GEOGloWS Evaluation for Nepal Flood AA Framework

Author

CHD Data Science

Published

May 1, 2026

GEOGloWS Evaluation

Purpose

This book evaluates the GEOGloWS River Forecast System (v2) as a potential data source for Nepal’s flood anticipatory action (AA) trigger framework.

The framework currently uses GloFAS (ECMWF/Copernicus) for trigger calibration and DHM (Department of Hydrology and Meteorology, Nepal) water-level observations for validation. Google GRRR is included here as a third reference signal for cross-comparison.

GEOGloWS offers a complementary global streamflow service with some distinct advantages — most notably an 86-year retrospective simulation (1940–present) and a free, unauthenticated API. This evaluation asks whether and how it can strengthen the existing framework.

Headline Findings

Bottom line: GEOGloWS is not feasible as the basis for an AA trigger framework. Two structural gaps make calibrated, RP-based triggering impossible without external data:

  1. No reforecasts. GEOGloWS does not publish a multi-year reforecast archive, which is required to calibrate lead-time-dependent skill metrics (e.g., exceedance probability at 3-day or 7-day lead). The live forecast archive is ~22 months, far too short.

  2. The published return-period thresholds are calibrated to the retrospective, but the operational forecast runs at a systematically different magnitude. Using the ~22-month archive of live forecasts, we find at the verified-location Chatara reach that the forecast distribution sits on the 2:1 line below retrospective — forecast ≈ retro / 2 — across all 15 lead days, including the AA action (day 3) and readiness (day 7) leadtimes. The ensemble mean has never reached the published RP2 (14,459 m³/s) at any lead time over the full archive; an ensemble-mean trigger using this threshold would not fire. See 6  Forecast vs Retrospective Magnitudes.

The remaining findings characterize the gaps in more detail:

  1. Against observed DHM danger-level crossings, GEOGloWS at RP2 detects 1 of 9 Chatara events and 0 of 7 Chisapani events (within a ±7 day window). GloFAS, on the same comparison, detects 2 of 9 and 5 of 7. GEOGloWS does not improve on GloFAS at this validation.

  2. At Chisapani (Karnali) the GEOGloWS-vs-GloFAS disagreement is structural. Annual-maxima are essentially uncorrelated (r = 0.07 over 1979–2023), and GEOGloWS’s mean discharge is ~28% of GloFAS’s. Caveat: the chosen reach has not been geographically verified against the DHM gauge location, so this disagreement may reflect a reach-mapping issue rather than a product difference.

  3. Built-in SFDC bias correction does not change the event-detection verdict at either station. Corrected counts are still 1/9 and 0/7.

  4. At Chatara (Koshi), the GEOGloWS retrospective tracks GloFAS well (annual-max r = 0.94). The forecast-vs-retro bias documented in finding 2 is therefore a real product-level issue, not an artefact of a wrong reach.

Validation Caveat

DHM water-level data ends in 2012 at Chatara and 2015 at Chisapani. All event-level validation in this book is bounded by those windows. Anything more recent in the GEOGloWS retrospective is uncheckable against DHM.

Structure

  1. GEOGloWS Overview — what the system is, how it works, API walkthrough
  2. Spatial Matching — finding the correct river reaches
  3. Threshold Comparison — return periods across sources, including bias-corrected GEOGloWS
  4. Time Series & Seasonal Patterns — retrospective overlay across all sources, with bias-correction before/after
  5. Event Detection — validation against observed DHM danger-level crossings
  6. Forecast vs Retrospective Magnitudes — the binding constraint: published RPs are retro-only, forecast runs at half retro magnitude across all lead times
  7. Where Does GEOGloWS Fit? — decision matrix and recommendations

Stations

Station River DHM Danger Level GloFAS RP2 GEOGloWS RP2 GEOGloWS River ID
Chatara Koshi 7.0 m 8,113 m³/s 14,459 m³/s 441135650
Chisapani Karnali 10.5 m 5,664 m³/s 6,745 m³/s 441112306