Flash flooding is one of the most damaging weather types, yet it remains challenging to quantify its severity. We propose a novel development – the Flashiness-Intensity-Duration-Frequency (F-IDF) curve – to quantify and spatially analyze flash flood intensity based on the frequency and duration of the event. As a proof-of-concept, we mapped Contiguous US (CONUS)-wide F-IDF values at 3,722 stream gage locations and explored their relations with 59 basin attributes. It is found that (1) Climatological precipitation amounts exhibit the most positive correlation with flashiness while a basin’s drainage area is the most negatively correlated; (2) Correlation of flashiness with basin attributes decreases with increasing F-IDF return periods and shorter event durations. Both aspects are attributable to the rainfall signal overwhelming the underlying basin attributes as the intensities become more extreme. This new term can have implications for hydrology, especially for hydrologic modelers, decision-makers, and emergency responders.

Effective flash flood forecasting and risk communication are imperative for mitigating the impacts of flash floods. However, the current forecasting of flash flood occurrence and magnitude largely depends on forecasters’ expertise. An emerging flashiness-intensity-duration-frequency (F-IDF) product is anticipated to facilitate forecasters by quantifying the frequency and magnitude of an imminent flash flood event. To make this concept usable, we develop two distributed F-IDF products across the contiguous US, utilizing both a Machine Learning (ML) approach and a physics-based hydrologic simulation approach that can be applied at ungaged pixels. Specifically, we explored 20 common ML methods and interpreted their predictions using the Shapley Additive exPlanations method. For the hydrologic simulation, we applied the operational flash flood forecast framework – EF5/CREST. It is found that: (1) both CREST and ML depict similar flash flood hot spots across the CONUS; (2) The ML approach outperforms the CREST-based approach, with the drainage area, air temperature, channel slope, potential evaporation, soil erosion identified as the five most important factors; (3) The CREST-based approach exhibits high model bias in regions characterized by dam/reservoir regulation, urbanization, or mild slopes. We discuss two application use cases for these two products. The CREST-based approach, with its dynamic streamflow predictions, can be integrated into the existing real-time flash flood forecast system to provide event-based forecasts of the frequency and intensity of floods at multiple durations. On the other hand, the ML-based approach, which is a static measure, can be integrated into a flash flood risk assessment framework for urban planners.

We present a near surface air temperature (NSAT) fused data product over the contiguous United States using Level 2 data from the Atmospheric Infrared Sounder (AIRS), on the Aqua satellite, and the Cross-track Infrared Microwave Sounding Suite (CrIMSS), on the Suomi National Polar-orbiting Partnership (SNPP) satellite. We create the fused product using Spatial Statistical Data Fusion (SSDF), a procedure for fusing multiple datasets by modeling spatial dependence in the data, along with ground station data from NOAA’s Integrated Surface Database (ISD) which is used to estimate bias and variance in the input satellite datasets. Our fused NSAT product is produced twice daily and on a 0.25-degree latitude-longitude grid. We provide detailed validation using withheld ISD data and comparison with ERA5-Land reanalysis. The fused gridded product has no missing data; has improved accuracy and precision relative to the input satellite datasets, and comparable accuracy and precision to ERA5-Land; and includes improved uncertainty estimates. Over the domain of our study, the fused product decreases daytime bias magnitude by 1.7 K and 0.5 K, nighttime bias magnitude by 1.5 K and 0.2 K, and overall RMSE by 35% and 15% relative to the AIRS and CrIMSS input datasets, respectively. Our method is computationally fast and generalizable, capable of data fusion from multiple datasets estimating the same quantity. Finally, because our product reduces bias, it produces long-term datasets across multi-instrument remote sensing records with improved bias stationarity, even as individual missions and their data records begin and end.