Tuesday, May 26, 2015

NDFD – The National Digital Forecast Database

The original of this article was posted in May, 2015 and updated, April, 2017.

Part 1. Background

The primary method of weather routing underway these days is based on vector wind forecasts in GRIB format displayed in echart navigation programs. Navigators who do not display the data directly in an echart program still use this data viewed in separate GRIB viewer software. One popular free program and data source is the WinFax program from Saildocs that includes both the GRIB viewer and a convenient link to download the data via email using HF radio or satellite phone when underway, or by Internet when available. Another popular free program is OpenCPN. Commercial GRIB viewers and data sources include Ocens, Expedition, LuckGrib,  and others.

The primary source of free data used in these programs is the numerical prediction forecast of the Global Forecast System (GFS) model from the National Weather Service (NWS). This is free, worldwide data. There are other (regional) model output options for selected regions that can improve on the GFS forecasts in coastal waters, but once offshore we must rely on a global model. There are global model data from other nations that are as good or sometimes better than the GFS, but these are not free data. There is also free global data from the US Navy (NAVGEM, formerly NOGAPS), but this would rarely be an improvement over the GFS. The regional model from NWS is called NAM; the one from the Navy is called COAMPS. The HRRR model is highest resolution with the most frequent updates.

The GFS model offers new wind forecasts every 6h, in 3-h steps, out to 10d, but none of the models are dependable on the level a sailing navigator might want beyond about 4 days (96h). The latest GFS model computes wind at grid points of 0.12º, but this high resolution is so far not typically available to sailors at sea. The most common sources of GFS data provides a wind vector every 0.5º, or about every 30 nmi, depending on Lat. The racing tactics program Expedition and others offers GFS data down to a 0.25º grid.

But here is the ongoing issue for those who want to do the best they can in wind forecasting. The GFS data are pure computer model output. The data you get this way has not been vetted in any way by human meteorologists once it left the computer. That is not to say it is wrong.  The computational science of the atmosphere is improving every year and is remarkably good artificial intelligence. In fact, the GFS winds will be very close more often than very wrong, but they can indeed be wrong, and can indeed be close, but not right, especially as we get beyond 48h.

So what do we do about this to optimize the forecasts? First we must remember that we do indeed have the actual forecasts from the human intelligence of professional meteorologists. We get these forecasts from the folks at the Ocean Prediction Center (OPC) and the National Hurricane Center (NHC). We also have regional forecast offices in Honolulu (HFO) and three in Alaska (ARH) that contribute to Pacific marine forecasts.

They provide (via Internet links or HF retransmission via USCG) surface analysis maps every 6h, as well as a 24h forecast and 48h forecast every 12h, and a 96h forecast every 24h. Thus we can compare their forecasts with what we see in the GFS forecasts as a way to judge if the GFS has withstood the scrutiny of the professionals. These meteorologists have, in preparing their own forecasts, studied all of the global models available, several of which are ranked higher than the GFS in overall performance skill. To make their judgments they include various ensemble studies of the models, which include looking at the output at a specific time based on different initialization times, as well as different input data. They can also vary the physics parameters of the solutions. The extent to which the results are independent of these variations in the calculations directs them to the best model to use for the situation at hand.

They also fold into this their experience with the climatic behavior in various regions as well as past performance of the models in particular circumstances. They can also use their experience to evaluate the ship and station observations that have seeded the models for the latest run, and they may have access to satellite wind and cloud image data that did not happen to show up in time to be assimilated into the last computation. In short, the background that goes into the forecasts they make is far more than just the output of one particular computer model, so we have every reason to believe that they could provide a better forecast.

With that said, we must add that the difference between what the meteorologists provide and what the GFS model alone provides depends on what part of the world you are in and when. Our recent study of this, for example, for the Transpac route, LA to Honolulu, during the summer, with a nicely formed and in place Pacific High, showed very little difference between human forecasts and pure GFS model predictions, and in such cases the NWS will indeed simply use the pure GFS themselves to create the isobars and subsequent wind fields.

But when things are not so benignly climatic, or at higher latitudes much of the time, or near coastal waters, the differences can be significant and it is our job as prudent navigators to make the comparison before relying on the GFS alone.

Up until late last year the process of comparing the two forecasts when underway was rather tedious, because the NWS analyses and forecasts were only available as graphic images of the weather maps. Thus we had to plot various positions on these and then extract the wind and pressure data using special tables. These data could then be compared to the GFS values for the same forecast times. The process and required tables are in the textbook Modern Marine Weather. But now we have a new digital solution that for many parts of the world not only offers an easy solution but in effect diminishes the requirement for the comparison in some cases... and that is the topic at hand—a long background introduction to put this new data into perspective.

Part 2. Oceanic NDFD

The answer is called NDFD, the National Digital Forecast Database, which is the vectorized versions of the NWS forecasts, which are now in a GRIB format that all mariners can download and run in their standard GRIB viewing software—parts of this system are actually some years old already; it is just our access underway that is new.  The raw data are readily available online, but in a complex format that requires special programs to interpret, and then it must be converted to the .grb format that mariners are used to. There may be other sources, but the only one I know of is an email request to Saildocs, and we are very grateful to them for providing this service. This option is not included within the WinFax Get Data option,  but it can be obtained by email request to query@saildocs.com with this in the body of the message: 

SEND NDFD:50N,45N,130W,120W|0.12,0.12|0,6..120

or vary this with the standard SailDocs conventions for region, and forecast span. The forecasts are available every 3h out to 72h, and then every 6h out to 168h (7d); however accuracy beyond 96h cannot be counted on regardless of who made it or how.  The longer run forecasts are still useful on some level for weather routing programs that must look ahead at something to make proposals for earlier times.

These data are called the "oceanic NDFD."  The resolution of the data is very high at 0.12º, which corresponds to roughly one wind arrow every 7 nmi, this is about four times finer grid than typical GFS data, and twice as good as best GFS available. Furthermore, it is not unvetted pure model output as the GFS is, but rather the digital forecast of professional human forecasters. A drawback for barometer oriented navigators like myself, however, is the absence of isobars. For now from the NDFD we just get wind speed and direction and significant wave height (SWH) for use at sea. The database itself includes air temp, humidity, and various other land oriented data.

What we do not have yet and very  much would like to see is that part of the NDFD that covers tropical cyclone surface wind forecasts of winds >34 kt, >50 kt, and >64 kt.  Unlike sea level pressure, these data are all in the NDFD already, but not yet available to mariners in conventional GRIB format. This is scheduled to come online operationally in June, 2017.

The other limiting factor to use of NDFD is the data are not available worldwide.  Figure 1 shows the regions covered by the oceanic NDFD. The borders do not have to be spelled out in the Saildocs request. That service will simply provide what is there within whatever you ask for.  I have been told by the NWS that it is in the planned expansion to extend the coverage from 140W on over to HI, but with such tight budgets these days we don’t know when that will happen. The first step is we need more vessels using it to learn its value.

Figure 1. Coverage of the "oceanic  NDFD"  wind and SWH data.  Wind arrows in this global view do not reflect what we obtain by download, which is one arrow every 0.12º.

This data set is a major breakthrough for sailors. It now covers a large extent of our sailing waters, and if the planned expansions go through it could be the main source of data for cruising and racing sailors in trans Pacific and trans Atlantic sailing. The key point here is that in principle the oceanic NDFD forecasts includes the skill of the ECMWF and UK Met models, which are ranked the top two numerical prediction programs. As we read in the official Forecast Discussions that accompany every forecast made by the NWS, oftentimes the NWS defers to one of these models for their forecasts. Whenever this happens we get that benefit in the NDFD. Normally only the boats with expensive contracts with private agencies have access to that data. (For a quick coarse comparison of the models see www.tinyurl.com/wxmod from weatheronline in the UK.)

Part 3. Regional NDFD

Beside the oceanic data discussed above, there is higher-res NDFD data for the continental US (CONUS) in two resolutions, and for specific US regions, namely AK, HI, PR &VI, and Guam. The areas are shown roughly in Figure 2.

Figure 2. The CONUS and regional NDFD coverage areas overlaid onto the oceanic coverage of Figure 1. 

Where there is overlap of the coverage, there will be two sources for NDFD data, and typically the CONUS or regional would be the preferred. The update times, forecast steps, and resolutions vary within these sets, so I defer this to a later article on the details of this data. See NDFD Oceanic, CONUS, and Regional. Also at the moment there are limited sources of all of this data, so we will keep this up to date in that article.

Part 4. Data Comparisons

These comparisons were done before we had GRIB access to the CONUS NDFD, which have higher resolution and more frequent updates. We will come back and make another comparison later on. This section was included when the data were new to mariners to show that it is indeed an improvement. This is even more pronounced with newer comparisons.

A rough comparison of the Ocean NDFD and GFS is shown in Figure 4, below; but this underestimates the actual differences in cases where GFS was not the basis of the NWS forecast.

Figure 4  Mean Average Error (MAE) for winds above 8 kts from land stations.  The GMOS values reflect the GFS data, but these Model Output Statistics (MOS) have been calibrated to climatic averages. The actual difference between raw GFS wind predictions and the corresponding NDFD predictions are notably higher.  This shows that even forcing a normalization to the GFS model output the NDFD still out performs the GFS above 72h. The plots give insight into the accuracy of the forecasts however they are made, which is remarkably good considering that the wind measurements themselves must be some ± (2 kts, 5º) at least.  It is not clear how ocean data would compare to these land data.

Also on the horizon is a new program called National Blend of Global Models (NBM). This will be an improved forecast system using both NWS and non-NWS models, along with state of the art ways to evaluate input observations plus enhanced ongoing verification to produce a top of the line forecast product. This work is for now focused on land based forecasts, but it will be extended to the ocean as well, and we will then have digital access to through the NDFD. The US did after all invent the concept of numerical weather prediction, so it looks like we may be working toward regaining that leading role.

Looking at this one and the Tehujuanepec one at the end show that the real challenge here is having the model forecasts in hand when we get a good scatterometer wind field to test


Below are a few comparisons of the model predictions for several regions.  [Note added, this was from 2015, and we have better data now and better ways to make these comparisons. See latest edition of Modern Marine Weather.]

The above is GFS (purple) compared to oceanic NDFD (black) for a TC centered at abut 20N, 125W, with an overlaid ASCAT satellite measurement of the actual winds. The forecasts were made 18h earlier and the agreement of both of them with the observed satellite winds is very good, though both underestimate the actual wind speeds near the storm—the red feathers are observed winds of 30 kts or more. The contour lines mark the wind speed boundaries, shown from Expedition, which allows the forecasts to be scaled to the precise time of the satellite pass, a very nice feature for these tests. These were based on the h18 forecast for both data sets. We still have to overlay these by hand and georeference then all to be the same.

It seems that with the wind speeds so very close in the two data sets that the NDFD must have been either pure GFS or it was in any event very close to what was used in the NDFD.

Looking at this one and the last one below for Tehujuanepec show that the main challenge in making this type of true test of the forecast is having model forecast data for the times of the satellite passes.  Thus we have started a new program of automatically downloading the GFS and NDFD forecasts once a day and then when we see a nice display of scatterometer winds showing interesting behavior, we can go back to see which model did the best in forecasting it.  This is an ongoing project now.


No comments:

Post a Comment