Process-Aware Interpolation Technique for Downscaling Hydrological Variables

Water is an essential resource for human society. Numerous approaches have been developed in order to assess the availability of water for our societies. Information on the availability of water is readily available from various databases, which, however, often come in spatial aggregation which is too coarse for detailed analysis or local use cases. In this research, we propose a Process-Aware Interpolation (PAI) technique based on previous research in advanced areal interpolation. In areal interpolation, the value of a variable from a source zone is reallocated among intersecting target zones. In PAI, ancillary information on process which causes the values being interpolated is used to improve the quality of interpolation.

We test the PAI methodology in a surrogate modelling context by downscaling runoff outputs from the Community Water Model (CWatM) at 30 arc-minute resolution and compare the downscaled output to CWatM model runs at 5 arc-minute resolution. We develop two surrogate models – simplified models emulating a more complex one - based on machine learning
(Random Forest Regression) and classical statistical methods (Ordinary Least Squares Regression). The surrogate models are used within the PAI framework as the ancillary information guiding the interpolation. The quality of the interpolation is assessed against a full run of CWatM at 5 arc-minute resolution, and compared to the surrogate models outside of the PAI framework as well as two simpler PAI benchmarks – a constant ancillary variable (rainfall) and an expert-knowledge based model.

We find that the developed surrogate models perform significantly better when used within the PAI framework than outside. Further, PAI with the simpler benchmarks can produce comparable quality interpolation to the PAI with surrogate models. The quality of the interpolation is, however, highly dependent on the quality of the source data.