Hydraulic variables obtained from remotely sensed data have been successfully used to estimate river discharge
(Q). However, most studies have used a rating curve based on a single hydraulic variable or the Manning
equation (multiplicative method). In this study, we developed a mathematically different approach to estimating
Q by applying the ensemble learning regression method (here termed ELQ), which is one of the machine learning
techniques that linearly combine several functions to reduce errors, over the Congo mainstem as a test-bed.
Using the training dataset (November 2002 ? November 2006) of water levels (H) derived from different Envisat
altimetry observations, the ELQ-estimated Q at the Brazzaville in-situ station showed reduced root-mean-square
error (RMSE) of 823m3 s?1 (relative RMSE (RMSE normalized by the average in-situ Q, RRMSE) of 2.08%)
compared to the Q obtained using a single rating curve. ELQ also showed improved performance for the validation
dataset (December 2006 ? September 2010). Based on the error analysis, we found the correlation
coefficients between input variables affect the performance of ELQ. Thus, we introduced an index, termed the
Degree of compensation (IDoC), which describes how ELQ performs compared to the classic hydraulic relation
(e.g., H-Q rating curve). The performance of ELQ improves when IDoC increases because the additional information
could be added in the ELQ process. Since ELQ can combine several variables obtained over different
locations, it would be advantageous, particularly if there exist few virtual stations along a river reach. It is
expected that ELQ can be also applied to the products of the Surface Water Ocean Topography (SWOT) mission,
which will provide direct measurements of surface water slope (S), effective river width (We), and H, to be
launched in 2021.