Photo: Don Buso
GAPS!
Why they occur, and what we can do about it.
Craig See, Ruth Yanai, John Campbell and Amey Bailey
Causes of stream flow gaps
25%
50%
6%
13%
6%
Equipment failure
Maintenance
Operator error
Ice in v-notch
Debris in v-notch
Misc.
68%
12%
8%
10%
1%
1%
Hubbard Brook Coweeta
Watershed 5 stream flow (ft3/s)
Watershed6streamflow(ft3/s)
Using lots of fake gaps to simulate
real ones
• For each real gap in the record,
randomly select a point in the
record to create a corresponding
“fake” gap:
o Same length of the real gap
o Stratify sampling by flow rate
o Calculate difference between
observed and expected
o 10,000 iterations to create an
error distribution
• For multiple gaps in a year, sum the
errors of all simulated gaps for each
iteration
Uncertainty in annual stream
export due to gaps
Wet year
~0.4%
Dry year
~1.9%
Advantages:
• Does not rely on statistical assumptions
(normality, homoscedasticity)
• Can be modified to fit most gap-filling methods
• Conceptually simple
Disadvantages:
• Requires a long-term record
• Hard to characterize extreme (infrequent) events
Code available soon!
www.quantifyinguncertainty.org
Uncertainty in annual stream
export due to gaps
-6.7% 6.7%
Annual stream flow uncertainty (mm)
Frequency
14.5% 14.5%
Wet Year (1996) Dry Year (2001)
Causes of precipitation volume
gaps at Hubbard Brook
Gaps in the Streamflow Record
days)
Hubbard Brook Coweeta

See asm 2015_gaps

  • 1.
    Photo: Don Buso GAPS! Whythey occur, and what we can do about it. Craig See, Ruth Yanai, John Campbell and Amey Bailey
  • 2.
    Causes of streamflow gaps 25% 50% 6% 13% 6% Equipment failure Maintenance Operator error Ice in v-notch Debris in v-notch Misc. 68% 12% 8% 10% 1% 1% Hubbard Brook Coweeta
  • 3.
    Watershed 5 streamflow (ft3/s) Watershed6streamflow(ft3/s) Using lots of fake gaps to simulate real ones • For each real gap in the record, randomly select a point in the record to create a corresponding “fake” gap: o Same length of the real gap o Stratify sampling by flow rate o Calculate difference between observed and expected o 10,000 iterations to create an error distribution • For multiple gaps in a year, sum the errors of all simulated gaps for each iteration
  • 4.
    Uncertainty in annualstream export due to gaps Wet year ~0.4% Dry year ~1.9%
  • 5.
    Advantages: • Does notrely on statistical assumptions (normality, homoscedasticity) • Can be modified to fit most gap-filling methods • Conceptually simple Disadvantages: • Requires a long-term record • Hard to characterize extreme (infrequent) events Code available soon! www.quantifyinguncertainty.org
  • 6.
    Uncertainty in annualstream export due to gaps -6.7% 6.7% Annual stream flow uncertainty (mm) Frequency 14.5% 14.5% Wet Year (1996) Dry Year (2001)
  • 7.
    Causes of precipitationvolume gaps at Hubbard Brook
  • 8.
    Gaps in theStreamflow Record days) Hubbard Brook Coweeta

Editor's Notes

  • #2 THis is a picture from one of the weirs at Hubbard Brook taken after an ice flow event. Its probably pretty obvious from the picture that the streamflow data at this time is probably unuseable. All longterm streamflow records contain gaps. I’m going to give a quick overview of a method ive been working on (or not working on) for the last few years on how to characterize the uncertainty surrounding filling gaps in the streamflow record. This is part of a larger project John Campbells been leading on estimating uncertainty in the net hydrologic flux of nutrients in small watershed studies.
  • #3 First I’ll talk a bit about where these gaps are coming from. This is a breakdown of the causes of gaps in the streamflow record at HB and Coweeta. Over half of the gaps at Hubbard Brook are due to equipment failure. As you can see, Equipment failure and Maintenance account for the majority of gaps in both sites. While there are things that can be done to cut back on these things, the reality is that some gaps are simply unavoidable.
  • #4 There are many methods that people use to fill gaps in streamflow. In paired watershed studies, by far the most common is some sort of regression model where you use data from a nearby stream to calculate the missing vaues. This is a simple linear regression from represent 15 minute observations from 1962-2001. WS5 on the X and WS6 on the Y. It looks messy until you consider that this roughly 1.5 milion datapoints. So the method we’ve developed is relatively simple: First, for every real gap in the record, you randomly go back into the record and create a fake gap. Obviously the “fake gap needs to be the same length as the real gap you’re trying to simulate. While the location of the “fake gap” is randomly selected, we do stratify the sampling by flow rate. (you don’t want to simulate a gap during a high flow event using a period of low flow). The results I’m presenting today we divided the entire record into octiles. Then you simply calculate the difference between the modeled gap and what was actually observed. You then have the computer do this many many times, and you end up with an error distribution for the gap you’re trying to simulate. If you want to calculate the error associated with multiple gaps over the course of a year, you simply sum the error associated with each simulated gap in each iteration. Sometimes they cancel eachother out, sometimes the error compounds.
  • #5 So then you end up with an error distribution for each year. So this is showing water years 1996 through 2009 on the X axis, and the error distribution in watershed mm on the Y axis. Calculating these distributions in terms of % error, I found it interesting that for a given year, annual discharge mattered as much or more than the number and duration of gaps . Because it was a wet year, the error in 1996 was a quarter of the error associated with 2001, even though the erro in mm wasn’t that much less.
  • #7 This shows the uncertainty (in mm) for each water year based on 1000 iterations.
  • #8 This shows a breakdown of causes of gaps in the precipitation record at hubbard brook. Here technician error and equipment failure make up most of the gaps. I should mention that precip volume gaps are exceedingly rare compared to other types, and the rain gauge network is pretty extensive, so there’s probably not much uncertainty introduced here.
  • #9 Craig See is working on a paper describing gaps in ecosystem monitoring data