5:15 PM - 7:15 PM
[ACG48-P06] Improving Flood Forecast in Large-Scale River Models through Data Assimilation of Satellite Observed Flood Extent
Keywords:flood forecast, data assimilation, river model, satellite
1. Introduction
River models are key for flood prediction, but models alone are not precise enough. Thus, data assimilation has been studied to improve their accuracy. River models need updated storage from observations as the initial value of calculation. Assimilation of Water Surface Elevation (WSE) has been advanced, but Flood Extent (FE) assimilation lags due to the non-unique water storage-FE relationship. This study estimated initial water storage from satellite FE data for assimilation into CaMa-Flood[1]. The effectiveness of assimilation was evaluated during the 2022 Pakistan flood.
2. Methods
Two water storage estimation methods were proposed. First, the inverse function method estimates storage using a storage-FE curve, based on the physics of CaMa-Flood that assumes WSE in each catchment is constant. Another approach, independent of model physics, uses the FLEXTH[2] algorithm. It estimates WSE from flood boundaries and calculates water depth using satellite FE and digital elevation model. This depth is integrated per catchment to estimate total water storage.
Estimated storage was assimilated using the Local Ensemble Transform Kalman Filter (LETKF). Two experiments were conducted: OSSE and real-data assimilation. In OSSE, ensemble members were generated by perturbing ERA5 runoff, and unperturbed simulation was treated as the "true" state for synthetic FE. In real-data assimilation, FE from synthetic aperture radar data (Global Flood Monitoring) was assimilated. Assimilation was applied only on August 30, 2022, at peak of flooding. A forecast simulation was run after that and observed and predicted FE were compared to assess the effectiveness of assimilation.
3. Results
Regarding storage estimation in OSSE, both methods provided estimations around true values (Fig. 1). The inverse function method closely matched true values because CaMa-Flood’s physics is also valid in the synthetic reality. The FLEXTH-based method deviated in catchments with discontinuous water depth due to its continuous estimation.
Assimilation improved OSSE F1 scores (Fig. 2). While the inverse function method worked as expected, FLEXTH also improved accuracy.
In real-data assimilation, although true storage was unknown, both methods provided comparable values, confirming validity. However, in the inverse function method, high internal elevation differences led to overestimated water depth (Fig. 3), causing larger storage estimates than FLEXTH. FLEXTH avoided extreme overestimation by ensuring spatially continuous water depth.
Forecast accuracy improved immediately after assimilating estimated storage from both methods. However, in the inverse function method, precision declined 12 days post-assimilation (Fig. 4), likely due to the overestimation issue. FLEXTH showed higher precision. Unlike OSSE, recall monotonically decreased after 24 days. This degradation is attributed to discrepancies between model physics and reality. ICESat-2 altimetry data suggests roads and channels hinder water movement within catchments (Fig. 5). Additionally, WSE of the lake was higher than floodplains, indicating embankments disrupted river-floodplain connectivity, preventing drainage. Since accuracy declines, better model physics are needed.
4. Conclusion and Future Outlook
This study demonstrated that assimilating water storage estimated from spaceborne FE improves model predictions. Future research should refine assimilation techniques by improving error covariance matrices and timing. Additionally, integrating optical and SAR satellites for high-frequency assimilation is a promising direction.
References
(1) Yamazaki, D., et al. (2011). Water Resources Research, 47(4).
(2) Betterle, A., & Salamon, P. (2024). Natural Hazards and Earth System Sciences, 24(8), 2817–2836.
River models are key for flood prediction, but models alone are not precise enough. Thus, data assimilation has been studied to improve their accuracy. River models need updated storage from observations as the initial value of calculation. Assimilation of Water Surface Elevation (WSE) has been advanced, but Flood Extent (FE) assimilation lags due to the non-unique water storage-FE relationship. This study estimated initial water storage from satellite FE data for assimilation into CaMa-Flood[1]. The effectiveness of assimilation was evaluated during the 2022 Pakistan flood.
2. Methods
Two water storage estimation methods were proposed. First, the inverse function method estimates storage using a storage-FE curve, based on the physics of CaMa-Flood that assumes WSE in each catchment is constant. Another approach, independent of model physics, uses the FLEXTH[2] algorithm. It estimates WSE from flood boundaries and calculates water depth using satellite FE and digital elevation model. This depth is integrated per catchment to estimate total water storage.
Estimated storage was assimilated using the Local Ensemble Transform Kalman Filter (LETKF). Two experiments were conducted: OSSE and real-data assimilation. In OSSE, ensemble members were generated by perturbing ERA5 runoff, and unperturbed simulation was treated as the "true" state for synthetic FE. In real-data assimilation, FE from synthetic aperture radar data (Global Flood Monitoring) was assimilated. Assimilation was applied only on August 30, 2022, at peak of flooding. A forecast simulation was run after that and observed and predicted FE were compared to assess the effectiveness of assimilation.
3. Results
Regarding storage estimation in OSSE, both methods provided estimations around true values (Fig. 1). The inverse function method closely matched true values because CaMa-Flood’s physics is also valid in the synthetic reality. The FLEXTH-based method deviated in catchments with discontinuous water depth due to its continuous estimation.
Assimilation improved OSSE F1 scores (Fig. 2). While the inverse function method worked as expected, FLEXTH also improved accuracy.
In real-data assimilation, although true storage was unknown, both methods provided comparable values, confirming validity. However, in the inverse function method, high internal elevation differences led to overestimated water depth (Fig. 3), causing larger storage estimates than FLEXTH. FLEXTH avoided extreme overestimation by ensuring spatially continuous water depth.
Forecast accuracy improved immediately after assimilating estimated storage from both methods. However, in the inverse function method, precision declined 12 days post-assimilation (Fig. 4), likely due to the overestimation issue. FLEXTH showed higher precision. Unlike OSSE, recall monotonically decreased after 24 days. This degradation is attributed to discrepancies between model physics and reality. ICESat-2 altimetry data suggests roads and channels hinder water movement within catchments (Fig. 5). Additionally, WSE of the lake was higher than floodplains, indicating embankments disrupted river-floodplain connectivity, preventing drainage. Since accuracy declines, better model physics are needed.
4. Conclusion and Future Outlook
This study demonstrated that assimilating water storage estimated from spaceborne FE improves model predictions. Future research should refine assimilation techniques by improving error covariance matrices and timing. Additionally, integrating optical and SAR satellites for high-frequency assimilation is a promising direction.
References
(1) Yamazaki, D., et al. (2011). Water Resources Research, 47(4).
(2) Betterle, A., & Salamon, P. (2024). Natural Hazards and Earth System Sciences, 24(8), 2817–2836.