Dynamic Stream Permanence Estimates at Local and Regional Extents

Video Transcript
Download Video
Right-click and save to download

Detailed Description

Speaker: Konrad Hafen, USGS 

Topic: Dynamic Stream Permanence Estimates at Local and Regional Extents 

Abstract: In the United States (US), the frequency and duration of surface water in a stream channel (i.e. stream permanence) determines if a stream is subject to regulation under the Clean Water Act. The most comprehensive dataset of stream permanence classifications for the US is the National Hydrography Dataset (NHD), which has been shown to exhibit high rates of disagreement with in situ stream permanence observations. Analysis of NHD stream permanence disagreements with in situ observations indicates that differences in climate conditions between observation years contribute to the NHD disagreements and supports the need for dynamic simulation of stream permanence based on climate. In this study, two differing hydrological models were implemented to generate dynamic stream permanence estimates at a regional and local extents. In the Pacific Northwest (PNW) the Thornthwaite monthly water balance model (MWBM) was implemented to generate stream permanence estimates for headwater streams in the NHD network from 1977-2019. In the HJ Andrews Experimental Forest and Willow and Whitehorse watersheds of Oregon, the Watershed Erosion Prediction Project (WEPP) hydrologic model was applied to simulate stream permanence from 2011-2017. On 40% of PNW headwater streams no MWBM parameter sets were greater than 65% accurate when compared to observations. On 60% of headwater streams the MWBM dynamically simulated stream permanence with varying precision. Additional data would encourage and inform further model development. WEPP simulations produced daily stream permanence accuracies up to 93% and annual stream permanence accuracies up to 87% but different parameter sets performed better for daily and annual time steps. These results indicate that, when implemented for stream permanence simulation, assessment of physically-based models should include both daily and annual accuracies. Additionally, future stream permanence data collection methods and methodologies should be strategically planned to capture the spatiotemporal dynamics of stream permanence that are required to effectively develop and evaluate models.



Date Taken:

Location Taken: OR, US


<Al Rea>Hi everybody, this is the Hydrography

Community Communications Call.

Just in case you're confused by the name

we used to call this the Advisory Call,

but there's some special rules about

advisory committees and federal government,

so we changed the name just to make it clear

that's not what this is.

I'm Al Rea I'm the Co-lead for the

hydrography program at USGS. And, uh.

Wanna introduce Konrad Hafen who is,

He's he's been with USGS for how long

three or four years now. Konrad,

I can't remember. <Konrad>four years I think. <Al Rea>yeah.

and he was working on his PhD.

He just recently finished.

Congratulations Konrad. <Konrad>Thank you.

He got his PhD. at the University of Idaho.

And he's gonna be presenting about

a lot of the research that he did

as part of his dissertation. Uhm

and he's now with the Idaho Water

Science Center as a Hydrologist. And

I'll go ahead and let you take it away,

Konrad. <Konrad>All right thanks Al.

So I worked on stream permanence

and stream permanence modeling

for my dissertation work,

and I'm going to present on that today,

like Al said, and so I'm considering

stream permanence and dynamically

permanent estimates at both

regional and local extents. And

I'll give you just a kind of quick overview

where this presentation is going to go.

I'm going to talk about kind of the

importance of stream permanence and why

why we're focusing on that.

And then I'm going to go into kind

of three case studies or three

different scientific studies.

One is the accuracy of existing stream

permanence classifications within

the National Hydrography Dataset.

The second is modeling stream

permanence on headwater streams

in the Pacific Northwest that are

associated with the NHD network,

and the third is implementing

a physically based model

at a more local scale to simulate

stream permanence and I'll kind of pull

in the broader context of these

contributions and what we can do

and expect as we move forward

with stream permanence modeling.

To start off I

want to give you an example from

my hometown in Santa Clara, Utah,

down in the southwest corner of Utah.

And this is the Santa Clara River.

you see here in the photo. And this is around

the time of peak flow and about the

magnitude of peak flow in a normal year.

The Santa Clara River goes dry

some point almost every year.

But it's also a desert river and

can have some flashiness to it.

So you can see the peak flow here.

In late 2004 there was a flood

event on the Santa Clara river.

You can see here December 24th.

this is a pretty substantial flood

for the river and by about two weeks

later that flood had continued to

grow to a very dangerous stage,

and in fact many others much

property damage in many homes

were lost throughout the region

in early 2005. And so this

just illustrates kind of our

complex relationship with water.

How we need enough but not too

much and we want to know when we're

approaching either end of that where

we have low flows, where we have high

flows. And the USGS gauge network was

developed largely for those purposes.

And this just shows a snapshot

of water conditions throughout the

nation that was taken this March. And

the purposes of this gauge network,

there are three primary purposes

that they mentioned in their report:

Water management,

so making sure we have enough

water and know how much water we

are is going to be available;

Design flood flow so we can plan

for floods and plan for safety;

and then to assess climate related

trends that may affect future water

availability and future floods.

The things and if we go to

more of the policy side and

determining how waterbodies are

regulated under a kind

of a different set of rules,

and so if we look at the Clean

Water Act jurisdiction

and where the Clean Water Act applies,

we're going to kind of focus on

two categories here, and the...

the regulatory information.

So much more nuanced,

just kind of breaking this down

to these two simple categories.

The first traditional navigable waters,

which include large rivers and lakes,

and the second includes perennial and

intermittent rivers and streams that

contribute surface flow to navigable waters.

So if we look at this,

the regulatory status depends

on the duration of surface

water presence is defined by intermittent

and perennial rather or yeah perennial

intermittent rather than

the magnitude of flow that we see

happening in those rivers and streams.

So we talk about stream permanence

that's what I'm talking about.

Is how often is surface water

present in the stream channel?

Now I'm sure you've heard

the definitions of perennial,

intermittent and ephemeral,

and depending on where you

look at those definitions,

there could be a little bit of variation.

I'm going to focus on two other definitions,

permanent and nonpermanent. Permanent

encompases streams that have continuous

surface water present throughout a

year and nonpermanent are any stream

channels that experience dry

condition anytime during the year.

So you can group perennial into the

permanent category and intermittent and

ephemeral into the nonpermanent category.

And I'm sure I'll probably

slip up and using

ephemeral or perennial at some point,

so just know that I'm referring

to the permanent nonpermanent

categories as we go on here.

And so we're talking about this.

We could have these two

competing measurements.

We have the magnitude,

or how much flow there is,

and then we have a flow state

whether flow is present or absent,

and we could...

we could pose the question which

of these is more important?

Is it more important to know

the difference in the magnitude

of flow or to know if there

is flow or there isn't flow?

And for different applications,

there's going to be

I guess either one of these

measurements could be better

than another especially water

availability or flood

analysis is more important

obviously to know the magnitude

of flow. When looking at ecology

of the water quality the magnitude

of flow can be important,

but also the presence of flow is

important and then for regulatory

regulatory determinations

and habit identification,

we might be more interested in the

presence versus absence of flow.

And our USGS gauges can kind of

get us to those top two breakups,

but don't get us as well,

at least spatially continuous data down

in that lower presence of

water classification.

Part of the reason for this is that

stream gage coverage is limited,

so this figure here is showing stream

gages in the Pacific Northwest region,

and the colors represent stream

gages on different stream orders.

So the light blue is first

order streams and up to the red,

which is on 9th order streams.

Now the red

Outlines here on this graph show

the proportion of the stream network

according to the medium resolution

NHD that occur in each

of those stream orders.

And so the main takeaway here is that

if we look at 1st and 2nd order streams,

a much larger portion,

ugh excuse me-

Looks like we've got our

slideshow exit out there-

but a much larger portion of the

stream flow or a much larger portion

of the stream segments occur on

these first and second order streams

then we have stream gauges.

And so they are underrepresented

in the gage network,

which is simply an artifact of the

priorities of that gage network,

which is not to monitor for

those ecological purposes.

So the National Hydrography Dataset

(NHD) as we all know it is

the best source of data for a lot

of these headwater streams and

the hydrography or the hydrology.

there is described by these

perennial or nonperennial

intermittent,ephemeral stream

classifications that we have from the NHD.

The one shown in here is that

the NHD presents these single

static classifications of perennial,

intermittent or ephemeral

but stream networks,

dynamic and discontinuous,

and this is some work from Godsey

and Kirchner in 2014,

and they surveyed streams in multiple basins,

and this is a snapshot from one of

those basins and you can just see that

depending on the time of year and the year,

the flowing extent of the stream

network changed,

and so that's not completely captured

through those static NHD classifications.

This presents some potential problems

for regulatory mapping because,

as we know from the USGS,

NHD is the most up to date and comprehensive

hydrography dataset for the nation.

But from the EPA,

the existing tools cannot accurately

map the scope of Clean Water Act

jurisdiction. And so in order to if we

want to map Clean Water Act jurisdiction

we need to find some new tools to

account for some of that discontinuity

and dynamism in stream networks. So

that brings us to well, what next?

How can we get there?

And I want to focus on on two things

that can maybe help us improve

our stream permanence modeling.

So the first is, since NHD is

not adequate for mapping Clean

Water Act jurisdiction,

can we find where and when an NHD

disagreements occur and find out

what might be causing those and

use that to springboard forward to

some modeling and then can we use

process based hydrological models to

affectively simulate extreme permanence?

So I first going to talk about some of the NHD

agreement and disagreement I mentioned

a minute ago and then move on to two

different modeling applications,

one modeling in the headwater

streams in the Pacific Northwest

and wanted at a more local extent.

So first let's talk about some factors

that can influence NHD agreement/

disagreement with other observations.

and I'll just note there's a

citation here, we published

this work about a year ago, so you can

you can see that publication if

you're interested in some of the

in some more details there.

So the National Hydrography Dataset

both in medium resolution and the

high resolution hydrography derived

from USGS Topographic maps and you

can see an example of one here

showing part of the University of

Idaho Experimental Forest up in northern Idaho.

And you can see their perennial streams

and nonperennial streams marked

those nonperennial streams encompass

both intermittent and ephemeral streams.

And those were digitized or processed

via various means into the network

we see today which has at which is

a digital representation of those

maps and we could see perennial

nonperennial streams here for a

different portion of that forest.

So the question we have is, well,

how accurate are these classifications?

They were made by survey crews who actually

observed conditions on the ground,

but often during only a single visit.

Other field studies have observed the

50% disagreement with NHD classifications,

and those and those disagreement

rates have been shown to vary

between different climate regions.

These studies often focus on smaller

study areas over larger extents,

and so it's a little uncertain exactly

what's causing some of those variations.

So objectives here were to quantify

agreement between the NHD stream

permanence classifications and other

observations using a large regional

dataset and then determine the effect

of climate differences between when the

NHD classifications were determined

when other observations were made

and how that affects the

probability of those two datasets

disagreeing with each other.

So the general approach was to compare

these other observations other stream

permanence observations to the NHD,

stream permanence classifications,

and link those to the Palmer Drought

Severity Index to describe the

climate at the time the NHD was,

NHD classifications were made and

the other observations were made,

and to determine how that difference

affects the probability of the observations

of disagreeing with each other.

The first step we took here was

to just create a map to visualize

the years that these topographic

map surveys were conducted. And

a couple of takeaways here.

This shows basically a one decade

period for each of these.

These symbol classes and a couple

of takeaways here is you'll notice

there are some regions like we

can see some over here in Idaho

where we have topographic map

surveys being conducted multiple

decades apart for adjacent maps.

And so that just indicates that

first the climate,

the climate conditions during

those surveys for adjacent maps

could be quite different because

they were conducted decades apart.

We can further take a look at this

when we take a look at the PDSI,

the Palmer Drought Severity Index

during the survey year. If

you take a look at some of these

and just to explain the scale,

little bit of value of negative two

indicates a period of drought of

negative two or less and a value of

negative four or less indicates a

period of extreme drought and you

can think the opposite for those

positive values where it's wetter

than normal year or an extremely

wetter than normal year.

And so we map this out,

you can just see the adjacent

maps were collected,

some of these,

like there's a couple of Idaho here.

You can see some over in Kentucky

here and throughout, throughout the

United States where pier where quad

maps that were surveyed in a very

wet year are adjacent to quad maps

that were surveyed in a very dry year.

And as the NHD classifications

derived from those,

we could expect just some variation based

on climate trends to come from that.

I want to give a shout out to the

people who provided data for this study.

So here I'm showing a map that

includes about 10,500 observations

of surface water presence in streams.

A lot of these came from the Idaho

Department of Environmental Quality.

Some came from state agencies in Oregon,

Washington, Tribal nations,

and some other Federal agencies,

including the EPA.

So each of these observations just

marks the location and classifies that

location to either have surface water.

present in the channel or absent in the

channel. And these were the observations

we compared to the NHD stream

permanence classifications.

This dataset is published as the

USGS data released this link below.

If you're interested to see those

data or use them for your own studies.

So we then took the NHD stream

permanent classifications and

link them to the nearest points,

did some quality control,

and were able to find out

where disagreements occurred.

And so here I'm showing you overall

disagreement between both versions:

the NHD medium resolution

and the NHD high resolution,

and so the takeaway here is that we

saw about 80% agreement between these

in situ observations made throughout

the Pacific Northwest and the NHD

stream permanence classifications.

But we did also see some disagreements,

and so we want to focus on where

those disagreements occur and

why they might be occurring.

Now with my stream order we start to see get

an idea of where these disagreements occur.

So here we have the number of observations.

The ones that agree and disagree

on the Y axis on the X axis.

I've broken this out by stream order. So you

can see that the highest disagreement rate,

which is about 45%,

occurs on 1st order streams,

followed by second order streams.

And after we get out of those,

those lower stream orders those disagreement,

the disagreement rate decreases.

And this is only for the high resolution

NHD, I didn't do this for

the medium resolution


And so that breaks out

kind of the where part.

The next question we have is how climate

is affecting these disagreements.

So if my observation, if I go

a stream during a wet year,

but the NHD classifications

made during a very dry year,

what is the probability that those are

going to disagree in this ground play

play in effect in these disagreements?

And that's why I pulled in the PDSI

calculated the difference in PDSI

between those two observation dates.

And I'm going to walk through

this chart here with you,

and then I'll show you this for

multiple stream orders. And so here we're

looking at this for 1st order streams,

and these are the results of

logistic regression model.

So let's start on the X axis here.

So on the X axis we have the PDSI

difference. And this just indicates

whether a field observation and

non NHD observation

was made in a wetter or drier year

than the NHD classification by

the field survey crew was made.

So on the right side of this axis

we see that the NHD was determined

in the wetter year than the field

observation or in other words,

the field observation was collected

in a drier year than the NHD.

And so we're looking at the far

right side of its axis,

this indicates where the difference

was very great and the field

observations are much drier.

And if we look on the left side of this axis,

that indicates it very great difference

between climate periods and the

field observations are much wetter

or the NHD was collected in

a much drier year.

And now you'll notice that as we look

at this graph that we have two lines.

We have a blue line in a dry and red line.

The red line indicates the probability

of a dry observation disagreeing

with an NHD classification.

That blue line indicates the probability

of a wet observation disagreeing

with the NHD classification.

And so you can see that as those

wet observations were collected

in very wet NHD years,

the probability of disagreement decreased.

And when those dry observations

were collected,

were matched up with an NHD stream observation

into very wet year,

there's a higher probability of disagreement.

And I'm going to move on here.

We can take a look at this for

all stream orders,

and we see these same strip the

same trends for all stream orders,

just with different magnitudes

for wet and dry observations

and obviously different different

levels of confidence

based on the number of observations

we have. And so this indicates that

this agreement is affected by the

climate conditions that are being

experienced when the NHD surveys

were made and when subsequent

field observations were made.

And so a couple takeaways from this.


we see greater disagreement

on low order streams.

And the climate is playing an effect

on when these disagreements occur.

And so this just gives the indication

that if we add climatic variables

to our stream permanence modeling,

or given the case that we need to

add climatic variables to improve

our stream permanence modeling,

and that's kind of these next

two studies will focus on is,

there's a couple different ways

to to include those variables.

So this study is going to focus

on modeling stream permanence

in headwater streams,

with the water balance model

in the Pacific Northwest.

I talked about how stream networks

are dynamic and discontinuous and

the point of this study is to try

on a monthly basis to capture some

of that dynamism and discontinuity.

I'm showing here an example of

a statistical model that does this,

called the probability of stream permanence

model or PROSPER.

And I'm showing these the PROSPER

outputs for the same

area where the NHD is shown and

right now on this slide those

estimates are shown for 2004,

which is a drier year going to be

to 2011 a wet year and 2015 to

dry year and this just shows how

a model can capture differences

as we flip through those between

years which the NHD stream permanence

classifications do not capture.

Now this is a statistical model.

We're looking to do this more with

the process based model,

which gives them more flexibility

in terms of predicting these

things out into the future and

not being limited by parameter

estimates for some covariates.


NHD plus already has.

I guess you'd call the quasi

dynamic component if you're

familiar with the EROM values,

the enhanced runoff method,

those flow values.

and so to get those,

the NHD plus implements the

USGS Thornthwaite Monthly Water

Balance Model, I'm going to call that

the monthly water balance model,

or MWBM,

and it's averaged for monthly and

annual flow across a 30 year period.

And so if you ever seen maps like this,

and I'm not saying these were

made with those EROM values,

but the EROM values provide data

that can help you make maps

like this and show these things

somewhat dynamically with monthly

averages of flow. And so our idea

was to use this same model of

the similar approach to do this,

not averaged over 30 year

period but on a monthly time step

that are for discrete time, ugh

for discreet year.

I'm limiting the modeling only the

headwater streams for a few reasons.


that's where we saw the greatest

disagreement with the NHD classifications.

It's where the fewest USGS gages existence,

so we have that.

We don't know what's going on in

those headwater systems as well.

And so if we can get those to

be to be correct then we have a

better chance to get be applying

that model downstream

to other areas and it also simplifies

the modeling because we don't have to route water

down through the stream network.

So the objectives here determine if

the MWBM can generate dynamic stream

permanent estimates on the NHD network

with better or similar accuracy to

the existing NHD classifications

and then assess the precision of those

estimates from the model on headwater

streams in the Pacific Northwest.

So this is just a quick diagram of

the monthly water balance model.

The inputs here are temperature

and precipitation.

These bold abbreviations indicate model

parameters that adjust processes or values.

And I've simply added on a flow threshold

parameter that uses the mean monthly

runoff... mean monthly

you can think of as mean monthly

discharged to break out streams into nonpermanent

and permanent classifications.

So the general approach to this

study was to identify regions that

responded similarly to model,

to changes in model parameters and

model those regions with different

parameter sets and assess the accuracy

of those parameter sets to observed data,

and I use the same data set from the

from the first study but limited headwater

streams and then identified the

suitable parameter sets and assess

model precision in each of those

calibration groups and they should

make a little more sense as they go

through some of the results here.

So first this shows parameter

sensitivity and so basically what I'm

showing here is the effect the relative

effect that changes in each model

parameter have on stream permanence

classifications for each headwater

stream segment in the Pacific Northwest.

And just the important takeaway here

is to notice that these relative

sensitivities vary based on parameter

and that different parameters have

different sensitivities in different regions.

The point of this was to get these

sensitivities and used and I use it

unsupervised classification to group

stream segments headwater stream

segments together based on how similar

the sensitivities were forgiven stream reach.

And when I do that,

I come up with these eight calibration


I call them. And you can see that there

is some spatial correlation between these.

But there's also some of them span

the entire Pacific Northwest region.

A couple of things to note.

You can see how this Group A

kind of this pink

this pink color represents a

lot of mountainous regions.

This group 5 green represents

a lot of these wetter forests

that aren't high Alpine areas,

and then we give some of the

the pink that represents some

of these drier areas that come

through Idaho on the Snake River

plain on the Columbia Plateau.

But this was done completely numerically

using a K-means unsupervised algorithm.

And this breaks,

this just shows you kind of the

distribution of those calibration

groups throughout the study area and

the number of field observations we had

to associate with each one of those.

And then perform the calibration of

the monthly water balance model for

each of those eight calibration groups.

I tested 1,000,000 parameter

combinations for each of them and

then recorded the overall accuracy,

the accuracy of wet observations, and

the accuracy of dry observations

against each parameter set.

These are ah parameter sets

was evaluated against the NHD

against the accuracies from our previous

study to identify which parameter

sets get good results.

And so for a parameter set to be

deemed suitable it needed to produce

at least 65% overall accuracy.

And be at least 60% accurate with

both wet and dry observations.

Over here in this table

the ID shows the idea of

the calibration group

and shows the number of single

parameter sets identified for each

calibration group and these other

columns show the range of accuracies

represented by those calibrate or

presented by those parameter sets.

You'll notice that three of these

calibration groups 1,7,8 did not

have any suitable parameter sets.

So we take those and we

take those parameter sets

and apply them all to the model or apply

them all with the model to the streams.

We can get an idea of how

precise the estimates are.

So this is showing the average model

precision over the study period,

which is 1977 to 2019.

And when I say model precision,

the way this was calculated is I ran.

this I ran the model for each

suitable suitable parameter set.

So if I go back here for example,

you'll see that the group two

had 32 suitable parameter sets.

So I ran the model at each

stream reached 32 times,

with each of those parameters sets and

then took the proportion of those that

predicted the same classification.

So we come back over here.

A blue color or a high precision

value indicates that all or nearly

all of the parameter sets predicted

the stream to be permanent,

for a given month, or for a given year.

As we average those we get this value.

If we see a low or red value that

indicates high precision for a

nonpermanent classification,

which means that all or nearly

all of those pseudo parameter sets

predicted the stream reach to be nonpermanent

for a given time period or

group of time periods and then a low

precision indicates that about half

of the parameters that particular one

classification together half the other.

So we're going to show you how this

can show some dynamism as I go through

and show this for individual years.

So this is the average across

the whole study period.

If we look at a dry year 1987,

you can see we have much higher precision

ah for nonpermanent streams and not

many streams exhibit high precision

for permanent classification.

If we go to a more normal year,

we see some more precision for

those permanent classifications.

As we went to a wet year,

we see where we get high precision

for those permanent classifications

and how we have some decreased

precision into nonpermanent category.

I'm going to pull this back and

will look at this kind of a summary

for the whole Pacific Northwest.

And so this is showing the percentage

of headwater streams that had a high

precision value from these classifications.

And so if you take a look at the black line,

this is basically the sum of

the dark blue and the dark red,

and indicates the percentage of headwater

streams that we could say had ah a

precise estimate for each of these years.

You can see that varies between

about 28% and about 45%.

So on this headwater

streams this model produces precise

results for up to 45% of stream segments.

Now I didn't do a formal accuracy analysis

here because I only had about 2000

headwater observations to work with,

and I'm using those observations

to model 1.3 million stream reaches,

and so if I pull any of those out for an

independent accuracy assessment really

limits the calibration step of the model,

and so I'm going to get into this

here just a sec with with these.

But we really need more data to

appropriately estimate the accuracy

of some of these models

for stream permanence classifications.

Another thing to know with the monthly

water balance model should notice

that three of those groups didn't

have good accuracies and that could

be that the model doesn't represent

some processes that are important to

hydrology in those headwater areas

and also the model doesn't include any

kind of topography which can be a very

important influencer of stream permanence.

This indicates we might need to use

a more detailed model or make some

adjustments to the monthly water

balance model in order to have

more confidence into the stream

permamence estimates it produces.

So it's going to bring us to

this final this final section,

where I used kind of full fledged process

based hydrologic model to estimate

stream permanence in some smaller

areas where we had a little more data,

and so for this I'm using the Watershed

Erosion Prediction Project model,

which is called WEPP as an abbreviation.

It accounts for topography,


soils and climate at the hillslope scale,

and it's designed for to be used

in small watersheds,

and it's been implemented for both

erosion and stream flow studies

across a variety of environments.

Another reason I chose this,

it's really easy to implement using

a cloud platform developed by the

University of Idaho called WEPPCloud.

So anyone can go on for any area in the

United States produce a web simulation.

But the objectives here to evaluate

WEPP and WEPP's performance for

estimating stream permanence in both

gauged and ungauged watersheds.


and then a secondary objective here is to

calibrate WEPP and ungauged watershed

using stream permanence observations

instead of stream flow observations.

And so the general approach for this

was to set up initial WEPP models online

with WEPPCloud for our study areas,

then calibrate web to observe stream flow,



or stream permanence data depending

on the study area,

which looked into here in just a sec.

To calibrate or to evaluate the accuracy

of daily and annual WEPP stream permanence

estimates from stream permanence data.

So I used two study areas here.

The first was the H.J. Andrews Experimental

Forest, which is in Western Oregon.

In this water I used eight gauged

watersheds from the H.J. Andrews Experimental Forest.

These all have a gauge

record from at least 1993,

and many of them go back even into

the 1950s. These are small watersheds.

They range from 10 hectares, 200 hectares,

and in addition to the gauge data,

we collect history permanence

observations through human observation,

which these those are shown here

with these blue, yellow,

and red triangles to indicate

where surface water was or was not

observed in the summer of 2020.

And then we deploy thermistors where

we put a temperature sensor in the

stream and adjacent one on the bank.

We can get an idea when the

stream goes dry by comparing

the two temperature time series.

So those were deployed in addition

to the existing gauges in this area.

This is the Willow-Whitehorse

Whitehorse watersheds which are

over in Eastern Oregon and where

the H.J. Andrews receives over 2000

millimeters precipitation a year,

The Willow-Whitehorse watersheds

receive about 300 millimeters a year,

so we have a very humid and a

very arid environment to compare.

In these areas thermistors been

deployed from 2011 and 2017 as part

of a bull trout habitat study.

And so we used only thermistor data

that's shown at these locations

in these watersheds here.

I'm going to walk through the

H.J. Andrew simulations

first and go through the results there,

and then I'll walk through the

Willow-Whitehorse methodology

and results from there.

So I calibrated WEPP.

Two observed stream flow into

H.J. Andrews watershed.

And adjusted four parameters that

affected the shape of the hydrograph

and the water balance for for the model.

And then I evaluated those simulations

to find the best simulation based

on three goodness of fit metrics,

which include percent bias,

the Nash-Sutcliffe Efficiency,

and the Nash-Sutcliffe Efficiency

performed on the log of discharge.

and the reason I used the NSE,

the log NSE was because that gives

a better indication of how well

the model discharge fits observed

discharge for low flow periods in

those low flow periods are going to be

the period that determines stream permanence

when the screen goes dry.

I selected parameter sets based on

those goodness of fit metrics where the

percent bias might need to be less than 25%.

The Nash-Sutcliffe Efficiency greater than 0.3,

and when those two conditions

were met by maximized,

I chose the parameter system,

maximized the log NSE.

A stream reach was classified as permanent

for a year when the simulated stream

flow was greater than zero for all

days from April 1st of October 31st.

I only considered annual stream permanence

into H. J. Andrews because all those thermistor

observations which gave daily data,

they all came back as wet,

so they didn't give us enough data

to actually evaluate how well the

model doing daily timestep there.

So we model this with WEPP

you can see here the results begin to

walk through the legend real quick.

so the same legend throughout

the rest of the talk.

The the kind of lighter blue and

red colors indicate where the model

correctly classified a permanent

permanent stream is permanent and nonpermanent

stream as nonpermanent.

The dark blue indicates where a nonpermanent

stream was modeled as permanent

and a dark red indicates where permanent

stream was modeled as nonpermanent.

So with the model initially we

had 61% accuracy for these eight

watersheds in the H.J. Andrews.

But what we found is we

look at this little more.

We found the four stream

reaches in these watersheds

One and two were misclassified

by our data collection.

So our data collection,

when we observed,

made human observations at points,

they went to a point and just

observed if they could see continuous

surface water along the channel,

and there were points on each of

these that observed continuous

surface water and no points that

were discontinuous surface water.


that was artifact of our random

sampling design. So we talked to

H.J. Andrew staff

we found out that these stream

segments these four stream segments

actually go dry every year and

so the model did a better job of

representing it then our field

observations did that resulted in

accuracy of 83% in the H.J. Andrews.

We'll move on to the Willow-Whitehorse simulations

here I'm in the Willow-Whitehorse.

There were no stream flow gauges

to calibrate the model to,

so instead I calibrated the Model

to based in agreement with wet

and dry observations from those

thermistors were made daily.

Here in the Willow-Whitehorse watersheds.

Because there were fewer

data to calibrate on,

I only altered parameters that

affected the shape of the hydrograph,

and then the evapotranspiration.

And I calibrated this based on

daily and annual accuracy to

stream presence observations,

and then I included a dry day

threshold in the calibration as well,

which basically sets the maximum

number of dry days that can be allowed

or or permanent classification,

and so if we have a dry day threshold of 1.

That means a stream gauge

can have one dry day,

but still be classified as

permanent for the year.

And then I evaluated those parameter

sets and I'll present kind of

the findings from that here.

Just as I go into these next slides,

I'm going to just inform you of this

suggested accuracy metric I used.

And it's basically penalizing

the overall accuracy,

which is the percentage of

correct classifications,

and adjusting that by the accuracy

with wet observations and the

accuracy with dry observations.

And the reason for that

is if we take a take a look

at some of these streams

some of these streams might

go dry for 20 days a year,

but if we're looking at a

200 day period of record,

we can still have 90% accuracy just

by classifying everyday as wet and

so 90% is a high accuracy value.

But it did not capture all

that dry period or so.

Using this is just accuracy

penalizes the model results

if we don't get both wet and

dry observations correct.

OK, so I'm going to walk through

these graphs 1st and then I'll

come back over to the table.

So on the Y axis here I'm showing

an accuracy value and those accuracy

values are for dry accuracy,

what accuracy and the adjusted

accuracy I just described.

On the X axis is the overall accuracy.

So this is going to

show a couple of things.

One is going to show how wet

and dry accuracy differ as the

overall accuracy increases and

how adjusted accuracy

kind of mirrors that overall accuracy

and it's also going to show you

kind of what the maximum accuracy

for each of these basins is.

And so we can see if we just take a

look at the Wiilow-Whitehorse watershed

number one up here at the top that

as the overall accuracy increases,

it's driven by an increase in the

accuracy with which observations

even though we see a very high

decrease for very extreme decrease in

the accuracy of observations and

the adjusted accuracy kind of mirrors that.

So they just accuracy peaks right here.

Do you know around .65 or around 65%.

And we see that kind of same pattern with

a different peak for all four of

those watersheds.

And this is for for daily accuracy,

so this is using daily observations

as the accuracy metric.

If you come over here and look at this table.

We can see for each watershed the

accuracy of those daily observations.

And so for watershed 1 63% accurate,

watershed 2 66%, watershed 4 64%,

at watershed 3 64%, and watershed 4 82%.

Now I'm a takeaway here is that

even though we had relatively high

daily accuracies,

if we look at the annual accuracy

to be predicted extreme to have

water all year or not,

we see those annual accuracies generally

are much lower than the daily accuracies.

So that indicates these daily

accuracies aren't capturing.

When extreme goes dryer,

when it stays wet very well.

Now we can take a look at kind

of the same graphs,

but doing this for annual accuracy.

And you'll notice the accuracy

values are a little better for

quite a bit of watershed 1

and about the same for watershed 4


lower watersheds 2 and 3.

I'm going to move on here and

go into this this dry day

threshold which adjust for that a

little bit. So we added a dry day threshold.

We can see a little bit of increase

in those accuracies getting up to 60

percent by about 2% for two and three,

and then by 10% for watershed

four and so adding that dry day

threshold for allowing a certain

number of dry days for the stream

to be permanent makes some sense

for some of these watersheds.

Ok now I'm going to summarize

these results for for a couple of

watersheds across the entire period.

This kind of kind of give you just an

overall idea of how these accuracies look,

and we can kind of break

this out a little more.

And so this shows the accuracies

annual and daily accuracies for each

of the stream reaches each of the

sample locations in watershed 1.

And so the daily accuracies sorry annual the

the annual accuracies are represented

by the colors in the grid.

So once again,

the lighter colors of blue and red

indicate a correct classification

of permanent stream classification

of a nonpermanent stream.

And the dark

red indicates a permanent stream was

modeled as nonpermanent and the dark

blue indicates nonpermanent stream

that was modeled permanent.

Then the numbers inside each

of those squares indicate that

daily accuracies for that year.

And I want to draw your attention to

a couple of these disagreements here.

So first in this dark red,

we see that even though we had

high daily accuracies for these

two stream reaches,

the model produced incorrect

classifications annually,

and so these were streams which

were permanent.

but were modeled as nonpermanent. And

we see the same thing down on this

one up for a nonpermanent stream

that was modeled as permanent

even though 83% of the days in

that year or 95% in this case

were classified correctly.

And then we went to watershed 2 and

show the opposite thing happening,

where these two stream reaches,

even though only you know 15 or

16 or 21% of those days were

classified correctly,

we still got the overall

annual classification correct.

This brings up some.

I think some important considerations

as we consider models in

evaluating stream permanence models.

There's a couple of takeaways

from the (inaudible).

Our current data collection methods and

current data may not actually describe

which reach-scale stream permanence well.

From Willow-Whitehorse,

we learned that it's important to

assess stream permanent models on

both daily and annual metrics.

Still getting the full picture

of what's happening with that

model at those stream reaches.

And there it raises a lot more questions

about how well do stream permanence

calibrations simulate stream flow

with the variation we're seeing,

and it would be interesting to

do this in the basin where,

we have some gauges and some more

stream permanence data to identify

if we can actually usually permenance

classifications to calibrate the stream flow,

or those data aren't available.

I want to wrap this up with just a

few a few thoughts made pulling in

some other other ideas for context.

But stream permanence is easy to observe,

but it can be difficult to simulate.

It's really easy to see if

there's water in a stream or not.

We have to keep building a

model (inaudible) stream permanence,

but we're somewhat limited by the

data we have to evaluate those models.

We can evaluate them in specific locations,

but evaluating these are regional

or sometimes local extents.

There just aren't enough spatially and

temporally then stated to get a good

picture of how to calibrate those models.

Part of that can be due to how

difficult is characterize

subsurface and channel characteristics.

For these applications were talking

about flow disappearing right.

We have a very very small amount

of flow and what causes that flow

to disappear and that could be those

could be very very small things that

are hard to capture in these models.

And as I mentioned before,

both temporal and spatial variability are

important as we consider these models,

which makes a much greater burden

of data to calculate those things.

Second, we're limited by data

as I mentioned before,

but if we just take a look at the data

used for these modeling applications,

we can get an idea of the spatial

and temporal differences with

stream gauges. And so on

the left I'll show you the data

set used to evaluate the NHD,

and we have 10,500 discreet

observation locations.

Which is 5 times more than

we have for USGS gauges

Really 2100 USGS gauges in the same area.


those USGS gauges collect 3/4 of

a million observations per year.

But we're just collecting daily data,

but we only have 10,005 total observations.

With those stream permanence observations.

We need to increase the spatial and

temporal density of these observations

to get better data to inform models.

And part of this is going to

involve deliberately collecting

data at the reach scale.

I just want to use an example for

how data collection schemes can

affect how we evaluate models.

And so if we randomly select

data data sampling locations,

we could say that all these streams

in this basin are permanent if we

always collect data at these locations

just by the look of the draw we

can always find a wet location.

Or we could select allocations.

I think these were all dry except

for main stem,

when in reality we need to be focusing

maybe on reaches and not just point

so we can get an overall picture of

what's happening with these reaches.

And one way to do this would be

to focus monitoring workstream

regions are most likely to go dry.

Identifying those locations

could involve some more work.

So moving forward I think there's a lot

to be excited about as we know

the USGS is working on collecting

lidar for the entire nation,

which hopefully will result

to updates for the NHD.

Data collection apps are being developed

which is pulling in citizens for citizen

science and collecting more data,

and recently I've been seeing more

focus on data collection headwater

streams which will give us

those data once again to to help

calibrate and validate these models.

So thank you for listening today and

I just want to make sure that all

these folks here get acknowledged.

You've helped out with this project

and in various ways,

and if there are any questions,

I am happy to take those.

<Al Rea>Great Konrad, thanks a lot.

This is really, really informative.

We have some questions in the

chat that basically are kind

of getting at the question of

What do you consider permanent

versus nonpermanent in terms of like

You know, is it

Like you can have pools of water in a

channel that maybe not may not be flowing.

Is that considered permanent?

Is it not? And the other idea that

was expressed was kind of about

Water may be flowing underground,

either in like sandy channel or in a

like in the mountains where they may have,

like, you know,

say global (inaudible) marine where the water is

flowing down under a boulder field

and you can actually hear it, but,

Can't actually see it on the surface,

so kind of

As far as like the work

that you've done, what

What would you be considering

permanent versus nonpermanent

in in those different scenarios?

<Konrad>That's a good question,

and it highlights some of the

complexities of collecting stream

permanence data. For this application,

and because of those considerations

you just mentioned,

we we focus simply on surface water,

and so if there was water

visible above the surface,

that stream reach would be...

We and I guess when you get

the permanent, permanence,

implies also at a time scale, right?

And so these observations are

based on surface water presence at

one point in time, and so so for us

a permanent stream needs to point in

time would be a stream that has surface

water continuously throughout the channel.

Any stream where there were dry portions

channel visible would have been

a nonpermanent classification.

I didn't do a good job of explaining this,

but the way we determined that I guess

determined the annual stream permanence,

was any any dry observation or

any stream gage indicated that

stream was not permanent that year.

And then we we determined permanence

by limiting that to any observations

that said the stream had surface water

in August or September we'll

use that as an indicator

the stream was permanent throughout the year,

which the stream could have potentially

been dry the day before the day

after, a month after month before

those observations were made

but based on the data we had

that was the best way we could

to break those up to describe

annual condition.

<Al Rea> OK. Any other questions for Konrad?

He presented a whole lot of

whole lot of information.

<Jim>Hello Al. -Yeah. This is Jim Mitchell.

In 93 or 94, when I was at

the Kansas Geological Survey, I

saw presentation by the Kansas Water

Quality folks where they went and they

took all of the stream gauges in Kansas and

essentially blanked out

the rest of the upstream network

from all the gauges that went

to zero at some point during

the year back in the early 90s.

In about the Western 2/3 of

the state just disappeared.

Now, most of that isn't,

you know, like headwater.

These are like,

you know,

lower or higher order reaches if you will.

Our explanation at the time was,

you know the groundwater dropping out

of the bottom and no more base flow.

I just wonder if the mechanisms for

these permanents are nonpermanents,

you know, really are different kinds

of things depending on where you

are high in the watershed or low.

<Konrad>Yeah, I'm I'm sure they are,

and that's kind of

why were trying to explore some

of these process based models.

We can represent those

things in those models.

But I also don't think as

I kind of showed here,

we haven't got the point

where we can model this consistently

enough to identify how much of an

impact those processes may be having.

So that's kind of the goal that we

can get these models to develop and

have the data develop these models,

we can hopefully represent these

processes and find out why these

streams might be going dry or the

thresholds that might affect that.

<Jim>Another observation that was made

was that in the in the Spring after

things thawed out and we started

getting growth and crops and things in,

you know,

in the plains that the winter time

flow if you will that happened

in you know ditches and

low order streams.

You know started to disappear again

if evapotranspiration in that case,

like pulling water out of the soil

and not contributing to any flow.

I think there's a lot of

things going on here.

I think slope, aspect,

a lot of other stuff could affect this.


definitely definitely.

<Al>Right Konrad, right.

You know another thing that you know

I've thought about with this is.

Any of the models that you've

been working with,

they really can't take into a into

account some very localized geologic

factors like you know you have a

fault or something, that's just,

you know, essentially snatching

the water right out of the channel.

Or, conversely,

you know may have a perennial Spring.

On a channel that

wouldn't normally have flow,

but then you've got this perennial

stream spring that's contributing flow

to that Channel. so it's it's like.

none of the models that I heard

you talk about could get into that

sort of specificity (struggling)

on sort of local on a local scale.

<Konrad>Right, yeah.

And Al you're exactly right, and I

and that's why I think,

like we use these models

for stream flow, right?

These models perform really

well for stream flow in

a lot of instances,

but I think taking that step to stream


I think we're starting to learn

a lot about those things that

are influencing stream permanence by

seeing why these models are wrong.

And how wrong they are.

And then also being able to kind of say,


we figured out the simple reason

that this might be the case.

And can we,

can you know can we go back and

adjust these models or develop some

new models that are going to count

with these things that that we find

it be important for stream permanence?

And then I think also going back to

looking at data collection. We've used

you know, we've you stream flow data

for developering hydrological models

for so long that we're modeling

something new that has different

characteristics and we need different

data to evaluate our models on and

so going back and saying OK well.

How do we collect data

to inform these models?

As as we move forward?

So I think there's still things

we can learn from these models,

even though right now I don't

think these process-

based models are producing,

You know necessarily results that can

be used as far as mapping stream permanence,

I think we can use them to further

understanding and further our

methodologies for finding

permanence models.

<Al>Yeah, great. So thanks,

thanks again, Konrad, for

your presentation it's

been really interesting.

With that I think

we're right out of time here and

We'll call it a day for today.

Thanks again Konrad. We'll see you again next month.

Bye Bye.