Evaluating the use of different distance measures in statistical downscaling of climate parameters using the K-NN method

Soroosh Sharifi, Philippe Ho, Saeed Golian

Tuesday 30 june 2015

14:20 - 14:35h at North America (level 0)

Themes: (T) Extreme events, natural variability and climate change, (ST) Hydrological extremes: floods and droughts

Parallel session: 6I. Extreme events - Flood Drought

General Circulation Models (GCMs) are the main tools available to the scientific community for modelling current and future climate conditions. They attempt to describe average conditions over a coarse resolution, and therefore, their simulations cannot be directly used in basin-scale studies such as developing climate change adaptation schemes and disaster risk reduction planning. One solution to overcome this limitation is to relate the large-scale atmospheric patterns to the local climate, using statistical tools in a process called statistical downscaling. This will enable deriving local-scale information of climate variables such as precipitation and temperature from the climate change scenarios generated by GCMs. The k-nearest neighbour (k-nn) algorithm is one of the most simple and impressive tools that has been extensively used in the past for statistical downscaling of large-scale GCM outputs. This non-parametric technique has the major advantage of avoiding the complex parameterization process of other statistical downscaling techniques. The only main issue in using this technique is the choice of a function (also referred to as metric or distance measure) which measures the distance between points in the state space for identifying the nearest neighbours. Conventionally, the Euclidean distance measure is used for such studies, however, it is reported that other measures such as the Mahalanobis distance have operational advantages and yield more accurate results. Modellers are often confused by the number of available distance measures and often select a measure without evaluating its correctness which may lead to inaccurate downscaling results. In this study it is aimed to evaluate the use of different distance measures in statistical downscaling of climate data. For this purpose, a catalogue of distance measures, including Euclidean, Standardized Euclidean, City block, Chebychev, Minkowski, Mahalanobis and Hamming will be used to perform downscaling on different Met Office Hadley Centre climate change scenarios for the River Severn basin. Evaluation and comparison of the results will allow the creation of a comprehensive guide on choosing the most appropriate distance measures for downscaling.