Uncertainty in Utah Hydrologic Data: Part 1 On The Snow Data Set by Randall P. Julander

In response to the post on June 2 2011

Comments On Western Water 2011

We are very fortunate to have a three part set of guest posts by Randall P. Julander of the US Department of Agriculture. The first of the three part guest posts appears below. Part 2 and 3 will appear this week. Randall has presented a very well documented report on a variety of issues associated with water resources and analysis that applies not only to Utah but throughout the western United States.

 

Uncertainty in Utah Hydrologic Data- Part 1 The Snow Data Set

A three part series that examines some of the systematic bias in Snow Course, SNOTEL, Streamflow data and Hydrologic Models

Randall P. Julander Snow Survey, NRCS, USDA

Abstract

Hydrologic data collection networks and for that matter, all data collection networks were designed, installed and operated – maintained to solve someone’s problem. From the selection of sensors to the site location, all details of any network were designed to accomplish the purpose of the network. For example the SNOTEL system was designed for water supply forecasting and while it is useful for avalanche forecasting, SNOTEL site locations are in the worst locations for data avalanche forecasters want such as wind loading, wind speed/direction and snow redistribution. All data collection networks have bias, both random and systematic. Use of any data from any network for any purpose including the intended one but especially for any other purpose should include an evaluation for data bias as the first step in quality research. Research that links a specific observation or change to a relational cause could be severely compromised if the data set has unaccounted systematic bias.  Many recent papers utilizing Utah Hydrologic Data have not identified or removed systematic bias from the data.  The implicit assumption is of data stationarity – that all things except climate are constant through time thus observed change in any variable can be directly attributed to climate change. Watersheds can be characterized as living entities that change fluidly through time. Streamflow is the last check paid in the water balance – it is the residual after all other bills have been paid such as transpiration, evaporation, sublimation and all other losses. Water yield from any given watershed can be impacted by vegetation change, watershed management such as grazing, forestry practices, mining, diversions, dams and a host of related factors. In order to isolate and quantify changes in water yield due to climate change, these other factors must also be identified and quantified. Operational hydrologic models for the most part grossly simplify the complexities of watershed response due to the lack of data. For the most part they operate on some snow and precipitation data as water balance inputs, temperature as the sole energy input, gross estimations of watershed losses mostly represented by a generic rule curve and streamflow as an output to achieve a mass balance. Temperature is not the main energy driver in snowmelt, short wave solar energy is. Hydrologic models using temperature as the sole energy input can overestimate the impacts of warming.

Snow Course Data

The snow data set most often used to demonstrate declining snowpacks west wide is the NRCS USDA SWE data set that extends back to the late 20’s and early 30’s in many areas westwide. This data collection network was installed to collect snow data and predict water supply. It was recognized very early on (in the 40’s and 50s’ – Snow Survey Measurement Handbook) by individuals in the program – that vegetative impacts would have a significant impact on snow accumulation – specifically as vegetation increased or matured near snow sampling sites that snow measurements of SWE would decline. Methods would have to be developed to compensate for this natural phenomenon in order to maintain consistency in statistical relationships to streamflow. This is an example of systematic bias in a data set that has to be identified site by site and removed from the data set in order to evaluate the impacts climate change may have had on snow accumulation. These vegetative impacts can be substantial and we have documented some of them at our longer term snow courses in Utah.

http://www.ut.nrcs.usda.gov/snow/data/long-term_snow_data_comp.html

Buckboard Flat is one such example.

The NRCS snow course at Buckboard Flat in southeastern Utah is very representative of April 1 SWE decline due to changes in site vegetative cover. In our assessment of site changes (http://www.ut.nrcs.usda.gov/snow/siteinfo/data_bias/externalinfluencesonsnowpack-2.pdf) we tried to put an estimate of the SWE change at each snow site due to changes in vegetation, site characteristics, weather modification, pollution and sensor changes.

In this graph, one can clearly see that SWE has declined. The 30 year running average at this site has decreased from an average of 14 inches to 12 inches over time. Unfortunately, this is the conclusion of nearly all research on this issue whereas it should be the start – this is where researchers jump to Climate Change as the reason for decline. The fact that SWE has declined at this site does not mean that it is all/mostly/some or even related at all to temperature increases.

Buckboard Flat snow course may well be the poster child for dramatic vegetative changes over the past 80 years.  There is a huge compendium of research on the impact of vegetation on snow accumulation and ablation, particularly when comparing open meadows to conifer coverage.  There is up to a 40% decline in SWE accumulation between the conifers and adjacent open areas.

Potential weather modification: none.

This 1936 photo is looking towards the north east along the main axis of the course.

This photo is looking in a more southerly direction opposite of the previous photo.

In these 2 photos of the Buckboard Flat Snow Course taken in 1936, the course is relatively open with young growing aspens nearby. Some conifers are apparent in the background interspersed in the aspen forest as is some brush species in the foreground.  There is an old trail in the first photo along which the snow course is laid out.

This is the Buckboard Flat Snow Course circa early 2000. Notice that the tree cover has increased dramatically and that the aspens are now a mixed cover with at least 50% conifer coverage if not more.

In this photo, looking the other way and after some tree removal, one can still see the remnants of the old trail in the early photograph. Notice again, the number and size of conifers along the course.

This photo is taken perpendicular to the course.

Where we have observed snowpack declines and site or vegetative changes are suspected, we make every effort to document the impacts by not only measuring the snow course, but to measure each point perpendicular to the course as well.

At this course by just moving 10 feet to either side, snowpack is reduced by 10%. The denser the vegetation at the point, the greater the reduction in SWE accumulation.

Another good example of the impact vegetation can have on SWE accumulation is our snow course at Garden City Summit. This course was originally in an aspen complex that subsequently has become nearly uniform conifer. We have observed a 20% decline in April 1 SWE values over time. Recently we installed a SNOTEL site at this snow course in an area some 10 yards removed from the course and the conifers. The result in SWE accumulation SNOTEL compared to snow course was illustrative of the impact these conifers have had on SWE.

The new SNOTEL is accumulating exactly the amount of SWE that went missing from the snow course due to conifer encroachment. This is clear evidence that the decline in SWE observed at the snow course is more likely due to vegetation change than climate change.

http://www.ut.nrcs.usda.gov/snow/siteinfo/data_bias/Garden_City_Summit.pdf

Climate change and declines in observed April 1 SWE in Utah

The signal of climate change on SWE accumulation due to temperature increases should be manifest at the lower elevations and lower latitudes first, then migrating to higher elevations and latitudes as temperatures increase.  In our analysis (An examination of the external influences imbedded in the historical snow data of Utah, Proceedings of the Western Snow Conference) we could find no such consistent signal. In fact, after adjusting the data for as many biases as we could identify (including subtracting SWE for weather modification) we found no statistically significant changes in the long term April 1 SWE for Utah. As we looked at various sites we found we had to explain why a site at 10,000 feet in northern Utah was losing snow and an 8000 foot site in southern Utah was not. Even if all sites had lost SWE, the question is still relevant – did this site lose more than another and why? The most consistent variable was site or vegetative changes. If the site had significant vegetation change, it was losing SWE, if the vegetation was stable, then the site showed no trend in April 1 SWE regardless of elevation, aspect or latitude. Thus the impacts of increased temperature may not be currently reflected in the snow data of Utah and it suggests that the data from all other areas needs to be re-evaluated to determine how much of the observed declines can be attributed to climate change and how much might be from other sources. 

As a personal opinion – this snow course data set contains valuable information about long term climate trends associated with snow accumulation and ablation but must be used with caution and understanding – specifically of the long term systematic biases in the data. The continued use of this data set by researchers not familiar with the purpose and operation of the data collection effort and its associated biases needs to stop. Further, research already published should be re-evaluated for its accuracy. Peer review processes seem to overlook whether a data set is sufficient in its published form for the kinds of long term trend analyses being done. That should be one of the first questions asked – has the author examined the data set for systematic bias and what methods were used to identify and numerically remove that bias?

The SNOTEL Data Network

 

The SNOTEL data network was designed for the same purposes as was the original Snow Course network and in most cases was co-located at a snow course with the ultimate purpose to replace the same with automated sensors.  This network originally provided daily data and is in many locations now producing hourly values. It should be noted that a man with 2 watches never knows what time it is and that is the case with snow courses and SNOTEL. They are different means of snow data collection and while close, are not the same.   With both daily and hourly data, we have the means to determine both the onset of snow accumulation and the end of ablation or melt out dates. These metrics seem to be a readily available means to determine if snowpacks are melting earlier than in the past. Here again, familiarity with the data set is essential in order to ascertain what may be ascribed to climate change and what must be filtered out as systematic bias.

One of the first considerations of this system is that it is expensive and as such, was put in locations far more susceptible to vegetation change than some of the snow courses. When a snow course was located in an open meadow it is likely the co-located SNOTEL will be nearby but hidden in the trees. SNOTEL data, over time will likely be compromised by vegetation change faster than many of the snow courses. Sensor change is also an issue with the SNOTEL data set. The original installation utilized snow pillows made of stainless steel, about 4 feet by 5 feet. These were about 1 inch thick and each could hold about 10 to 15 gallons of antifreeze. Pillow configurations at various sites ranged from 2 to as many as 4 of these pillows plumbed together. These stainless steel pillows were notorious for data fluctuations depending on temperature and possibly barometric pressure. It was common to dampen these diurnal fluctuations by adding a 1 inch thick layer of pea gravel on top of the pillows. When subsequently hypalon pillows were developed and shown to be less susceptible to the irritating data anomalies, they became the pillow of choice. These pillows were black, 10 feet in diameter and were filled with 150 to 200 gallons of antifreeze. These essentially became a huge heat sink in the summer with lots of energy stored in the soil beneath. What this did was delay the onset of snow accumulation compared to the original stainless steel pillows. Also, hypalon pillows melt out faster than their stainless counterparts giving the illusion that there is a trend or a greater trend toward earlier melting snowpacks. Hypalon pillows accumulate less snow than stainless steel pillows and the effects are greater at lower elevations – all what one would intuitively expect as impacts due to climate change but in reality all/most/some is due to sensor change.

In this graph of collocated steel and hypalon pillows at Parleys Summit SNOTEL (elevation 7500 feet) that the hypalon is running 4 inches of SWE less than its steel counterpart and melts out 2 weeks earlier as well. The impacts of this sensor anomaly are much greater at lower elevations and far less at the higher elevations – likely due to those sites having far less opportunity to become a significant heat sink.  Various areas of the SNOTEL system have vastly differing stainless to hypalon conversions. In the Utah, Nevada and California, we replace stainless with hypalon upon pillow failure so many of these sites still have stainless pillows. In other areas such as Montana, all stainless were replaced with hypalon over a short period (several years) of time which conceivably introduced a step function in the data set.

Parleys Summit SNOTEL – hypalon pillow center, steel pillows right.

http://www.ut.nrcs.usda.gov/snow/siteinfo/data_bias/soiltemps-steelandhypalon.pdf

There is yet another factor in melt out dates that needs normalization if it is to be used as a climate metric. Melt out dates are in fact a good metric for drought because snowpacks with a smaller mass melt out earlier than to their larger counterparts for equal amounts of energy. The analogy here is: take a 10 lb block of ice and one ice cube, put both on hot asphalt and the ice cube melts out faster every time. Thus drought or small snowpacks will melt sooner than larger ones. If the metric is to be temperature related – then the energy budget relative to the size of the snowpack is important. Analyses of melt out dates that end during the drought of 2000-2004 are predisposed to declining results.

In our analyses of melt out dates (without any normalization procedure), once the pillow data have been adjusted for type, we have found no significant difference at any of the SNOTEL sites in Utah.

SNOTEL temperature.

There were limitations in the amount and type of data that the early SNOTEL system could process and transmit.  Thus early on, only the SWE, precipitation and current air temperature data were initially collected.  The observation times of these early data occurred without any uniformity.  This situation was adequate for SWE and precipitation measurements but gave a much less than desired result for temperature.  The data poll for individual sites typically started at midnight, 6:00 am, noon, 6:00 pm and could last for up to five or more hours.  Thus air temperature data would be reported at the time an individual station was contacted and might vary from as early as shortly after midnight to as late as 5:00 am.  Polls were conducted four times daily but in that early period, a station might report between zero and four times per day and the time stamp on those data would be dependent on when the station reported.  Temperature data were to be used in a relational context to calculate or predict snow melt rates, predict the onset of melt and generally be used in a water supply context.  As the data collection, processing and transmitting electronic components were improved, additional sensors were added to the system.

In the mid 1980’s, with the advent of better electronics, daily minimum, maximum and average temperature data were added as standard data collection to SNOTEL.  Unfortunately, temperature sensors were not uniformly installed across the entire network and in fact, a rather poor job was done particularly in the mounting of these sensors.  The first temperature sensors were generally Climatronics or Climet thermistors and were in small ~3X3 inch aluminum box shaped shields or in aspirated housings.  Most of these sensors were mounted on or very close to the brown SNOTEL shelters.  Others were mounted to the antenna tower.  Various mounting configurations were used, mostly dictated by the ease of installation and not to any technical standards.  In some cases, they were mounted horizontally across the face of the shelter about six to 12 inches below the shelter roof and about 24 inches from the side and in all cardinal directions.  In other cases, an “S” shaped aluminum tube was used to mount the sensor vertically to the side of the shelter which put the thermistor about six inches from the shelter side and up to 12 inches above the roof line.  In yet another configuration, the sensor was mounted vertically above the shelter at a distance of between three and six feet.  Clearly any mounting configuration that put the sensor near the brown radiating shelter would have a net warming bias on the early dataset.  Some of these early sensors were mounted to remote antenna towers then moved to a shelter mount to be moved again to the antenna tower and finally moved to the meteorological tower.  Much of the early data from these sites are compromised by inappropriate mounting, sensor moves and various changes in sensors and aspirators.  Another more isolated and easily identifiable problem with this mounting scheme is that occasionally, a shelter and its temperature sensor was completely buried in snow or the roof snow load encased the sensor in which case, the air temperature sensor became a snow temperature sensor.  Later, some sensors were moved to the antenna towers, which in most cases is a better location.  However, some towers were directly adjacent or attached to the shelter and the temperature data at these sites could be compromised to some degree, other towers, remote from the shelter should have reasonably consistent data.

In this Photo, the SNOTEL site at Beaver Divide, Utah is shown with a standard YSI thermistor, a three gill aluminum aspirator and the “S” mount attached to the side of the shelter and extending above the roof line.  This configuration is perhaps the worst of all mounting scenarios.

The next Photo shows the Rock Creek SNOTEL site and the impact on snow the brown shelter can have via long wave radiation. Notice the snowpack has melted to a distance of about 2 feet from the edge of the shelter and that the temperature sensor mounted horizontally across the top of the shelter in a northeastern direction is in a direct line above that obvious impact.

In the next Photo, (Buck Flat, Utah) there is a standard YSI thermistor mounted on a remote tower.  This sensor has subsequently been moved from this location to the meteorological tower some 20 feet distant, but the overall impact of this move would be small.  Data from these sites would be the most consistent in relation to current standard location and mounting practices.

Temperature sensors gradually migrated to YSI thermistors with a range of aspirators including the most commonly used, a silver three vent aluminum model.  There were tower mounted sensors and aspirators that were white, wind directed models and a variety of other configurations.  In the mid 1990’s, snow depth was added to many sites as a standard sensor and this began the installation of standardized meteorological data collection towers.  At that time, there was still a mix of temperature sensors mounted on shelters and on antenna towers with a wide variety of aspirators.  Since snow accumulation is variable across the West, tower height is also variable with most meteorological towers being in the 10, 20 and 30 foot ranges, depending on snow depth.  Sensor mountings are therefore also variable in height being at about seven, 17 and 27 feet respectively.  The majority of all sensors are at the 17 foot height.  That stated, during a significant portion of the year, the ground surface level is constantly changing due to the accumulation and ablation of the snowpack and the respective height of any individual sensor may range from 17 feet to as low as five feet or less.

In this photo, (Cascade Mountain, Utah) the current standard temperature mounting configuration is shown with the sensor at 17 feet, mounted three to six feet from the tower and in a white, six gill aspirator.

In the mid 1990’s as the installation of meteorological towers progressed, another sensor change was made from the standard YSI sensor to the extended range YSI sensor in order to capture temperature readings to minus 40 degrees F.  The coldest sites were the first to get the extended range sensor.  Personnel in theIdahoregion had the foresight to run both the standard YSI and the extended range YSI side by side for several years with identical mounting and aspirator configurations and noticed a plus one degree C difference in a very large portion of the observed temperature range of the extended sensor compared to the standard sensor.

This graph displays a comparison of side by side mounted standard and extended range YSI temperature sensors, the results which show a one to two degree C difference between the two sensors, with the current extended range sensor warmer than the standard sensor.  This difference is most noted at the lower end of the temperature scale whereas at the upper end, the difference becomes much less.  This is consistent at all of the Idaho sites tested.

Physical site changes will continue to pose some problems in overall data consistency. Vegetation grows, and at times, dies yielding an ever changing solar view and site characteristics.  At some sites, this could be dramatic and at others, not likely to be much of a source of an inhomogeneous dataset.  The removal of one or several trees at a site can impact the canopy, solar window and evapotranspiration characteristics which could change the temperature regime over some or all of a daytime pattern including nighttime pattern.  The same could be true of growing vegetation altering the periods of full sun or shade at any given site.

In this photo, (Camp Jackson, Utah) one can clearly see the proximity of the vegetation to the tower and the temperature sensor as opposed to the vegetation in the previous photo.  Vegetation height and proximity is constantly changing at some sites, while at others it tends to be relatively stable.  At this site, the dominant species isAspen(Populous Tremuloides) and has the added feature of being deciduous which changes the overall solar input as leaves are generated in the spring and subsequently lost in the fall.  At other sites such as Big Flat, Utah, which is in a mature Spruce and Fir (Picea and Abies) forest, the vegetation is and has been very stable.  However, should the current beetle and bud worm infestation spread with subsequent high spruce mortality, currently experienced in southern Utah, the vegetation at this site could change dramatically and hence, the solar window.

Another source of potential inconsistency in the dataset is that of data editing and quality control.  For the most part and with the exception ofIdaho, the SNOTEL temperature dataset has undergone little in the way of systematic data quality control and verification.  The data editing that is done is primarily the removal of howlers and screamers and focused on the daily maximum, minimum and average.  Some areas such as Idaho have done more and have concentrated on a serially complete dataset complete with estimated data points but the techniques of data estimation and editing have been far from standardized system wide at this point.  The NRCS is attempting to use spatial climate station comparison methodology to resolve suspected data and fill-in missing data by assigning quality control flags that are quantified by confidence probabilities.

An un-quantified source of data error is in the electronics of the system and could be either random, systematic or a combination of both. There have been a series of joint transceiver/receiver/data processors in combination with the series of different temperature sensors. These include the Secode, MCC 550A, MCC 550B and the current version, the MCC 545 coupled to a Cambell CR10X data logger. Each of these systems measure voltages from each sensor which are then equationally converted to meaningful engineering units.  Errors may occur due to: error in the thermistor, resistive errors such as line loss, ground potentials, exitation voltages, and errors associated with the data logger reading the voltages. These errors may occur from something as mundane as the type of cable used or in the connections from the cable to an interface. Drift in the voltage reading device could be the source of some un-quantified error.

In summary, the historic temperature data from the SNOTEL network have some significant systematic and random bias.  This bias includes poor mounting techniques, sensor changes, location changes, aspirator changes, vegetation changes and electronic errors.  Vegetation changes can be in the form of 20 to 30 years of growth or instantaneous change due to fires, disease, or insects.  Documentation of these changes has been inconsistent system wide and currently resides mostly on paper records difficult to access and digitize.  Much of the very early record, from mid 1980’s to the early 1990’s will be difficult to salvage and much of those data are compromised by poor sensor mounting techniques and are suspected to record much warmer temperatures than actually occurred.  General use of these early records as comparisons to current conditions is discouraged.  Some records, particularly those sensors that were mounted to remote towers early on will have reasonable quality, subject only to observation time, aspirator and sensor changes and possibly some vegetation change.  Useful metadata on dates of sensor, aspirator and location changes would facilitate the potential construction of a reasonable, corrected dataset for these specific sensors.

Currently, NRCS is moving to standardized temperature sensor mountings that are on a meteorological tower in a specified data collection area at a height of 7, 17 or 27 feet and a distance of four to six feet horizontally from the tower.

The following illustrates a serious potential systematic error in the SNOTEL temperature data set.

This graph shows the minimum daily temperatures from the Thunder Basin SNOTEL site inWashingtonfor the years 1987 through 2003.  The issue is a flat line of temperature data at 32 degrees starting in late November and continuing into April.  The “minimum minimums” look good with a large range of data from -27 degrees up to near freezing conditions, however, the “maximum minimums” hit a ceiling at 32 degrees. 

This chart shows the maximum data for the same site and time period. Notice that there is no ceiling in these data, both the “maximum maximums” and the “minimum maximums” show a very jagged edge on top and bottom with a large amplitude and defined sinusoidal pattern through the year.  Obviously the sensor itself is sensing exactly what it sees – in other words, the sensor is operating correctly and the ceiling in the minimum temperature trace must be an artifact of the environment of the sensor.

The sensor environment.

This sensor is mounted at about 10 feet, on the shelter to the lee side of prevailing winds and just below the top of the shelter itself.  When there is significant snow accumulation, this sensor is about 12 inches from an open ice box.  Snow on top of this shelter could be several feet to potentially 4+ feet deep which means that the air temperature is actually measuring temperature near the snow surface.  In the daytime, solar radiation from the shelter would allow increased maximum temperatures but at night, absent solar heating and given some protection from warm wind impacts, this sensor would not register values much above the snow surface values.  The minimum minimums are free to fall, but the maximum minimums are constrained to near 32 degrees.

This sensor was moved to a location on the antenna tower at a height of 19 feet. This means that it is about 7 feet higher than the shelter roof at this point and given a snow depth on top of the roof of say, 5 feet, may still be compromised to some degree.

An analysis of the data post 2003 shows that the current mounting configuration is much better than the old.

This chart clearly shows that the 32 degree ceiling of the previous charts no longer exists. Given the current sensor location this site may still have some ceiling, the question becomes what that ceiling may be and what time duration it may exist.

This sensor mounting technique was very common throughout the SNOTEL system for many years.  This specific problem of a minimum temperature “ceiling” has just been identified.  There is a significant potential for this systematic bias to be misinterpreted.  In the observed data, there can be this minimum temperature ceiling and when removed, as it has been by moving these sensors to a meterological tower remote from the shelter, there will be a net increase in minimum temperatures system-wide.

Conclusions

These data sets, used appropriately, can be of great value in determining various aspects of climate change. These data contain numerous sources of systematic bias that must be identified and removed prior to making conclusions about the magnitude and rate of change in any given variable. The shotgun approach of taking all sites and making general conclusions about trends is simply not an acceptable method as the biases described in this document for Utah are most certainly in all other areas as well.

Comments Off on Uncertainty in Utah Hydrologic Data: Part 1 On The Snow Data Set by Randall P. Julander

Filed under Climate Change Metrics, Guest Weblogs

Comments are closed.