|
Fig. 1 | Percentage of headwater streams by length in level 4 HydroBASINS
across the globe using the MERIT Hydro-based stream network (with a 5 ha
drainage threshold) as used in the Hydrography90m global hydrography
dataset.
Headwater streams are operationally defined here as Strahler stream
orders 1 and 2. For level 4 HydroBASINs data and the Hydrography90m global
hydrography dataset, see ref.
116
and ref.
103
, respectively.
Fig. 2 | Comparing observed and predicted headwater flows with those of
larger rivers, with flows normalized by area. a
, The observed streamflow
(
Q
) from exemplar headwater streams and large rivers demonstrates greater
headwater streamflow variability and flashiness compared with the less variable
flows of large rivers.
b
, The daily model performance (Nash–Sutcliffe efficiency,
NSE) across drainage basin percentiles for data-driven DL long short-term
memory (LSTM) neural networks predicting streamflow (Kratzert et al.
117
using
CAMELS data; Ouyang et al.
118
using US Geological Survey Geospatial Attributes
of Gages for Evaluating Streamflow (GAGES-II data
119
)) and US-based national
scale process-based models (NWM
92
and the NHM
93
). Lines of best fit are shown
for each of the four models, and drainage basin percentiles demonstrate a
comparable generalized summary of model results. Model performance (NSE)
in both the LSTM models and process-based models improves with increasing
drain ge area (here, drainage basin percentiles). One exception is the largest
20% of basins, where heavily managed flows are not well reproduced because of
the complexity of coordinated water storage, release, transfer and diversion. For
detailed results, see Supplementary Figs. 2–4.
Fig. 3 | Percentage of headwater streams by length in US Geological Survey
Hydrologic Unit Code (HUC)12 watersheds across the conterminous United
States.
Headwater streams are operationally defined and mapped here as
Strahler stream orders 1 and 2, based on the NHDPlus High Resolution V2
33
. State
boundaries were derived from the US Census Bureau
120
. Despite using different
base layers, Figs.
1
and
3
similarly demonstrate that headwaters generally account
for >70% of watershed stream networks. However, while similar in most regions,
comparisons between Fig.
1
and Fig.
3
also illustrate physiographic differences
in the landscape, information that may not be fully captured when using a flow
accumulation threshold approach to the stream network as in Fig.
1
.
Fig. 4 | Percentage of US Geological Survey (USGS) stream gauges across the
conterminous United States with at least 5 years of recent data (2018–2023)
that are considered headwaters, as operationally defined by Strahler stream
orders 1 and 2, based on the NHDPlus High Resolution (V2) dataset.
The stream
gauge locations were derived from the National Water Information System
121
.
The flowlines in the figure are from NHDPlus High Resolution data and are from
stream orders ≥7 for graphical purposes. The state boundaries were derived from
the US Census Bureau
120
. The NHDPlus High Resolution (V2) dataset can be found
at ref.
33
. The currently operating USGS stream gauges in this figure all have an
end date after 2019 with at least 5 years of data.
Fig. 5 | Simple conceptualization of data availability balanced against
hydrological process heterogeneity at different scales of flow regime
modelling.
We are required to capture a relatively dense level of process
heterogeneity (compared with catchment- and basin-scale models) to get an
accurate headwater flow regime model, yet the spatial density of data required
to do this is limited—except for a handful of highly instrumented headwaters,
nationally and around the globe. For our purposes, plots are small, highly
instrumented parts of the landscape unrelated to drainage areas; hillslopes are
sloped areas of the land draining to streams; headwaters are defined herein;
catchments are small-to-medium drainage areas or watersheds (~1–1,000 km
2
);
and basins are large drainage areas >1,000 km
2
).
|
