Triple Collocation for Streamflow Estimation
Assessing and optimizing high-resolution global river streamflow estimates with triple collocation analysis
Overview
This study evaluates three global hydrological models (CWatM, PCR-GLOBWB, H08) at both high (5 arcmin) and low (30 arcmin) resolutions, using discharge data from 1,707 GRDC stations worldwide and 62 stations in Thailand. It introduces Triple Collocation (TC) as a statistically robust method to fuse multi-model outputs, improving streamflow accuracy without requiring dense gauge networks.
Why It Matters
Accurate streamflow estimates are critical for flood forecasting, drought mitigation, and sustainable water management. By demonstrating that high-resolution models and multi-model fusion significantly improve accuracy, this study provides practical guidance for policymakers and water managers, especially in ungauged or poorly monitored regions. The Thailand case study shows how global models can be tailored for national-scale decision-making.
Figure 1: (a) Global distribution of GRDC gauge stations (blue dots) that provide data between 2000 and −2022. The focused study areas are shown in dotted circles with different colors. (b) The inset (at the bottom-left corner) shows the distribution of the measurement sites in Thailand (not included in GRDC).
How It Works
The study runs three open-source models (CWatM, PCR-GLOBWB, H08) with ERA5/ERA5-Land forcing, then applies Triple Collocation (TC) to combine outputs. TC estimates optimal weights based on error covariance, producing a fused streamflow dataset. Performance is compared against simple averaging and individual models, using both global GRDC and local RID observations.
Figure 2: The distribution of GRDC stations that are used in this study.
Key Advances
- First global-scale application of TC for streamflow.
- Resolution analysis: Demonstrates clear gains from 5 arcmin vs. 30 arcmin.
- Model comparison: CWatM consistently outperforms others, especially at high resolution.
- Fusion benefit: TC improves accuracy beyond averaging, especially with H08 as reference.
Figure 3: Global distribution of KGE values under different simulation scenarios, (a) CWatM (05 arcmin), (b) CWatM (30 arcmin), (c) H08 (05 arcmin), (d) H08 (30 arcmin), (e) PCR-GLOBWB (05 arcmin), (F) PCR-GLOBWB (30 arcmin), (g) Average (05 arcmin), (h) Average (30 arcmin), (i) TC (05 arcmin), (j) TC (30 arcmin).
Best Practices
High-resolution models should be prioritized for complex basins, and multi-model fusion (TC) should be applied to reduce uncertainty. Validation against both global and local datasets ensures transferability. Consistent forcing inputs (ERA5/ERA5-Land) are essential to isolate model differences.
Challenges
Despite improvements, challenges remain. Sparse and uneven global gauge coverage limits calibration, while shared forcing data can introduce inter-model error dependence. High-resolution simulations demand significant computational resources, and model transferability is uncertain in regions with limited observational data.
Figure 4: Boxplots of model performance metrics (KGE) (see NRMSE, NSE, and R2 in supplements) for different hydrological models and spatial resolutions in the global. The label format is Model_Resolution, where A, C, H, P, and T represent Average, CWatM, H08, PCR-GLOBWB, and TC.
Findings
The results show that CWatM at high resolution consistently outperforms PCR-GLOBWB and H08, particularly in Europe. Both TC and averaging improve accuracy compared to individual models, but TC delivers the highest overall performance. Resolution analysis confirms that finer grids yield more accurate estimates in most regions, validating the push toward high-resolution hydrological modeling.
Figure 5: Regional-average KGE computed from different model simulation scenarios. The bar chart represents the normalized KGE values, allowing relative performance comparisons by bar length.
Figure 6: Streamflow estimates from different model simulation scenarios (red) and in situ data (black) over three river basins, the Danube River (left column), Rhine River (middle column), and Mississippi River (right column).
Key Takeaways
- High-resolution models (5 arcmin) are more reliable than coarse ones.
- Multi-model fusion (TC) reduces uncertainty and bias.
- CWatM is the most robust performer globally.
- Regional case studies (Thailand) confirm practical utility.
- TC is a transferable, statistically grounded method for global hydrology.
Please Find the full publication here: https://doi.org/10.1016/j.jhydrol.2026.135122
RECENTLY PUBLISHED
GRACE/-FO DATA ASSIMILATION– Best Practices & Future Directions
Weight-Supported Random Forest Downscaled GRACE(-FO) Data
GRACE/-FO DATA ASSIMILATION– Best Practices & Future Directions
RACE and GRACE‑FO satellite data assimilation has transformed monitoring of Earth’s water cycle. By merging coarse satellite observations with fine‑scale hydrological models, it improves detection of groundwater loss, snowpack changes, floods, and droughts. This approach bridges science and practice, delivering actionable insights for water management, climate resilience, and disaster preparedness. With next‑generation missions, low‑latency products, and machine learning hybrids on the horizon, GRACE/-FO assimilation is set to become an even more powerful tool for safeguarding global freshwater resources.
Weight-Supported Random Forest Downscaled GRACE(-FO) Data
This study investigates groundwater depletion in the Lower Indus Basin (LIB), where intensive irrigation has led to unsustainable groundwater extraction. To address the coarse spatial resolution of satellite-derived water storage data, a Weight-Supported Random Forest (WSRF) model was developed to downscale data to a finer 0.1° resolution using precipitation, evapotranspiration, vegetation index, and soil moisture as predictors. The results showed high agreement with groundwater well observations (R² > 0.85) and revealed detailed spatial patterns of groundwater depletion from 2003–2023, highlighting the combined influence of climate variability and irrigation practices.