Sanjeev Panta
University of Louisiana at Lafayette
| Subject Areas: | Data Science |
Recent Activity
ABSTRACT:
The National Water Model (NWM), NOAA's operational physics-based hydrologic model, provides continental streamflow estimates but carries a systematic local bias, a consistent over- or under-estimation that varies by site and season and limits local use. We tested whether different machine learning algorithms (a simple recurrent network, a gated recurrent unit, a long short-term memory network, and a Transformer) can reduce this bias and asked where correction is most useful across three sites in three states (NC, VA, and SD). We trained these four sequence models under two setups: a residual setup that learns the difference between NWM discharge and observed USGS discharge and adds the learned correction back to NWM, and a direct setup that predicts observed discharge directly rather than a correction to NWM, with NWM still among its inputs. Records were split 70% training, 15% validation, 15% testing. We combined the models by simple averaging, error-weighted averaging, and constrained stacking, tuned and scored with time-ordered (walk-forward) cross-validation on a withheld recent block. The workflow was applied to three unregulated USGS gauges spanning NWM skill (Kling-Gupta Efficiency, KGE) from 2010–2020 at hourly resolution: Watauga River (high skill), New River (moderate), and Little Spearfish Creek, a groundwater-fed karst spring (poor). Correction gains increased as NWM skill decreased: KGE improved by +0.12 at Watauga, +0.31 at New River, and +6.90 at Little Spearfish, with corrected KGE reaching 0.83, 0.78, and 0.50 respectively. The residual setup performed best where NWM was reliable. Overall, ML correction generalized across regimes and added the most value where NWM skill was lowest, though at the karst spring it hit a ceiling set by driving information absent from the inputs.
Contact
| (Log in to send email) |
| All | 0 |
| Collection | 0 |
| Resource | 0 |
| App Connector | 0 |
Created: July 17, 2026, 4:15 p.m.
Authors: AKINADE, BASIT · Omar, Ahmed · Panta, Sanjeev · Pineda-Castellanos, Saddy Rafael · Javidian, Mohammad Ali · Mehan, Sushant
ABSTRACT:
The National Water Model (NWM), NOAA's operational physics-based hydrologic model, provides continental streamflow estimates but carries a systematic local bias, a consistent over- or under-estimation that varies by site and season and limits local use. We tested whether different machine learning algorithms (a simple recurrent network, a gated recurrent unit, a long short-term memory network, and a Transformer) can reduce this bias and asked where correction is most useful across three sites in three states (NC, VA, and SD). We trained these four sequence models under two setups: a residual setup that learns the difference between NWM discharge and observed USGS discharge and adds the learned correction back to NWM, and a direct setup that predicts observed discharge directly rather than a correction to NWM, with NWM still among its inputs. Records were split 70% training, 15% validation, 15% testing. We combined the models by simple averaging, error-weighted averaging, and constrained stacking, tuned and scored with time-ordered (walk-forward) cross-validation on a withheld recent block. The workflow was applied to three unregulated USGS gauges spanning NWM skill (Kling-Gupta Efficiency, KGE) from 2010–2020 at hourly resolution: Watauga River (high skill), New River (moderate), and Little Spearfish Creek, a groundwater-fed karst spring (poor). Correction gains increased as NWM skill decreased: KGE improved by +0.12 at Watauga, +0.31 at New River, and +6.90 at Little Spearfish, with corrected KGE reaching 0.83, 0.78, and 0.50 respectively. The residual setup performed best where NWM was reliable. Overall, ML correction generalized across regimes and added the most value where NWM skill was lowest, though at the karst spring it hit a ceiling set by driving information absent from the inputs.