Regionalization of a Distributed Hydrology Model Using Random Forest

Distributed hydrology models are suitable tools for understanding the hydrological processes, which take place on heterogeneous media under ever-changing internal (e.g. land use change) and boundary conditions (e.g. climate change). The generally accepted practice for applying such models is to calibrate their parameters using observed data. Still in many locations, even in developed countries, observed data is lacking or unreliable. Regionalization is a way around this problem. In this research, we built a Random Forest (RF) model to regionalize the parameters of a distributed hydrology models (Hydrotel), which is the operational model at Quebec government. Using the RF model, the following three hypotheses were tested regarding the efficiency and spatio-temporal variability of the proposed regionalization technique: (1) A finer time-step adds more information to the calibrated parameters and therefore improves the efficiency of the regionalization method; (2) The parameters approximated by RF are spatially consistent and therefore transferrable across spatial scales (i.e. from lumped to sub-catchment to hydrological response units); (3) Using more spatially representative predictors (i.e. by refining the spatial resolution of CDs) to reflect heterogeneity of the catchment will improve the performance of regionalization at internal ungauged locations. All these hypotheses were tested on three groups of nested catchments at 3- and 24-hour time-steps. The results show that for simulations at sub-daily time-steps, the calculated loss of regionalization efficiency (with respect to calibration) is less than that of the 24-hour time-step (12% improvement). Approximating the parameters at different levels of spatial discretization demonstrates that the parameters are spatially consistent as the distribution of parameters and catchment descriptors are spatially correlated. Finally, we found a consistent improvement of simulations when we replace lumped with fully distributed parameters, for simulations with a 24-hour time step. This improvement in the efficiency is higher for catchments with a higher degree of spatial heterogeneity (up to 12%). However, no significant improvement in the efficiency of simulation from lumped to distributed parameters has been observed when the time-step of the simulation was reduced to 3-hour.