PROSPER-rfops: Random Forest Operational Predictions of Streamflow Permanence for the Pacific Northwest (HU17)

The PROSPER-rfops_HU17 rasters provide annual estimates of the probability of streamflow permanence for years 1989 through 2022 in the Pacific Northwest region of the United States, produced by the PRObability of Streamflow PERmanence Random Forest Operational (PROSPER-rfops) model. The model domain encompasses 12 four-digit Hydrologic Unit Code (HUC4) boundaries: 1701 (Kootenai-Pend Oreille-Spokane), 1702 (Upper Columbia), 1703 (Yakima), 1704 (Upper Snake), 1705 (Middle Snake), 1706 (Lower Snake), 1707 (Middle Columbia), 1708 (Lower Columbia), 1709 (Willamette), 1710 (Oregon-Washington Coastal), 1711 (Puget Sound), and 1712 (Oregon Closed Basins). The PROSPER-rfops_HU17 is a raster-based empirical model whose outputs represent the probability that unregulated and minimally impaired stream channels in the Pacific Northwest maintain year-round flow. Model predictions are provided at a 10-meter spatial resolution and are assigned to pixel cells on the channel network defined by the High-Resolution National Hydrography Dataset Plus (NHDPlus HR) stream grid, using a minimum upstream drainage area threshold of 0.01 square kilometers. PROSPER-rfops follows the general approach of the PROSPER Upper Missouri (PROSPERum) model (Sando and others, 2022) adapted to the Pacific Northwest. Explanatory variables include monthly and annually aggregated climatic conditions and static physiographic variables accumulated upstream of each stream pixel as flow-conditioned parameter grids (FCPGs). Key differences from PROSPERum include: (1) the model domain is the Pacific Northwest (HU17); (2) topographic wetness index (TWI) and plan curvature are added as physiographic explanatory variables; (3) the annual National Land Cover Database (NLCD) product is used in place of the episodic NLCD, providing land cover values for every model year; and (4) the random forest model is trained using empirical field observations of streamflow permanence and implemented in Python using the scikit-learn library (Pedregosa and others, 2011) rather than in R. This data release operates within an active, operational framework and is updated annually with a one-year lag. The current release includes annual mean probability rasters and annual ensemble standard deviation rasters for each HUC4 and year, as well as temporal summary statistics rasters computed across the full 1989–2022 period of record for each HUC4.