Using satellite products to improve random forest-based error correction of hydrological streamflow model

Abstract

Hydrological models are important tools for making streamflow predic- tions and studying the effects of climate change on water resources around the world. This study aims to improve a global hybrid hydrological stream- flow framework based on PCR-GLOBWB and Random Forest (RF), with the addition of satellite products. The selected satellite products are liq- uid water equivalent from GRACE (Gravity Recovery and Climate Ex- periment), snow cover fraction from JASMES (JAXA Satellite Monitoring for Environmental Studies) and soil moisture from ESA CCI SM (Euro- pean Space Agency Climate Change Initiative Soil Moisture). These prod- ucts are selected because of their relevance for streamflow predictions. Five global and seven Local (Australia, Canada and the United States) model configurations were used for the RF model. The differences be- tween configurations are based on the inclusion of state variables from PCR-GLOBWB, catchment attributes, satellite products, lagged variables for meteorological input and satellite products of 4 and 12 months and exclusion of state variables from PCR-GLOBWB related to satellite prod- ucts. The results showed that the global and local configurations did not improve compared to the benchmark model (based on 51 predictors from PCR-GLOBWB) and the addition of lag was not significantly effective ei- ther. The configuration with only satellite and meteorological input and satellite products did, however show good performance with only six pre- dictors. After adding static variables to the previous configuration there was equal performance to configurations with state variables from PCR- GLOBWB. These result mean that good performance can be achieved with- out the need for a hydrological model and with a limited number of vari- ables.