H040-0004
Rivers discharge estimations from SWOT-like data by a hybrid data-driven physically-based algorithm
Rivers discharge estimations from SWOT-like data by a hybrid data-driven physically-based algorithm
Tuesday, 8 December 2020
Poster
Abstract:
We present the new version of the HiVDI (Hierarchical Variational Discharge Inference) algorithm [1] enabling the estimation of discharge and bathymetry of rivers from altimetry measurements, more particularly from the forthcoming SWOT space mission. The present approach is hybrid in the sense it is based on a purely data-driven stage, next on physically-based stages. Following [2,3], hierarchical flow models and Variational Data Assimilation (VDA) remains key ingredients. HiVDI algorithm is now based on three stages: 1) a first purely data-driven estimation obtained by an Artificial Neural Network; the ANN is trained from the altimetry measurements and rough drainage area values; 2) a first physically-consistent estimation of discharge is obtained from a dedicated satellite-scale flow model; 3) an advanced VDA process based on the dynamic Saint-Venant’s model provides estimations of discharge Q(x,t), bathymetry b(x) and effective friction parameter K(x;h(x,t)).
The final estimation of Q(x,t) is highly accurate for rivers presenting features within the learning partition. For rivers far outside the learning partition, the space-time variations of discharge is accurate too ; however, the global estimation still may present a bias. It is shown that if the estimation is based on the flow models only, the inversions are well-defined but with an intrinsic bias; the bias scales the global estimation. For rivers outside the learning partition, any mean value (eg. annual, seasonal) enables to remove the bias.
Finally, this new hybrid inversion strategy seems to provide much more accurate estimations compared to previous ones; examples are shown for 29 heterogeneous Pepsi2 [4] river portions.
The final estimation of Q(x,t) is highly accurate for rivers presenting features within the learning partition. For rivers far outside the learning partition, the space-time variations of discharge is accurate too ; however, the global estimation still may present a bias. It is shown that if the estimation is based on the flow models only, the inversions are well-defined but with an intrinsic bias; the bias scales the global estimation. For rivers outside the learning partition, any mean value (eg. annual, seasonal) enables to remove the bias.
Finally, this new hybrid inversion strategy seems to provide much more accurate estimations compared to previous ones; examples are shown for 29 heterogeneous Pepsi2 [4] river portions.
References
[1] K. Larnier, J. Monnier. Submitted.
[2] K. Larnier, J. Monnier, P.-A. Garambois, J. Verley. Inv. Prob. Sc. Eng. (IPSE) 2020.
[3] P. Brisset, J. Monnier, P.-A. Garambois, H. Roux. Adv. Water Ress. 2018.
[4] Pepsi datasets from the Discharge Algorithm Working Group (DAWG), Science Team NASA-CNES et al. SWOT mission.