SPoRT Embarks on a Project to Produce Stream Height Forecasts using Machine Learning

The use of various soil moisture parameters from the SPoRT-LIS for monitoring drought conditions and assessing flood risk has been ongoing for years, and has been demonstrated for efficacy at other collaborative offices.  The use of soil moisture output for drought analysis is relatively straight-forward.  However, the use of the data for assessing flood risk has always been a bit more complicated and has involved the development of significant thresholds of rainfall and soil moisture that lead to flooding based on forecaster experience.  This method lends itself to some degree of subjectivity and has always been less quantitatively robust than preferred.  Nevertheless, the SPoRT-LIS data have provided valuable information regarding the state of soil moisture and the potential for flooding in near real-time.  There are older posts on the blog that describe the application of the data for flood risk assessment.

Now, the SPoRT group is embarking on a new project, employing a machine learning technique to provide a more quantitative measure of the relationship between soil moisture values, rainfall and stream height.  This new methodology involves the use of a Long Short-Term Memory (LSTM) Network.  The LSTM model for a particular drainage basin can be trained using a history of SPoRT-LIS soil moisture values, gauge height observations, and precipitation from the Multi-Radar Multi-Sensor Quantitative Precipitation Estimation data set.  Using this training data set, the forecast model for each basin is then run in real-time using the most recent gauge height observations, SPoRT-LIS soil moisture at various depths, and quantitative precipitation forecasts (QPF).  For this initial version of the gauge height forecasts, we’re using QPF values from the GFS model and the WPC.  One of the advantages of this type of modeling is that the great majority of computational power is on the front-end to train the model, while running the model in real-time is computationally much less expensive than running a hydrologic model.  Thus, we can run multiple precipitation scenarios for any basin quickly in real time.  For example, for the 40+ basins in our initial evaluation, the amount of processing time needed to run each basin at 6-hourly time steps out to 5 days with two different precipitation schemes takes just about 5 minutes!

The SPoRT group is working with some of our collaborative NWS offices (Huntsville, Nashville, Morristown, Sterling) and the Lower Mississippi and Mid-Atlantic River Forecast Centers (LMRFC and MARFC, respectively) for this initial test and evaluation of the stream height forecasts.  Although SPoRT has produced models for several thousand basins in the Southeast CONUS domain, this initial evaluation will involve a sub-set of streams, shown in the image below.

Image 1.  Gauge locations (black dots) for the initial evaluation of real-time gauge height forecasts from NASA SPoRT.

So, one might be asking…why is SPoRT engaged with stream height forecasting?  It’s important to remember that the SPoRT paradigm involves working closely with collaborative partners and assessing forecaster needs.  One of those needs involves having a better sense of flood risk at mid to long timescales during the 7-day forecast period.  Let us explain.  Operational gauge height forecasts from the RFC may not incorporate precipitation into the hydrologic models beyond one or two days due to operational and model limitations and constraints.  However, this can be problematic if an area is expecting heavy rainfall in the period beyond a day or two.  Take for example the current flooding event occurring across parts of the Tennessee Valley.  Operational gauge height forecasts for the Flint River (at Brownsboro, AL) from the afternoon of February 3rd indicated no rise forecast for the river (Image 2).  This is because the hydrologic models were not incorporating precipitation into the models as it was before the 48 hour cutoff.

Image 2.  Graph of gauge height observations (dotted blue line) and forecasts (dotted red line) for the Flint River at Brownsboro.  The forecast was valid approximately 1440 UTC 3 Feb 2020.  Observations are current through about 22 UTC 5 Feb 2020.  Horizontal bars at the top of the image indicate flooding thresholds (yellow=Action Stage, orange=Minor Flood, red=Moderate Flood)

However, heavy rain was expected in the region, which would certainly lead to some river rises.  The question is…how much?  Our old rules of thumb would have suggested flooding likely, based on soil moisture values and expected rainfall.  But again, the old rules didn’t indicate the time frame for flooding or the degree of flooding…it was generally just a qualitative “likely” or “not likely”.  So, to help alleviate this gap in knowledge, the new methodology provides objective, deterministic forecasts of stream or gauge height.  The image below shows the gauge height forecasts from the SPoRT LSTM models valid at about the same time.  Notice that the forecasts based on both GFS and WPC QPF scenarios indicated flooding was likely, while the higher precipitation from the GFS suggested flooding would reach Moderate Stage.  So, it’s easy to see here one of the advantages this type of modeling can have for overall hydrologic forecasting and situational awareness for the threat of flooding.

Figure 3.  SPoRT LSTM gauge height forecasts for the Brownsboro River at Brownsboro.  The black line shows observations, up to analysis time at 12Z 3 Feb 2020.  The blue dashed line contains gauge height forecasts based on GFS QPF, while the red line contains forecasts based on WPC QPF.  The blue vertical bars indicate 6-hourly GFS QPF, while the red bars indicate 6-hourly WPC QPF.

This post has become rather long.  So, we’re going to leave it here for now.  We’ll be providing more information about this project and discussing other advantages and limitations of our stream height forecasts in some upcoming posts as we continue this evaluation over the next couple of months.

– Kris and Andrew

Comparison of Quality Flags for NUCAPS Gridded Products

Gridded NUCAPS products developed as part of a multi-organization JPSS PG/RR project are currently being evaluated at the Hazardous Weather Testbed (HWT) Experimental Warning Program (EWP).  The project contains contributions by researchers from UW/CIMSS, CSU/CIRA/ UAF/GINA, and SPoRT.  NUCAPS soundings are retrieved temperature and moisture soundings from the Suomi-NPP CrIS and ATMS sounders.  The evaluation of NUCAPS at HWT is aimed at providing upper air temperature and moisture information in the pre-convection environment to better understand variables that are necessary for convection and severe weather.  The Gridded NUCAPS products allows for isobaric plan views of temperature and moisture that forecasters can use to gain confidence in the model output

Forecasters at the HWT-EWP posted some input on the use of the Gridded NUCAPS products.  On the Satellite Proving Ground at HWT Blog (http://www.goesrhwt.blogspot.com/2016/04/nucaps-planviews.html), a forecaster noted:

“[Gridded NUCAPS] would be beneficial in the forecasting environment as added temperature data would be available in between standard upper-air launches.  This could serve as a good proxy to help judge the strength of a capping inversion, while also possibly serving as an additional information source during winter wx events.

However, the forecaster also noted that the amount of missing data included in the product limits its utility.  Currently, the Gridded NUCAPS contains only the highest quality (i.e., “best”) data that comes from a combination of both microwave and infrared (top image below).  In this image, the dark blue pixels represent the data that are discarded due to QC issues.  However, this quality control can be strict at times and leave out “good” data that can still be useful to the forecasters.  When these “good” data are included, there are much more useful data (bottom image below) without any noticeable discontinuities or oddities in the data.

SPoRT plans to use the feedback from HWT-EWP participants to test pushing the inclusion of the “good” quality data to the Gridded NUCAPS product to provide forecasters with more data for their analysis.

NUCAPS.2016.04.21.1902329.853mbtemp_QCeq1

853 hPa Gridded NUCAPS temperature product from 21 April 2016 at 1902 UTC including only the highest quality flags.  Dark blue pixels denote discarded data that results in data gaps.  Note that a lot of over-land observations are discarded.

NUCAPS.2016.04.21.1902329.853mbtemp_QCle2

853 hPa Gridded NUCAPS temperature product from 21 April 2016 at 1902 UTC including both “best” and “good” quality flags.  Dark blue pixels denote discarded data.  Missing pixels generally correspond to thick cloud features.

 

Snowfall Rate Provides Guidance for New Mexico Snow Event

Forecaster Jennifer Palucki from Albuquerque, New Mexico submitted a nice case study to our online evaluation form being used during the current 2016 NESDIS Snowfall Rate Evaluation.  Here are some of her discussion and impressions of using the product:

A very well defined band of snow developed along a frontal boundary extending from the southern Sangre de Cristo Mountains, toward Las Vegas, and continued southeastward toward Melrose. Initially the southeast part of the band was rain, but as temps dropped it changed to snow. At 0052z (552pm MST; see image below) the merged SFR likely did very well distinguishing where there was snow and no snow, however, in areas that there was snow, amounts were way underdone. At 545pm, approximately 4″ of snow had fallen in Sapello in the southern Sangre de Cristo Mtns. Snow likely started around 1 or 2pm, which is an average of about 1″/hr compared to the 0.3″/hr the SFR product was showing with an 18:1 ratio. Thus, the amounts via the SFR product were largely underdone. It was still snowing heavily according to the spotter at 545pm. At 645pm, approximately 1.5 inches of snow was reported in Las Vegas. The SFR product was showing around 0.1″/hr for this area.

ABQ_160203_0052Z_annotated_zoom

NESDIS SFR Product at 0052 UTC on 03 February 2016 showing light snow over Las Vegas, NM.

Another pass at 0330z (830pm MST; see image below), the SFR product missed the southeastern extent of the snowfall, and again had amounts that were likely underdone. A report of 0.5 inches of snow in the last hour was reported at 841pm in Taos. The SFR product showed around 0.02 liquid equivalent, or around 0.3″/hr snowfall rate given 18:1 ratio (which should be close to the snow ratios in that area).

ABQ_160203_0330Z_annotated

NESDIS SFR Product at 0330 UTC on 03 February 2016 showing some heavier snow over Taos, NM.

Really like using this product to gather intel on where it is snowing in areas without radar coverage. Do have some concerns about the amounts, especially in these scenarios where the heavier amounts are likely isolated. In this case, the band was very narrow, likely no more than 10 to 15 miles wide.

Life of Winter Storm Jonas as seen by the NESDIS Snowfall Rate Product

Winter Storm Jonas tracked across the eastern United States this past weekend dropping near-record amounts of snowfall in a track from West Virginia through southern New York.  Two things about this storm are particularly interesting:  1) the heavy amounts of snow that fell for long periods of time and 2) the relatively narrow swath of the heaviest snows.  Below is the 48-hour snow accumulations from the National Weather Service ending Sunday, January 24.  It is striking that New York City received on the order of 30 inches of snow, while areas less than 100 miles to the north received little if any snow.

NWS_SnowTotals

48-hour snowfall totals ending Sunday, January 24, 2016 (from NWS Central Region).  Contours are every 3″ with the darkest reds indicating 30″ of snow.

Select Eastern Region WFOs are currently evaluating the NESDIS Snowfall Rate product, which uses passive microwave observations from 5 sensors, to observe total column snowfall rates.  Below is a series of images from this past weekend showing the SFR product displayed as a 10:1 solid/liquid conversion.  The darkest greens indicate snowfall rates at the top of the sensor detection range at approximately 2″/hr.  Depending on the actual solid/liquid ratio in individual areas, rates may have been higher.

SFR_Collage_first4

SFR_Collage_second4

NESDIS SFR Product showing the evolution of Winter Storm Jonas from late on Friday through early Sunday.  The darkest greens indicate solid snowfall rates of around 2″/hr.

In the images, the NESDIS SFR product shows very good agreement with the location and track of the heaviest snows (greens) compared to the heaviest totals in the ground reports.  Additionally, the SFR product does well in picking up the abrupt northern edge of the snowfall (especially across southern New York).

UPDATE:  The Sterling, VA WFO included mention of the SFR product in a forecast discussion to confirm snowfall rates that would cause white out conditions:

Sterling_AFD

NESDIS Snowfall Product Captures Unfolding Winter Weather in the South

Beginning in the morning hours of 22 January 2016, rain began to change to snow across Mississippi, Tennessee, and Alabama.  The NESDIS Snowfall Rate, which is currently being evaluated by a handful of Weather Forecast Offices, has the ability to differentiate rain from snow.  This ability was particularly important for the large winter storm impacting much of the eastern half of the United States.  The animation below shows the 10:1 Solid SFR Product with METAR station observations indicating temperatures and precipitation.

AL_TN_SFR_Example_20160122_07-19Z_slower

The animation shows the evolution of snow across the area beginning with snow in Western Tennessee and Eastern Mississippi at around 1200 UTC (6:00a local time).  Also of note at that same time is that the SFR Product indicates relatively heavy snow (~1.5 in./hr. solid snow) directly over the Nashville area; however, the METAR site at the airport is still reporting rain.  In the following hour (1300 UTC; not shown in the loop here because there was no SFR product valid near 1300 UTC) Nashville was reporting snow.  Thus, the SFR product was seeing in-cloud snow in that area that began to reach the ground within an hour of the observation.  This is one way forecasters can use the product to view in-cloud snow to determine the potential for snow to reach the ground.

Later in the period, a similar set up appears in the Huntsville area at the Madison County Executive Airport (KMDQ).  The 1853 UTC SFR product shows light snow over Madison County, but the 1900 UTC METAR was not yet reporting any snow.  However, the 2000 UTC METAR showed snow beginning to fall across the Huntsville area.  The change over to snow falling across Western Madison county into Central Madison county was between 1830 and 1900 UTC, verified as I drove home from work.

The NESDIS SFR product will continue to be evaluated as blizzard conditions begin to set up along parts of the East Coast.

Latest Version of NESDIS Snowfall Rate Product in AWIPS

Researchers at the National Environmental Satellite and Information Service (NESDIS) have recently wrapped up development of the latest iteration of their Snowfall Rate (SFR) product to aid WFOs in situational awareness of snowfall events and snowfall forecasting.  The developers within NESDIS have teamed up with SPoRT, utilizing SPoRT’s unique transition, training, and evaluation capabilities to deliver the SFR products to several WFOs in the CONUS and Alaska.  During the evaluation period this winter, I will be evaluating the SFR and merged SFR products for use mainly here in the Tennessee Valley (provided the atmosphere obliges), but I will also be looking at the product CONUS-wide (and perhaps AK too, as the opportunity affords).

The SFR products are being delivered in two main versions: a merged snowfall rate product (merged polar-orbiter and radar data) and a product that contains only data from polar orbiters.  Through collaboration with researchers and forecasters (especially at the Boulder NWS office), SPoRT is including SFR data with liquid to snow ratios of 10:1, 18:1, and 35:1.  These data are being ported in AWIPS II workstations at the NWS offices.  In the merged product, the polar swath data are complimented with NSSL’s Multi-Radar/Multi-Sensor (MRMS) precipitation data, and update much more frequently (every 10 minutes).  Swaths containing polar orbiter data of course come in as associated polar orbiter swaths cross a region, with updates from about every 30 minutes to as long as ~4-5 hours over any location.

A look at the products the past several days has brought the opportunity for some initial evaluation.  So far, the SFR product looks rather promising.  Here’s a quick look at the product as a snowstorm was ongoing yesterday evening (Dec 15th) across the northern/central Rockies and the Northern Plains.  The loop below (Image 1) shows data from 0110Z through 0410Z 16 Dec 2015.

NESDIS Merged Snowfall Rate Product (showing 10:1 liquid to snow ratio) 0110Z to 0410Z 16 Dec 2015

Image 1.  NESDIS Merged Snowfall Rate Product (showing 10:1 liquid to snow ratio) 0110Z to 0410Z 16 Dec 2015

The loop above shows the Merged Snowfall Rate product (displaying 10:1 liquid to snow ratio).  Most of what you see is the MRMS precipitation during the loop.  At the end of the loop however, you will notice a sudden expansion of the apparent snowfall over the region as an insertion of snowfall rate derived from a polar orbiter swath is incorporated into the product.  So, let’s take a closer look at that single image containing the polar orbiter data (Image 2).

 

NESDIS Merged Snowfall Rate product (10:1 ratio) with polar orbiter data insertion, 0410Z 16 Dec 2015

Image 2.  NESDIS Merged Snowfall Rate product (10:1 ratio) with polar orbiter data insertion, 0410Z 16 Dec 2015

In the image above, you will notice that MRMS data remain and replace satellite retrievals where these data are available.  That is, the MRMS data take precedence over the satellite data in the merged SFR product.  However, data are inserted for locations where snowfall is detected by satellite instruments and radar (MRMS) data are not available.  For large areas of Wyoming and Colorado, where radar coverage is certainly more limited, notice that the insertion of polar-orbiter data allowed for a more thorough and proper analysis of locations likely experiencing snowfall.  Many of the surface observations (in yellow) likewise corroborate the snow that was occurring, particularly for locations in Wyoming, where coverage from radar data alone was very lacking.  However, there are some surface observations that do not corroborate where the SFR product is indicating snowfall.  Multiple reasons for this apparent discrepancy may exist, but it’s important to remember that the polar orbiting satellite instruments are detecting snowfall in the clouds.  Some of this snowfall may not be reaching the surface due to sublimation aloft.  Also, the snowfall could be very light and patchy in some instances with detection issues at some of the automated ground observation sites.

Now, let’s take a quick look at the polar orbiting data alone (Image 3).

NESDIS SFR product (liquid to snow ratio 10:1) 0345 UTC 16 Dec 2015

Image 3.  NESDIS SFR product (liquid to snow ratio 10:1) 0345 UTC 16 Dec 2015

The resolution of the polar orbiting data still allowed for the detection of banded structures across parts of the Dakotas that were evident in the MRMS data.

Further evaluations and posts about this product will be forthcoming as we progress through the winter.  Perhaps I’ll have the chance at some point to evaluate the product here in the Tennessee Valley…that is, if the current mild Eastern U.S. pattern changes.

 

Follow-up to Sep 2 and Sep 3 Dust and Flash Flood concerns

As mentioned in the previous post, we were concerned about the potential of organized thunderstorm outflow creating favorable conditions for blowing dust, as well as previous heavy rain activity increasing the potential for flash flooding in portions of Southeast Arizona yesterday (Sep 2) and today.  We have been trying to use LIS data as part of our process in determining the threat of both of these problems.  Yesterday was a mixed bag as I briefly outline below with less than full cooperation from the atmosphere.

Convective initiation occurred in locations we expected (similar to the WRF output posted previously), however we were unable to get the chain reaction of outflows that we hoped for.  We had problems in many valley locations (especially the Tucson Metro area) with early debris cloud and then anvils from early convection blocking solar insolation. We also underestimated the eastern extent of a modest low level drying trend filtering in from western Pima county.

Below is a loop of velocities with some observation and warning overlays early yesterday afternoon. The initial 45 kt outflow from the SVR southeast of Tucson attenuated rapidly as it approached Tucson.  Readings from Davis-Monthan AFB on the southeast side of Tucson at 2137Z showed a gust out of the SE at 30kts, and shortly later Tucson International Airport (a couple of miles further west) registered 23 kts.  By the time the outflow was west of Tucson it was difficult to detect.  There was no additional activity along this outflow in Pima County and it certainly wasn’t strong enough on it’s own to generate any dust problems by the time it got to areas we were concerned about.  The initial area that it started southeast of Tucson does have dust issues at times, but referring back to the soil moisture imagery from the previous post, things were pretty wet there.

sep2velsvr

It was an active day however.  Below is a composite post of the Severe Thunderstorm (yellow) and Flash Flood (green) products we issued yesterday.  When you compare to the 09z LIS output posted yesterday, the Flash Flood warnings were issued in an area with relative soil moistures above 70 percent (posted again for convenience).

severe and flash sep 2

soilmstrdusttue-20150902_090000

We once again have a favorable atmospheric profile for strong storms today (Sep 3), but with a little more convective inhibition to overcome and continuing issues with cloud cover.  We do have a stronger impulse embedded in the southwesterly flow that will push into our area late today and this evening.  Most standard and mesoscale model output (including latest UofA WRF and national HRRR) show increased coverage and organization of thunderstorm activity by late afternoon, especially west of Tucson.  This seems very reasonable, keeping the aforementioned caveats in mind.

A look at today’s LIS output shows that soil conditions are even more favorable in the areas of concern west and northwest of Tucson.  Especially in eastern Pinal and Maricopa counties with widespread values below 15 percent:

sep3mrnsoil-20150903_090000

Some of the special communications to our partners and the general public yesterday morning extended into mentions of issues for today as well.  We will adjust our weather story and social media posts to reflect the latest information, but our message is similar to yesterday.  We will again coordinate with Phoenix about any possible coordinated dust headline later this morning.  An example of the partner email we sent yesterday below:

BriefingUAPD09022015