Dynamic coupling of support vector machine and K-nearest neighbour for downscaling daily rainfall

- Coupled framework of support vector machine and K nearest neighbour is proposed for downscaling.
- Ensemble approach by considering various plausible model parameter combinations.
- Dynamic framework eliminates relationship stationarity between predictors and predictand.
- Hybrid model captured extreme events and variability better and also the PMFs of wet and dry spells.

SummaryClimate change impact assessment studies in water resources section demand the simulations of climatic variables at coarser scales from dynamic General Circulation Models (GCMs) to be mapped to even finer scales. Related studies in this area have mostly been relying on statistical techniques for downscaling variables to finer resolution. This demands a careful selection of a suitable downscaling model, to alleviate the downscaling uncertainty. In this study, it is proposed to develop a dynamic framework for downscaling purpose by integrating the frequently used techniques, K-Nearest Neighbour (KNN) and Support Vector Machine (SVM). In order to give flexibility in future predictors-predictand relationships and to account the sensitivity in model parameters, it is also proposed to generate an ensemble of outputs by identifying various plausible model parameter combinations. The performance of this framework for downscaling daily precipitation values at different locations is compared with simple KNN and SVM models. The proposed hybrid model is found to be better in capturing various characteristics of daily precipitation than individual models, especially in simulating the extremes, both in magnitude and duration. The mean ensemble is found to be efficient than single best simulation with optimum parameter combinations. The efficacy of hybrid SVM-KNN ensemble downscaling model is established through detailed investigations. The future downscaled projection for mid-century and late century employing this hybrid model indicates an increased variability in future precipitation, though the intensity varies for various locations. The developed methodology hence ensures lesser downscaling uncertainty and also eliminates the inherent assumption of relationship stationarity considered in many downscaling models.