Tiger-HLM CONUS Evaluation: Model Parameters, Streamflow Metrics, and Simulation Sets

Description This repository contains the supporting datasets, model parameters, spatial hydrography files, streamflow evaluation metrics, gauge mappings, and compressed output hydrographs for the continental-scale hydrologic evaluation using Tiger-HLM hydrologic model. Tiger-HLM is a GPU-accelerated conceptual rainfall-runoff and routing model designed for rapid, continental-scale streamflow simulation. The accompanying software release can be found here. To avoid masking precipitation biases, model parameters are estimated a-priori without calibration, drawing on globally available geospatial products, Hydrography90m stream topology, ERA5-Land reanalysis snow physics, and regional hydraulic geometry scaling laws to establish physical initial conditions. These parameters and datasets are meant to serve as an initial benchmark of Tiger-HLM streamflow simulations and should be improved upon in future versions. This dataset provides complete global runoff and routing parameter maps, regional hydrography networks and parameter sets across the Contiguous United States (CONUS), evaluation metrics comparing model performance against USGS streamflow observations across three time periods (CONUS), and simulated hydrograph outputs across nine hourly gridded precipitation products (CONUS). The CONUS is broken up into three domains for simulations (West, Central, East) and all files included are named accordingly. Model Simulations & Forcing Datasets 1. Primary Long-Term Simulation Set (AORC): Forcing: Analysis of Record for Calibration (AORC) 1-km hourly precipitation and daily temperature. Period: Water Years 1982–2024 (simulation run from 1980 through 2024). Spin-up: Year 1980 was run repeatedly for 10 cycles to initialize soil and groundwater storage tanks prior to evaluation. Purpose: Serves as the primary benchmark and enables direct comparison against retrospective National Water Model (NWM) simulations. 2. Precipitation Sensitivity Ensemble: Forcings (9 Products): Gauge-based (NLDAS, Stage IV), satellite-derived (CMORPH, IMERG), dynamically downscaled reanalysis (CONUS404), global reanalysis (ERA5-Land), and multi-source blended products (AORC, EM-Earth, MSWEPv3). Period: Simulation run from 2002 to 2019; evaluation period covers Water Years 2004–2019 (16 years). Temperature: AORC temperature was applied across all runs to isolate streamflow sensitivity strictly to precipitation forcing choice. 3. Output Resolution: 15-minute instantaneous discharge at every channel link. Daily aggregated mean discharge time series. Water Year (Oct 1–Sep 30) annual maximum instantaneous discharge values. Observation Data & Network Mapping Simulated streamflow was benchmarked against USGS streamgage observations retrieved via the dataRetrieval Python package across three distinct evaluation periods: Long-Term Primary: 1 October 1981 – 30 September 2024 (42 Water Years). NWM Benchmark Period: 1 October 1983 – 31 December 2016 (~33 years) for direct comparison to retrospective NWM evaluations (Towler et al., 2023). Precipitation Comparison: 1 October 2003 – 30 September 2019 (16 Water Years). Quality & Topology Filtering Rules: Gauge Network Matching: USGS streamgages were snapped to the nearest Hydrography90m reach link. For evaluation, gauges were retained only if the relative difference between USGS-reported drainage area and Hydrography90m drainage area was ≤ 5%. Completeness Thresholds: Instantaneous and daily metric evaluations required ≥ 25% complete data for the period of record. Peak flow analysis required ≥ 30 years of record for long-term periods and all 16 complete years for the sensitivity period. All metrics are provided and should be filtered by the user. Analysis & Evaluation Metrics To evaluate model performance across diverse hydroclimatic regimes without bulk metric bias, four main metrics plus annual peak flow errors were computed (metrics.zip): Kling-Gupta Efficiency (KGE): Evaluates overall skill score combining correlation (r), variability ratio (α), and bias ratio (β):KGE = 1 - √((r - 1)² + (α - 1)² + (β - 1)²)(Lower benchmark bound: -0.41, representing mean flow skill score). Volumetric Bias (Percent Bias): Measures systematic volume over- or under-estimation:Volumetric Bias = Σ(Q_sim - Q_obs) / Σ(Q_obs) Normalized Root Mean Square Error (NRMSE): Normalized by mean observed discharge; sensitive to large magnitude errors and peak flow events:NRMSE = (√(mean((Q_sim - Q_obs)²))) / mean(Q_obs) Normalized Mean Absolute Error (NMAE): Normalized by mean observed discharge; treats errors linearly across the hydrologic regime:NMAE = (mean(|Q_sim - Q_obs|)) / mean(Q_obs) Annual Maximum Discharge NMAE: Evaluates the generation of annual peak flood magnitudes by taking NMAE between the maximum simulated instantaneous discharge per water year and the observed annual peak flow. Summary of A-Priori Parameter Derivations Hillslope Domain: Hydrography90m subbasins; mean subbasin area ~0.18 km², mean channel reach length ~0.47 km. Runoff Parameters (12 total): Hmax from Stocker et al. (2023) root storage (5-km, 6x6 smoothed); Infiltration rates (I2, I3) set to 15% of surface/100-cm Ksat (Gupta et al., 2021); Interflow (C3) from 30-cm Ksat; Baseflow (C4) from WaterGAP storage/discharge ratios; Relative soil moisture ET points, Sw & Ss, based on soil types from HWSD v2.0; Global Land Cover 2020 to Manning's n mapping and Geomorpho90m slopes. Snow Module: Degree-day method with Tmelt = 0°C; Gridded melt factor (M) derived from ERA5-Land SWE reductions on active melt days (M = ΔSWE / (T_daily - T_melt)); Tacc threshold mapped on a 9-km grid. Routing: Channel power laws (v = v0 * qλ1). Gauged HUC4 parameters fitted via linear regression on log-transformed USGS NWIS velocity-discharge data (R² ≥ 0.3, N ≥ 5). Ungauged basins estimated using a 500-member XGBoost multi-output regressor ensemble (R² = 0.83 train, R² = 0.49 test) trained on 8 HUC4 geophysical/climatological predictors. Repository Contents & Structure The repository is organized into five primary zipped archives: 1. hydrography.zip GeoPackage vector files defining stream channels and regional domain boundaries derived from Hydrography90m: CONUS_[Region]_Links.gpkg: Vector lines defining reach-level stream topology, network hierarchy, and reach IDs (East, Central, West). CONUS_[Region]_Region.gpkg: Polygon boundaries defining the three regional modeling domains. 2. parameters.zip Contains ready-to-use parameter CSV files, raw spatial inputs, and machine learning pipeline artifacts: parameters/runoff/: Ready-to-use CSV tables (CONUS_[Region]_runoff_params_V1.csv) containing 12 runoff parameters for every hillslope subbasin. parameters/routing/: CSV tables (CONUS_[Region]_routing_params_V1.csv) with v0 and lambda1 routing parameters per channel reach. parameters/conus_hsl_attrs.csv: Physical attributes for all CONUS hillslopes and links. parameters/inputs/runoff_parameter_maps/: Raw global geospatial rasters (.tif, .nc, .nc4) used to compute static storage, infiltration, interflow, aquifer time, ET points, land cover n, terrain slope, and snow parameters. parameters/inputs/routing_parameters_gpkg/: XGBoost routing parameter predictors (huc4_predictors.gpkg), USGS NWIS fits (huc4_fit.gpkg), serialized models (.pkl), and train/test split definitions (ML/xgb_results2/). final HUC4 ML predictions (huc4_xgb_predictions.gpkg), final HYBAS4 Global ML predictions (huc4_xgb_predictions.gpkg) 3. metrics.zip Evaluation CSV tables containing metric scores (KGE, Volumetric Bias, NRMSE, NMAE) and annual peak values: AORC_daily_metrics_1981_2024.csv / AORC_inst_metrics_1981_2024.csv / AORC_peak_metrics_1981_2024.csv: Metrics for the primary 42-year AORC simulation. AORC_*_1983_2016.csv: Metrics for the retrospective NWM comparison period (~33 years). [Product]_[daily|inst|peak]_metrics.csv: Metrics across all 9 precipitation forcing products for the sensitivity period (WY 2004–2019). [Product]_peak_values.csv: Extracted annual peak discharge values and timings per water year. 4. usgs_mapping.zip Mapping cross-walk tables: gauge_mapping.csv: Cross-walk table mapping USGS station IDs directly to Hydrography90m link identifiers (for analysis we suggest it be filtered to ≤ 5% drainage area difference). network_mapping.csv: Approximate NHDPlus to Hydrography90m network conflation mapping. 5. hydrographs.zip Compressed NetCDF (.nc) time series files containing simulated 15-minute discharge for all CONUS stream reaches at USGS gauge locations, split by region (East, Central, West) and precipitation forcing product across both simulation periods.