LSTM input data: DROMEDARY basin sample with dynamic land use (Francois et al., 2026)

This resource contains NeuralHydrology-style inputs for continental-U.S. streamflow modeling with time-varying cropland/land-cover (CDL) fractions and static basin attributes. It supports the LSTM experiments that pair DayMet meteorology, USGS streamflow, Gages II basin descriptors, and USDA NASS Cropland Data Layer (CDL) categories aggregated to each basin.

Contents overview

The basin list included here has 3246 gauges (see time_series/list_DROMEDARY_basins.txt).

time_series/

Files

• {USGS_ID}.nc — NetCDF-4 dataset for basin USGS_ID (8-digit USGS streamgage identifier, zero-padded where applicable). There is one file per basin in the sample.
• list_DROMEDARY_basins.txt — Text list of all basin IDs (one ID per line), matching the expected .nc filenames without the extension.
• list_all_gages.py — Optional helper script: if run from inside time_series/ after all .nc files are present, it scans *.nc and regenerates list_DROMEDARY_basins.txt.

Temporal coverage and coordinate

• Time dimension: date — daily timestamps from 2008-01-01 through 2023-12-31 (inclusive), aligned across variables.
• Spatial unit: Each file corresponds to a single gaged catchment; variables are basin averages (meteorology from basin-averaged DayMet; CDL classes as percent of basin area).

Variables in each NetCDF

Meteorology (DayMet, basin average) — variable names in the file:

Units follow the DayMet product conventions for gridded DayMet variables (see the DayMet documentation for exact units and definitions).

CDL-derived land use / land cover (daily time series) — each field is the percent of basin area in that CDL aggregate class (the classes sum to 100% at each time step). Values come from annual CDL summaries; on the daily time axis they stay fixed for long stretches and change roughly around the turn of the calendar year (end of December into January). Variable names use underscores (e.g. Grassland_Pasture, Developed_Open_Low for “Grassland/Pasture”, “Developed Open/Low”).

Corn, Cotton, Rice, Sorghum, Soybeans, Oilseed, Barley, Spring_Wheat, Winter_Wheat, Other_Cereals, Alfalfa, Other_Hay, Nuts, Peas_Beans, Tree_Crops, Melons, Berries, Herbs, Roots, Vegetables, Double_Crops, Aquaculture, Fallow, Developed_Open_Low, Developed_Med_High, Forest, Wetlands, Shrubland, Grassland_Pasture, Open_Water, Perennial_Ice_Snow, Barren

Streamflow (target)

attributes/

File: static_attributes.csv

• Rows: One per basin; the first column is the USGS basin ID (same identifier as the {USGS_ID}.nc filenames). When reading with pandas, use index_col=0 and treat IDs as strings (zero-pad to 8 digits if needed).
• Columns: Concatenation of:
• Gages II basin characteristics (e.g. drainage area, gage coordinates, dam density, withdrawals, hydro modification and morphology indices, elevation, slope, soil hydrologic group fractions HGA … HGVAR, and encoded basin CLASS).
• Climatology computed from DayMet basin-average daily time series over 2008–2023: long-term mean annual precipitation and PET, aridity index (PET / precipitation), precipitation seasonality (PREC_SEAS), snow fraction (SNOW_FRAC), and metrics of high/low precipitation frequency and duration (HPF, HPD, LPF, LPD).
• CDL snapshot for a single reference year (2015): the same cropland/land-cover class columns as in the dynamic NetCDF files, representing percent of basin area for that year (used as static inputs in configurations that do not feed the full daily CDL stack).

Column names match the headers in the CSV (e.g. DRAIN_SQKM, LAT_GAGE, LNG_GAGE, DDENS_2009, … through the CDL aggregates). For soil-group codes (HGA, HGB, …), percentages describe the share of each NRCS hydrologic soil group within the basin (see Gages II documentation).

Data provenance (short)

• Streamflow: USGS NWIS (processed Gages II–style daily extracts used in the project).
• Meteorology: DayMet, spatially averaged to the basin (basin-mean CSVs / workflow described in the repository).
• Static physiographic attributes: Gages II basin characteristics (USGS Gages II).
• Land cover: USDA NASS Cropland Data Layer (CDL), aggregated to catchments and organized into the crop/land-cover classes above (https://www.sciencedirect.com/science/article/abs/pii/S0168169912000798?via%3Dihub).

Using this dataset with NeuralHydrology

• dataset: generic
• data_dir to this dataset root
• dynamic_inputs, static_attributes, and target_variables consistent with the variables above (training may use a subset of the columns present in the files).

Point data_dir to the folder that contains time_series/ and attributes/ as siblings.

Citation and contact

When publishing this resource (e.g. on HydroShare), cite the associated paper / DOI and this dataset’s HydroShare identifier once it is assigned. For questions about field definitions or preprocessing, email baptiste@tova.earth