Séminaire
End-to-end learning of river discharge prediction largely resolves the equifinality of distributed river discharge prediction systems over Japan
Tristan Hascoet (RIKEN)
Data-driven models that jointly train runoff generation and river routing to reproduce in-situ discharge achieve high river discharge prediction accuracy while remaining spatially,consistent and mass-conservative. To tackle the problem of equifinality in hydrology modelling, we ask whether ML systems trained end-to-end can nonetheless identify a physically consistent causal structure.
Description
- Data-driven models that jointly train runoff generation and river routing to reproduce in-situ discharge achieve high river discharge predictionaccuracy while remaining spatially consistent and mass-conservative. Yet in hydrology accurate prediction need not recover physically true parameters: because discharge observations are sparse and downstream, many combinations of upstream parameters in a distributed system yield the same basin-level output: the problem of equifinality.
- We ask whether ML systems trained end-to-end can nonetheless identify a physically consistent causal structure. Over 700 Japanese catchments we jointly learn an LSTM runoff module paired with various linear time-invariant routing kernel (Hayami, Muskingum, linear-storage, pure-lag) using an unconstrained network mapping river-channel features (channel length, slope, upstream area, sinuosity) to routing parameters (reach celerity, diffusivity, etc.), and probe identifiability across random seeds, spatial discretizations, and temporal resolutions using statistical tests and symbolic regression of the learned relationships.
- We find that spatio-temporal resolution governs identifiability. Daily simulation disentangles runoff generation processes from the first-order routing control (the effect of channel length on travel time and peak delay) but cannot recover the second-order modulation of wave celerity by channel slope and discharge magnitude, while hourly simulation restores correct identification of this second-order effect.
Informations supplémentaires
Lieu
LMD-Ecole Polytechnique. Salle Lanceau
