Statistical downscaling of daily maximum and minimum temperatures in the Central Pampas region

BETTOLLI ML, PENALBA OC, RAGGIO GA

Lima

Workshop; WCRP VAMOS/CORDEX Workshop on Latin-America and Caribbean CORDEX LAC: Phase I - South America; 2013

Empirical downscaling is a widely used technique for exploring the regional and local-scaleresponse to large-scale circulation. Different statistical techniques have been explored inmany regions of the world. However, statistical downscaling (SD) has not been extensivelydeveloped for South America regions.The La Plata Basin in Southeastern South America, is the second most important basin of thecontinent after the Amazon basin. Its morphologic and climatic variety has generated adiversity of interests related to its hydrological resources. The central Pampas is extendingitself towards the south of the basin and is the most important region of agriculturalproduction in Argentina. Therefore, the exploration and development of statisticaldownscaling techniques are of special interest for the region.The purpose of this work is to downscale daily maximum and minimum temperatures at localscales in the Central Pampas using the analogue statistical downscaling method. To this end,daily mean fields of the NCEP-NCAR Reanalysis 1 in the area bounded by 60°S, 15°S, 90°Wand 44.5°W were used as predictor variables. Daily maximum (Tmax) and minimum (Tmin)temperatures from 25 meteorological stations located in the Central Pampas region were alsoused for the period 1979-2000. This information was provided by the Argentina’s NationalWeather Service. The statistical downscaling is based on the analogue method which consistsin finding, for each day in the record, the most similar days (called analogue fields). Thissimilarity between the fields was assessed by the Euclidean distance. For each day, thememory effect was filtered (in a window of ± 5 days) and the first 10 analogue fields wereretained. With the surface information (Tmax and Tmin) of each of these 10 analogues, fivedifferent types of estimations were performed: 1) the average, 2) the median, 3) the firstanalogue field, 4) the weighted average giving more importance to first analogues and 5) thetemperature value of a random analogue.For the two variables and for each type of estimation, different validation measures wereconsidered. The bias, the absolute error (AE) and the mean square error (RMSE) wereanalyzed. The correlation between the estimated and the observed time series was alsoperformed. The analysis was focused on evaluating the ability of the method in representingspatial and temporal variability of the variables as well as the annual cycle.The results show that the best predictor of Tmax and Tmin was air temperature at 850 hPa.The method has a good performance at estimating spatial distributions, seasonal cycles andtemporal variability. Both variables and all estimations showed annual cycle correlationsgreater than 0.98, though winter temperatures tend to be overestimated. The estimations ofTmin were better than for Tmax. However, for both temperatures, AE and RMSE errorsshowed values between 0 and 1°C, specially for the estimations (1), (2) and (4). Winterinterannual correlations were mostly between 0.7 and 0.81, while summer interannualcorrelations showed values between 0.64 and 0.87. In all cases, the spatial distribution ofestimated fields is in agreement with the observed fields. The statistical model reproduces thespatial structure of isoline distributions and also the gradients.