11:15 〜 11:30
[MGI27-09] Leveraging Japan's National Streamflow Records for End-to-End Data-Driven Hydrological Modeling at National Scale
キーワード：Deep Learning, River Discharge Prediction, Hydrological Modeling
Accurate and timely river discharge simulation is critical for water resource management, flood risk mitigation, and climate adaptation—particularly in Japan, where complex terrain and frequent extreme weather events present create unique hydrological challenges. While traditional physics-based hydrological models remain valuable, they often incur high computational costs and require significant domain expertise for calibration while being limited by their input (e.g. runoff) biais. Recent advances in machine learning (ML) have shown promise for improving both accuracy and efficiency of the rainfall runoff modeling. However such development remains large scale and at point site (i.e. catchment) resolution, failing in considering spatial water travel through river network. Furthermore end-to-end ML solutions tailored to Japan’s diverse hydrological conditions remain scarce.
In this work, we construct a national-scale dataset designed for data-grounded hydrological modeling. Central to our approach is I) a “gauge- and dam-aware” segmentation (Based on river gauge and dam discharge measurement from the Ministry of Land, Infrastructure, Transport and Tourism (MLIT)) of high-resolution, hydrologically corrected digital elevation maps into a comprehensive network of hydrological catchments. We then interpolate meteorological (P and T observations from various sources including JMA and ECMWF) and hydrological catchment descriptors onto these catchments.
Building on this dataset, we propose the first National-scale, end-to-end differentiable and GPU-accelerated hydrological modeling framework for Japan. Our model encompasses runoff generation, routing, and dam modules within a unified architecture. To achieve efficient GPU-based computation, we refactor the traditional Muskingum routing in both time and space, leveraging block-sparse causal convolutions. This design also allows us to infer individual catchment impulse response functions
Applying our model to the newly constructed dataset yields unprecedented accuracy in national-scale river discharge modeling, with a median Nash–Sutcliffe Efficiency (NSE) of 0.81. This performance reflects robust predictive skill, improved generalization across time and space, and a marked improvement over existing state-of art global ML (NSE=0.53) and physical based (NSE=0.30) river discharge products. This new paradigm unveils water management at National scale. Error analysis points to four main sources of uncertainty: snowmelt processes, measurement errors in heavy rainfall events, unknown dam operations, and inaccuracies in river discharge observations.
Finally, the high accuracy of our system reveals previously unquantified uncertainties. In particular, we detect inconsistencies in in-situ discharge measurements that imply calibration discrepancies. These findings highlight both the advantages of the multiple observation synergy leveraged by data-grounded modeling to pinpoint neglected inconsistency and sources of error.
We conclude by emphasizing the need for novel, principled methods that incorporate physical constraints to address observational and calibration uncertainties in future large-scale hydrological modeling efforts.
In this work, we construct a national-scale dataset designed for data-grounded hydrological modeling. Central to our approach is I) a “gauge- and dam-aware” segmentation (Based on river gauge and dam discharge measurement from the Ministry of Land, Infrastructure, Transport and Tourism (MLIT)) of high-resolution, hydrologically corrected digital elevation maps into a comprehensive network of hydrological catchments. We then interpolate meteorological (P and T observations from various sources including JMA and ECMWF) and hydrological catchment descriptors onto these catchments.
Building on this dataset, we propose the first National-scale, end-to-end differentiable and GPU-accelerated hydrological modeling framework for Japan. Our model encompasses runoff generation, routing, and dam modules within a unified architecture. To achieve efficient GPU-based computation, we refactor the traditional Muskingum routing in both time and space, leveraging block-sparse causal convolutions. This design also allows us to infer individual catchment impulse response functions
Applying our model to the newly constructed dataset yields unprecedented accuracy in national-scale river discharge modeling, with a median Nash–Sutcliffe Efficiency (NSE) of 0.81. This performance reflects robust predictive skill, improved generalization across time and space, and a marked improvement over existing state-of art global ML (NSE=0.53) and physical based (NSE=0.30) river discharge products. This new paradigm unveils water management at National scale. Error analysis points to four main sources of uncertainty: snowmelt processes, measurement errors in heavy rainfall events, unknown dam operations, and inaccuracies in river discharge observations.
Finally, the high accuracy of our system reveals previously unquantified uncertainties. In particular, we detect inconsistencies in in-situ discharge measurements that imply calibration discrepancies. These findings highlight both the advantages of the multiple observation synergy leveraged by data-grounded modeling to pinpoint neglected inconsistency and sources of error.
We conclude by emphasizing the need for novel, principled methods that incorporate physical constraints to address observational and calibration uncertainties in future large-scale hydrological modeling efforts.