GC31C-1196
On a Phase Space Reconstruction Approach to Improve the Statistical Downscaling of Regional Precipitation

Wednesday, 16 December 2015
Poster Hall (Moscone South)
Dhanya C T, Indian Institute of Technology Delhi, New Delhi, India
Abstract:
Climate change impact assessment studies associated with water resources and its management, presently, rely on a great extent on statistical downscaling techniques, to undertake fine resolution hydrologic modelling over a region. Though, recent literature is flooded with many new downscaling methodologies with varying complexities, the inefficiency of these downscaling techniques in accurately capturing the regional extremes, limits the employment of these downscaling techniques in any impact studies in water resources. It merely increases the monstrosity of uncertainty by adding one more factor: downscaling uncertainty. While chaotic nature of climate systems and existence of attractors are irrefutable, the knowledge on system dynamics is not incorporated in the statistical downscaling methodologies so far. A few questions rise from this context: (1) How possibly can we incorporate the information about system dynamics in statistical downscaling techniques? (2) Will it help in accurately simulating the characteristics of actual system, in statistical terms? (3) If so, will such as approach be able to capture the extremes better?. The present study attempts to address the above questions by adopting a phase space reconstruction approach, which reproduces the attractor of the system and inherent dynamics. The dominant variables to replicate the system dynamics are selected as surface air temperature, specific humidity, precipitable water, long wave radiation, and vertical velocity, in addition to rainfall, which is the variable to be downscaled. A multi-variate phase space reconstruction approach, taking into account the relevance and information contained in one variable, is adopted, to estimate the dimension of system and reconstruct the attractor. The downscaling of future rainfall is done by applying a local prediction approach, enforcing the known information of future dominant variables and employing classification and regression trees (CART). The present methodology is expected to simulate accurately the variations in downscaled rainfall.