Why one dataset of 6,830 catchments matters more to flood AI than any new model
Caravan stitched regional streamflow collections into one standard format. For machine-learning hydrology, that is the bottleneck it removed.
- Caravan bundles forcing, observed discharge and catchment attributes for 6,830+ basins in one schema.
- It merges CAMELS US, AUS, BR, CL and GB with HYSETS and LamaH-CE, so a model can train across continents.
- Data-driven rainfall-runoff models improve with more catchments; harmonised data is what makes "more" possible.
Most discussion of AI in hydrology centres on architectures. The quieter story is data plumbing. A neural network that learns how rain becomes river flow needs to see many rivers, and until recently those rivers lived in separate national collections with different variable names, units and time conventions.
Caravan, published in Scientific Data in 2023 by Kratzert, Nearing, Addor and colleagues, addressed that directly. It is an open community dataset that provides meteorological forcing, streamflow and catchment attributes in a single standard, designed for large-sample hydrological modelling.
What "large-sample" buys a model
The curated 4TU.ResearchData collection extends Caravan to 6,830+ basins by folding in CAMELS US, AUS, BR, CL and GB, plus HYSETS and LamaH-CE. For a model, each basin is another example of how soils, slope, climate and land cover shape a hydrograph. Training across thousands of them is how a model learns patterns that transfer to rivers it has never seen.
Try it yourself
Caravan is distributed as per-basin time series plus an attributes table. A first look takes a few lines of Python:
import pandas as pd
# One basin's daily forcing + discharge (Caravan netCDF/CSV export)
ts = pd.read_csv("camels_01013500.csv", parse_dates=["date"])
attrs = pd.read_csv("attributes_caravan_camels.csv", index_col="gauge_id")
print(ts[["date", "total_precipitation_sum", "streamflow"]].tail())
print(attrs.loc["camels_01013500", ["area", "p_mean", "aridity"]])What to watch
- Coverage gaps: the bundled collections lean toward the Americas, Europe and Australia. Regional extensions for Africa and South Asia would change what models can learn.
- Benchmarks: a shared dataset makes model comparisons fair. Expect more papers to report results on Caravan splits.