Auteur : Donghwan Kim

Co-Auteur :  Hyongki Leeb, Edward Beighley , Raphael M. Tshimanga

Review : Advances in Space Research

Lien : https://doi.org/10.1016/j.oneear.2020.02.008.

Abstract

The Congo River is one of the least studied basins although it is the world’s second largest in size (~3.7 million km2) and discharge (Q) (~40,600 m3 s−1). Using remote sensing data and a hydrologic model, previous studies have successfully estimated Q of the Congo River at Brazzaville-Kinshasa stations with an accuracy of 10–40%. However, those studies depended on only a hydrologic model or estimated Q using remotely sensed data with a single rating curve. Recently, Kim et al. (2019) has also successfully applied the ensemble learning regression method, which is one of the machine learning techniques, to estimate Q (termed as ELQ) by linearly combining several rating curves over different locations. The study has estimated Q at the Brazzaville station with relative root-mean-square error (RRMSEs) of 7.17/5.53% for training/validation datasets whose temporal resolutions are 35-day for the period from 2002 to 2010. However, ELQ still requires in-situ Q data in order to train base learners and obtain their weights. In this study, we present a study estimating daily Q by applying ELQ with satellite altimetry data and the hydrologic-hydraulic Hillslope River Routing (HRR) model for the Congo River. We parameterized our model with the HRR-derived Q () in training/validation datasets without the aid of in-situ Q. ELQ-derived Q () has not been calibrated or scaled with in-situ Q data in training/validation datasets. However, the HRR model has been calibrated to mean monthly historical gauge measurements (1903–1990).  showed RRMSEs of 15.72/18.00% for training/validation datasets compared with daily in-situ Q data at the Kinshasa station spanning from November 2002 to September 2010. We used the Basic Ensemble Method (), which employs the uniformly distributed weights in ELQ process, that can provide improved estimates of Q.  using  showed RRMSE of 11.32/9.05% on average for training/validation datasets. Moreover, we introduced the method generating more accurate ELQ using the ensemble mean of  () which yielded RRMSEs of 7–10%. This study demonstrates that ELQ can provide more accurate daily Q using satellite altimetry data and a hydrologic model for poorly gauged river basins.

Keywords :  Congo River, ELQ, Ensemble learning regression, Basic Ensemble Method (BEM), Hillslope River Routing (HRR)