A High-Resolution Spatiotemporal Dataset for Urban Streamflow and Flood Modeling for seven US cities, 2007-2021

Accurate localized streamflow and flood prediction remains a critical challenge in hydrology. Existing open datasets often rely on coarse, large-scale catchments that obscure local dynamics, or on purely grid-based structures that are computationally expensive and misaligned with natural hydrological boundaries. To address this gap, we developed a high-resolution, multi-modal benchmark dataset comprising 4,099 hydrologically delineated sub-watershed “chips” (approximately 2 km² each) spanning seven flood-prone inland U.S. cities (Springfield, Missouri; Phoenix, Arizona; Tulsa, Oklahoma; Louisville, Kentucky; Memphis, Tennessee; Kansas City, Missouri; and Tucson, Arizona). The dataset provides a continuous 15-year daily record (2007–2021) that integrates dual-source meteorological forcings (Daymet V4 and NCEP Stage IV), daily U.S. Geological Survey (USGS) streamflow observations, annual land-use and land-cover fractions, and a rich set of static hydro-geomorphological, soil, and anthropogenic attributes. By using hydrologically delineated chips rather than arbitrary square grids, the dataset offers spatially contiguous sub-watersheds within each city, facilitating user-defined reconstruction of physical flow topologies and upstream–downstream relationships. Target variables include a binary flood-occurrence label and daily streamflow expressed as both discharge and specific runoff, making the dataset directly suitable for spatiotemporal deep learning models such as Long Short-Term Memory (LSTM) networks and readily extensible to Graph Neural Networks (GNNs) through user-specified graph edges between chips.