCM2, UKMO/HadCM3, DOE-NCAR/PCM, and MPI-DMI/ECHAM4-OPYC3. Only the simulations based on SRES A2 and B2 emission scenarios were considered in our calculations. Available mean monthly values of predictors for the period 1951–2100 were used. All the projections of GCM results indicate an increase in T and D and decrease in P in the 21^st century at all 4 locations. The expected range of changes in T, D, and P is wide due to the different response of GCM to identical changes in the atmospheric composition, and represents a source of uncertainty in empirical downscaling results. Another important source of uncertainty in empirical downscaling studies, especially when temperature dependent predictors are used, is the problem of extrapolation. By using the proper mathematical approach for EM development we only reduce a part of the uncertainty related to the quality of empirical models that also strongly depend on the quality of input data and predictor selection. The N–PLS regression seems to be a suitable choice of mathematical method, as the feature selection from a large number of predictor time series is not predictand independent. Finally, any climate change and impact studies for the future are affected by many other uncertainties that we have to be fully aware of, while interpreting their results.     

多模式集合预报及其降尺度技术在东亚夏季降水预测中的应用

利用动力季节模式输出的匹配域投影技术和多模式集合预报技术对多个国家和城市的站点月平均降水进行预报。预报变量是北京1个站、韩国60个站和曼谷地区8个站点的月平均降水,预报因子是从多个业务动力季节预报模式输出的多个大尺度变量。模式回报数据和站点观测降水数据时段是1983—2003年。降尺度预报降水的技巧是在交叉验证的框架下进行的。匹配域投影方法是设定一个可以活动的窗口在全球范围内大尺度场上进行扫描,寻求与目标站点降水最优化的因子和最相关的区域,目标站点的降水变率就是由该匹配域上大尺度环流场信息决定的。最终预报是用多个降尺度模式预报结果的集合预报(DMME)。多个降尺度模式预报结果的集合预报能显著地提高站点降水的预报技巧。北京站,多个降尺度模式预报结果的集合预报的预报和观测降水的相关系数可以提高到0.71;韩国地区,多个降尺度模式预报结果的集合预报平均技巧提高到0.75;泰国,多个降尺度模式预报结果的集合预报技巧是0.61。     

Statistical Downscaling of Pattern Projection Using Multi-Model Output Variables as Predictors

A pattern projection downscaling method is employed to predict monthly station precipitation. The predictand is the monthly precipitation at 1 station in China, 60 stations in Korea, and 8 stations in Thailand. The predictors are multiple variables from the output of operational dynamical models. The hindcast datasets span a period of 21 yr from 1983 to 2003. A downscaled prediction is made for each model separately within a leave-one-out cross-validation framework. The pattern projection method uses a moving window, which scans globally, in order to seek the most optimal predictor for each station. The final forecast is the average of the model downscaled precipitation forecasts using the best predictors and is referred to as DMME. It is found that DMME significantly improves the prediction skill by correcting the erroneous signs of the rainfall anomalies in coarse resolution predictions of general circulation models. The correlation coefficient between the prediction of DMME and the observation in Beijing of China reaches 0.71; the skill is improved to 0.75 for Korea and 0.61 for Thailand. The improvement of the prediction skills for the first two cases is attributed to three steps: coupled pattern selection, optimal predictor selection, and multi-model downscaled precipitation ensemble. For Thailand, we use the single-predictor prediction, which results in a lower prediction skill than the other two cases. This study indicates that the large-scale circulation variables, which are predicted by the current operational dynamical models, if selected well, can be used to make skillful predictions of local precipitation by means of appropriate statistical downscaling.     

Seasonal reversal at Miryang Eoreumgol (Ice Valley), Korea: observation and monitoring

Climate change information required for impact studies is of a much finer scale than that provided by Global circulation models (GCMs). This paper presents an application of partial least squares (PLS) regression for downscaling GCMs output. Statistical downscaling models were developed using PLS regression for simultaneous downscaling of mean monthly maximum and minimum temperatures (T _max and T _min) as well as pan evaporation to lake-basin scale in an arid region in India. The data used for evaluation were extracted from the NCEP/NCAR reanalysis dataset for the period 1948?C2000 and the simulations from the third-generation Canadian Coupled Global Climate Model (CGCM3) for emission scenarios A1B, A2, B1, and COMMIT for the period 2001?C2100. A simple multiplicative shift was used for correcting predictand values. The results demonstrated that the downscaling method was able to capture the relationship between the premises and the response. The analysis of downscaling models reveals that (1) the correlation coefficient for downscaled versus observed mean maximum temperature, mean minimum temperature, and pan evaporation was 0.94, 0.96, and 0.89, respectively; (2) an increasing trend is observed for T _max and T _min for A1B, A2, and B1 scenarios, whereas no trend is discerned with the COMMIT scenario; and (3) there was no trend observed in pan evaporation. In COMMIT scenario, atmospheric CO₂ concentrations are held at year 2000 levels. Furthermore, a comparison with neural network technique shows the efficiency of PLS regression method.     

基于时间尺度分离的中国东部夏季降水预测

基于时间尺度分离,利用NCEP第2代气候预测系统 (CFSv2) 每年4月起报的夏季月平均预测资料, 结合实际观测资料和再分析资料,对江淮流域及华北地区夏季降水距平百分率进行降尺度预测。将预测量和预测因子分为年际分量和年代际分量,在两个时间尺度上分别建立降尺度模型,两个预测分量之和为总预测量。对1982—2008年拟合时段的夏季降水距平百分率的回报结果表明:降尺度预测结果相对于原始模式结果预测技巧显著提高。降尺度预测与实况降水在江淮流域和华北地区的空间相关系数最大值超过0.8,多年平均值也分别提高到0.53和0.51;时间相关在每个站点也显著增强,相关系数为0.38~0.65。对2009—2013年进行独立样本检验,结果表明:降尺度模型能较好地预测出该时段的降水异常空间型态。同时,该模型对2014年夏季降水长江以南偏多、黄淮地区偏少的分布形势也有一定预测能力。     

Statistical downscaling of daily climate variables for climate change impact assessment over New South Wales, Australia

This paper outlines a new statistical downscaling method based on a stochastic weather generator. The monthly climate projections from global climate models (GCMs) are first downscaled to specific sites using an inverse distance-weighted interpolation method. A bias correction procedure is then applied to the monthly GCM values of each site. Daily climate projections for the site are generated by using a stochastic weather generator, WGEN. For downscaling WGEN parameters, historical climate data from 1889 to 2008 are sorted, in an ascending order, into 6 climate groups. The WGEN parameters are downscaled based on the linear and non-linear relationships derived from the 6 groups of historical climates and future GCM projections. The overall averaged confidence intervals for these significant linear relationships between parameters and climate variables are 0.08 and 0.11 (the range of these parameters are up to a value of 1.0) at the observed mean and maximum values of climate variables, revealing a high confidence in extrapolating parameters for downscaling future climate. An evaluation procedure is set up to ensure that the downscaled daily sequences are consistent with monthly GCM output in terms of monthly means or totals. The performance of this model is evaluated through the comparison between the distributions of measured and downscaled climate data. Kruskall-Wallis rank (K-W) and Siegel-Tukey rank sum dispersion (S-T) tests are used. The results show that the method can reproduce the climate statistics at annual, monthly and daily time scales for both training and validation periods. The method is applied to 1062 sites across New South Wales (NSW) for 9 GCMs and three IPCC SRES emission scenarios, B1, A1B and A2, for the period of 1900–2099. Projected climate changes by 7 GCMs are also analyzed for the A2 emission scenario based on the downscaling results.     

夏季地面温度场月预报的统计模式

用15年北半球夏季地面温度的资料分析了月平均场与该月前后一周左右的逐日平均场在时间和波数上的相关。发现越新、越靠后的日平均场与该月的平均场相关越强。用第0天外推的未来30天平均场与实况的相关比用前30天平均场作外推的相关要显著得多。由此得出结论认为各种预报方法的效果检验应与前者定义的惯性预报作比较。文中提出了几个用于作月预报的统计模式，分别用不同时间或不同波数作预报因子。其结果均使预报准确率有所提高。既用不同时间又用不同波数作预报因子的模式更为理想。文中最后讨论了一个源于动力学考虑的统计模式。近1400个月预报例子的结果与实况的平均相关系数高达0.75，显著优于惯性预报。     

High-Frequency and Low-Frequency Variability in Stochastic Daily Weather Generator and Its Effect on Agricultural and Hydrologic Modelling

The high-frequency and low-frequency variabilities, which are often misreproduced by the daily weather generators, have a significant effect on modelling weather-dependent processes. Three modifications are suggested to improve the reproduction of the both variabilities in a four-variate daily weather generator Met&Roll: (i) inclusion of the annual cycle of lag-0 and lag-1 correlations among solar radiation, maximum temperature and minimum temperature, (ii) use of the 3rd order Markov chain to model precipitation occurrence, (iii) applying the monthly generator (based on a first-order autoregressive model) to fit the low-frequency variability. The tests are made to examine the effects of the three new features on (i) a stochastic structure of the synthetic series, and on (ii) outputs from CERES-Wheat crop model (crop yields) and SAC-SMA rainfall-runoff model (monthly streamflow characteristics, distribution of 5-day streamflow) fed by the synthetic weather series. The results are compared with those obtained with the observed weather series.Results: (i) The inclusion of the annual cycle of the correlations has rather ambiguous effect on the temporal structure of the weather characteristics simulated by the generator and only insignificant effect on the output from either simulation model. (ii) Increased order of the Markov chain improves modelling of precipitation occurrence series (especially long dry spells), and correspondingly improves reliability of the output from either simulation model. (iii) Conditioning the daily generator on monthly generator has the most positive effect, especially on the output from the hydrological model: Variability of the monthly streamflow characteristics and the frequency of extreme streamflows are better simulated. (iv) Of the two simulation models, the improvements related to the three modifications are more pronounced in the hydrological simulations. This may be also due to the fact that the crop growth simulations were less affected by the imperfections of the unmodified version of Met&Roll.     

基于广义线性模型和NCEP资料的降水随机发生器

天气发生器可以用来插补历史缺测气象数据或生成未来天气情境, 近年来被普遍应用于对气象变量的降尺度研究, 为陆面的水文、 生态模拟提供外强迫输入。广义线性模型 (GLM) 是近年来用于建立大尺度气象变量与地面气象因子之间的一种有效方法, 基于GLM的天气发生器具有一定的应用前景。本文以NCEP再分析资料中的单格点气温、 500 hPa位势高度、 位温、 相对湿度、 海平面气压等5个变量作为影响降水变化的大尺度因子建立模拟逐日降水量的广义线性模型。模型中对降水概率的描述采用Logistic模型模拟, 而对降水量则分别试用Gamma分布、 指数分布、 正态分布和对数正态分布来模拟, 试图比较和揭示这些基于不同理论分布的模型的能力。模型中待定参数的估计及对研究区逐日降水量的模拟采用了完全相同的实测逐日降水数据和同期NCEP再分析资料。参数的最大似然估计用遗传算法来实现, 对山东省临沂地区10个主要气象观测站降水资料的研究表明, Gamma分布模型的拟合效果最好, 对数正态分布次之, 指数分布再次, 正态分布最差; 参数估计分月获取的拟合效果略好于不分月的。模型逐日降水模拟表明, 对降水发生概率的模拟会低估各月的多年平均值, 基于指数分布的GLM会低估各月总降水量期望 (为月内每日降水量期望之和) 的多年平均值, 而基于对数正态分布的GLM则会在降水量较大时产生高估现象。由对应的天气发生器模型生成的随机模拟降水序列表明, 基于对数正态分布的模型会高估月降水量较大时的多年平均, 而基于指数分布及Gamma分布的模型则模拟效果较好。总体上看, 这种基于NCEP再分析资料和GLM的天气发生器对降水变率具有很强的解释和模拟能力。     

Daily relative humidity projections in an Indian river basin for IPCC SRES scenarios

A two-stage methodology is developed to obtain future projections of daily relative humidity in a river basin for climate change scenarios. In the first stage, Support Vector Machine (SVM) models are developed to downscale nine sets of predictor variables (large-scale atmospheric variables) for Intergovernmental Panel on Climate Change Special Report on Emissions Scenarios (SRES) (A1B, A2, B1, and COMMIT) to R _H in a river basin at monthly scale. Uncertainty in the future projections of R _H is studied for combinations of SRES scenarios, and predictors selected. Subsequently, in the second stage, the monthly sequences of R _H are disaggregated to daily scale using k-nearest neighbor method. The effectiveness of the developed methodology is demonstrated through application to the catchment of Malaprabha reservoir in India. For downscaling, the probable predictor variables are extracted from the (1) National Centers for Environmental Prediction reanalysis data set for the period 1978–2000 and (2) simulations of the third-generation Canadian Coupled Global Climate Model for the period 1978–2100. The performance of the downscaling and disaggregation models is evaluated by split sample validation. Results show that among the SVM models, the model developed using predictors pertaining to only land location performed better. The R _H is projected to increase in the future for A1B and A2 scenarios, while no trend is discerned for B1 and COMMIT.     

统计降尺度法对华北地区未来区域气温变化情景的预估

迄今为止，大部分海气耦合气候模式（AOGCM）的空间分辨率还较低，很难对区域尺度的气候变化情景做合理的预测。降尺度法已广泛用于弥补AOGCM在这方面的不足。作者采用统计降尺度方法对1月和7月华北地区49个气象观测站的未来月平均温度变化情景进行预估。采用的统计降尺度方法是主分量分析与逐步回归分析相结合的多元线性回归模型。首先，采用1961～2000年的 NCEP再分析资料和49个台站的观测资料建立月平均温度的统计降尺度模型，然后把建立的统计降尺度模型应用于HadCM3 SRES A2 和 B2 两种排放情景， 从而生成各个台站1950～2099年1月份和7月份温度变化情景。结果表明:在当前气候条件下，无论1月还是7月，统计降尺度方法模拟的温度与观测的温度有很好的一致性，而且在大多数台站，统计降尺度模拟气温与观测值相比略微偏低。对于未来气候情景的预估方面，无论1月还是7月，也无论是HadCM3 SRES A2 还是B2排放情景驱动统计模型，结果表明大多数的站点都存在温度的明显上升趋势，同时7月的上升趋势与1月相比偏低。     

Probabilistic estimates of future changes in California temperature and precipitation using statistical and dynamical downscaling

Sixteen global general circulation models were used to develop probabilistic projections of temperature (T) and precipitation (P) changes over California by the 2060s. The global models were downscaled with two statistical techniques and three nested dynamical regional climate models, although not all global models were downscaled with all techniques. Both monthly and daily timescale changes in T and P are addressed, the latter being important for a range of applications in energy use, water management, and agriculture. The T changes tend to agree more across downscaling techniques than the P changes. Year-to-year natural internal climate variability is roughly of similar magnitude to the projected T changes. In the monthly average, July temperatures shift enough that that the hottest July found in any simulation over the historical period becomes a modestly cool July in the future period. Januarys as cold as any found in the historical period are still found in the 2060s, but the median and maximum monthly average temperatures increase notably. Annual and seasonal P changes are small compared to interannual or intermodel variability. However, the annual change is composed of seasonally varying changes that are themselves much larger, but tend to cancel in the annual mean. Winters show modestly wetter conditions in the North of the state, while spring and autumn show less precipitation. The dynamical downscaling techniques project increasing precipitation in the Southeastern part of the state, which is influenced by the North American monsoon, a feature that is not captured by the statistical downscaling.     

Statistical downscaling of regional daily precipitation over southeast Australia based on self-organizing maps

This paper presents a novel statistical downscaling method based on a non-linear classification technique known as self-organizing maps (SOMs) and has therefore been named SOM-SD. The relationship between large-scale atmospheric circulation and local-scale surface variable was constructed in a relatively simple and transparent manner. For a specific atmospheric state, an ensemble of possible values was generated for the predictand following the Monte Carlo method. Such a stochastic simulation is essential to explore the uncertainties of climate change in the future through a series of random re-sampling experiments. The novel downscaling method was evaluated by downscaling daily precipitation over Southeast Australia. The large-scale predictors were extracted from the daily NCAR/NCEP reanalysis data, while the predictand was high-resolution gridded daily observed precipitation (1958?C2008) from the Australian Bureau of Meteorology. The results showed that the method works reasonably well across a variety of climatic zones in the study area. Overall, there was no particular zone that stands out as a climatic entity where the downscaling skill in reproducing all statistical indices was consistently lower or higher across seasons than the other zones. The method displayed a high skill in reproducing not only the climatologic statistical properties of the observed precipitation, but also the characteristics of the extreme precipitation events. Furthermore, the model was able to reproduce, to a certain extent, the inter-annual variability of precipitation characteristics.     

Hydrologic Implications of Dynamical and Statistical Approaches to Downscaling Climate Model Outputs

Six approaches for downscaling climate model outputs for use in hydrologic simulation were evaluated, with particular emphasis on each method's ability to produce precipitation and other variables used to drive a macroscale hydrology model applied at much higher spatial resolution than the climate model. Comparisons were made on the basis of a twenty-year retrospective (1975–1995) climate simulation produced by the NCAR-DOE Parallel ClimateModel (PCM), and the implications of the comparison for a future(2040–2060) PCM climate scenario were also explored. The six approaches were made up of three relatively simple statistical downscaling methods – linear interpolation (LI), spatial disaggregation (SD), and bias-correction and spatial disaggregation (BCSD) – each applied to both PCM output directly(at T42 spatial resolution), and after dynamical downscaling via a Regional Climate Model (RCM – at 1/2-degree spatial resolution), for downscaling the climate model outputs to the 1/8-degree spatial resolution of the hydrological model. For the retrospective climate simulation, results were compared to an observed gridded climatology of temperature and precipitation, and gridded hydrologic variables resulting from forcing the hydrologic model with observations. The most significant findings are that the BCSD method was successful in reproducing the main features of the observed hydrometeorology from the retrospective climate simulation, when applied to both PCM and RCM outputs. Linear interpolation produced better results using RCM output than PCM output, but both methods (PCM-LI and RCM-LI) lead to unacceptably biased hydrologic simulations. Spatial disaggregation of the PCM output produced results similar to those achieved with the RCM interpolated output; nonetheless, neither PCM nor RCM output was useful for hydrologic simulation purposes without a bias-correction step. For the future climate scenario, only the BCSD-method (using PCM or RCM) was able to produce hydrologically plausible results. With the BCSD method, the RCM-derived hydrology was more sensitive to climate change than the PCM-derived hydrology.     

华东地区月平均气温统计降尺度方法比较

用中国地面气象观测站的逐日气温观测资料和NCEP/NCAR再分析资料,分别使用基于多元线性回归(MLR)和3种主成分分析(PCA)的统计降尺度方法,对1959-2008年的华东地区的月平均气温分两个时段进行统计降尺度分析并加以检验,比较了不同降尺度方法的结果。结果表明:对于华东地区气温的统计降尺度预报,基于MLR的统计降尺度方法相对于3种PCA方法而言,对单站年际变化模拟方面有一定优势。PCA方法应用于统计降尺度时,预报因子的区域选择是影响统计降尺度结果的重要因素之一。对于温度进行统计降尺度分析时,预报因子中包含温度因子是非常必要的;所试验的4种降尺度方法,对各站点多年平均情况的模拟要好于对区域平均的年际变化的模拟。     

A Comparison of Four Precipitation Distribution Models Used in Daily Stochastic Models

Stochastic weather generators are statistical models that produce random numbers that resemble the observed weather data on which they have been fitted; they are widely used in meteorological and hydrologi- cal simulations. For modeling daily precipitation in weather generators, first-order Markov chain-dependent exponential, gamma, mixed-exponential, and lognormal distributions can be used. To examine the perfor- mance of these four distributions for precipitation simulation, they were fitted to observed data collected at 10 stations in the watershed of Yishu River. The parameters of these models were estimated using a maximum-likelihood technique performed using genetic algorithms. Parameters for each calendar month and the Fourier series describing parameters for the whole year were estimated separately. Bayesian infor- mation criterion, simulated monthly mean, maximum daily value, and variance were tested and compared to evaluate the fitness and performance of these models. The results indicate that the lognormal and mixed-exponential distributions give smaller BICs, but their stochastic simulations have overestimation and underestimation respectively, while the gamma and exponential distributions give larger BICs, but their stochastic simulations produced monthly mean precipitation very well. When these distributions were fitted using Fourier series, they all underestimated the above statistics for the months of June, July and August.     

Statistical downscaling of historical monthly mean winds over a coastal region of complex terrain. II. Predicting wind components

A regression-based downscaling technique was applied to monthly mean surface wind observations from stations throughout western Canada as well as from buoys in the Northeast Pacific Ocean over the period 1979–2006. A predictor set was developed from principal component analysis of the three wind components at 500?hPa and mean sea-level pressure taken from the NCEP Reanalysis II. Building on the results of a companion paper, Curry et al. (Clim Dyn 2011, doi:10.1007/s00382-011-1173-3), the downscaling was applied to both wind speed and wind components, in an effort to evaluate the utility of each type of predictand. Cross-validated prediction skill varied strongly with season, with autumn and summer displaying the highest and lowest skill, respectively. In most cases wind components were predicted with better skill than wind speeds. The predictive ability of wind components was found to be strongly related to their orientation. Wind components with the best predictions were often oriented along topographically significant features such as constricted valleys, mountain ranges or ocean channels. This influence of directionality on predictive ability is most prominent during autumn and winter at inland sites with complex topography. Stations in regions with relatively flat terrain (where topographic steering is minimal) exhibit inter-station consistencies including region-wide seasonal shifts in the direction of the best predicted wind component. The conclusion that wind components can be skillfully predicted only over a limited range of directions at most stations limits the scope of statistically downscaled wind speed predictions. It seems likely that such limitations apply to other regions of complex terrain as well.     

时间尺度分离在华南夏季极端高温预测中的应用

基于高温日数存在受不同物理因子影响不同时间尺度变率的特征,应用滤波对华南夏季高温日数进行时间尺度分离,得到高温日数的年代际分量和年际分量。统计分析高温日数总量、年代际分量和年际分量在各自对应时间尺度上的影响因子,采用"向前"交叉检验逐步回归法,分别建立高温日数总量、年代际分量和年际分量的回归模型。高温日数总量的回归模型即为高温日数不区分时间尺度的直接回归模型,而两个分量回归模型拟合结果的叠加,即为高温日数时间尺度分离统计模型对总量的拟合。利用十折交叉检验法,对高温日数直接回归模型和时间尺度分离统计模型的拟合结果进行比较:相比高温日数直接回归模型,时间尺度分离统计模型的年代际分量均方根误差由2.6降低到2.3,与观测数据的相关系数由0.69提高到0.73（显著性水平α=0.01）;年际分量均方根误差由3.2降低到2.9,与观测数据的相关系数由0.4（α=0.1）提高到0.48（α=0.01）;高温日数总量均方根误差由4.1降低到3.7,与观测数据的相关系数由0.48提高到0.62（α=0.01）。1979~2010年拟合时段华南夏季高温日数的回报结果表明:两模型回报结果与观测数据均存在明显相关（α=0.01）,直接回归模型的相关系数为0.57,时间尺度分离统计模型提高到0.72。2011~2013年独立检验时段的预测结果表明:直接回归模型预测结果的平均均方根误差为26.4%,时间尺度分离统计模型降低到12.3%。初步结果表明,两模型对华南夏季高温日数均有一定的预测能力,而时间尺度分离统计模型的预测结果有所改进。     

中国气温的降尺度研究

Monthly mean temperatures at 562 stations in China are estimated using a statistical downscaling technique. The technique used is multiple linear regressions （MLRs） of principal components （PCs）. A stepwise screening procedure is used for selecting the skilful PCs as predictors used in the regression equation. The predictors include temperature at 850 hPa （7）, the combination of sea-level pressure and temperature at 850 hPa （P＋T） and the combination of geo-potential height and temperature at 850 hPa （H＋T）. The downscaling procedure is tested with the three predictors over three predictor domains. The optimum statistical model is obtained for each station and month by finding the predictor and predictor domain corresponding to the highest correlation. Finally, the optimum statistical downscaling models are applied to the Hadley Centre Coupled Model, version 3 （HadCM3） outputs under the Special Report on Emission Scenarios （SRES） A2 and B2 scenarios to construct local future temperature change scenarios for each station and month, The results show that （1） statistical downscaling produces less warming than the HadCM3 output itself; （2） the downscaled annual cycles of temperature differ from the HadCM3 output, but are similar to the observation; （3） the downscaled temperature scenarios show more warming in the north than in the south; （4） the downscaled temperature scenarios vary with emission scenarios, and the A2 scenario produces more warming than the B2, especially in the north of China.     

Evaluation of machine learning tools as a statistical downscaling tool: temperatures projections for multi-stations for Thames River Basin, Canada

Many impact studies require climate change information at a finer resolution than that provided by global climate models (GCMs). This paper investigates the performances of existing state-of-the-art rule induction and tree algorithms, namely single conjunctive rule learner, decision table, M5 model tree, and REPTree, and explores the impact of climate change on maximum and minimum temperatures (i.e., predictands) of 14 meteorological stations in the Upper Thames River Basin, Ontario, Canada. The data used for evaluation were large-scale predictor variables, extracted from National Centers for Environmental Prediction/National Center for Atmospheric Research reanalysis dataset and the simulations from third generation Canadian coupled global climate model. Data for four grid points covering the study region were used for developing the downscaling model. M5 model tree algorithm was found to yield better performance among all other learning techniques explored in the present study. Hence, this technique was applied to project predictands generated from GCM using three scenarios (A1B, A2, and B1) for the periods (2046–2065 and 2081–2100). A simple multiplicative shift was used for correcting predictand values. The potential of the downscaling models in simulating predictands was evaluated, and downscaling results reveal that the proposed downscaling model can reproduce local daily predictands from large-scale weather variables. Trend of projected maximum and minimum temperatures was studied for historical as well as downscaled values using GCM and scenario uncertainty. There is likely an increasing trend for T _max and T _min for A1B, A2, and B1 scenarios while decreasing trend has been observed for B1 scenarios during 2081–2100.