I was wondering how you people analyse your GPS logs to get plausible speed numbers out of it?

I have little confidence in the max and average speed readings I see from viewing GPS speed logs on the devices or web sites. Every speed log I have seen has the odd random peak, and a heap of peaks and troughs caused by gybes, easing off through waves, whatever. Even 10 second averages are pretty dodgy when there's the odd 40+ knot booboo sitting in the midst of the data.

My method is to pull the track logs out, filter out the peaks and troughs, maybe do some best fit lines, eyeball it and come to some plausible numbers. I generally only do that if the session "felt fast" and if I thought there might be some change to the usual numbers.

There's nothing unusual about the filtering, it's been done for decades in surveying, various kinds of GPS instruments and web servers. It's pretty standard data analysis to filter out statistical anomalies and GPS signal quality stuff. How do you do it?

Select to expand quote

Gorgo said..
I was wondering how you people analyse your GPS logs to get plausible speed numbers out of it?

I have little confidence in the max and average speed readings I see from viewing GPS speed logs on the devices or web sites. Every speed log I have seen has the odd random peak, and a heap of peaks and troughs caused by gybes, easing off through waves, whatever. Even 10 second averages are pretty dodgy when there's the odd 40+ knot booboo sitting in the midst of the data.

My method is to pull the track logs out, filter out the peaks and troughs, maybe do some best fit lines, eyeball it and come to some plausible numbers. I generally only do that if the session "felt fast" and if I thought there might be some change to the usual numbers.

There's nothing unusual about the filtering, it's been done for decades in surveying, various kinds of GPS instruments and web servers. It's pretty standard data analysis to filter out statistical anomalies and GPS signal quality stuff. How do you do it?

Gorgo what device are you using? and how are you wearing it?

For accuracy I wear a 10hz ublox device mounted on my helmet. And use doppler data for processing.
Accuracy for 2s is then within small fractions of a knot.
The beauty about ublox and locosys devices is their accuracy report. You get a +/- number, if that number is low you can be fairly sure of the result.
Track logs are less accurate than doppler data, especially for short times. I'll look for some examples
Here's a comparison between my 10hz logger on my head and a GW60 on my arm, wearing two devices as a cross reference can also help determine accuracy. So the right hand device is my logger on head, left hand numbers are GW60. So both have +/- numbers for 2s around 0.1kts but my logger is slightly better, and the 2s speed differences between them are around 0.01kts. The 10s numbers are much better



But a change in sats will affect any GPS. A watch is more prone to this if you use the under arm grip, as it changes sky view when you gybe. Any crash will probably submerge the GPS obscuring sats, so that will produce big errors. Any decent software will filter those out. Main filters being number of sats, typically set to 5, m/s2 this varies with sample rate, typically 4 for 2hz up to 16 for 10hz, and SDoP/sAcc 4, but I think it can go lower than that for devices that aren't watches. It's only that high because of the poor accuracy of watches in a gybe.

Thanks for the reply. I am using a Garmin Fenix 3 on my left arm under the wetsuit sleeve. I guess I can try mounting it on a helmet to see if there is a considerable difference. That's not my main focus but it would be interesting to try.

I was more interested in knowing if people were using instruments or software that automatically did the filtering. It seems it depends on how far you get into it and what sort of devices you're using.

I am deeply suspicious of the various claims of "... 40 knots!!!!!! Woohoo!!!" with a picture of a watch screen. I spent about 8 years working on a mapping project and I know how hard it is to get true accuracy out of a GPS. It's good to know that GPS speed enthusiasts are going the extra mile to ensure a degree of accuracy.

Under your wetsuit sleeve is a big problem, Not too bad when it's dry, but as soon as it's wet, it's almost obscure to radio waves, you're lucky to get any results at all.
You'll get much better results on top of the wetsuit, if the band isn't long enough, you can get band extenders.
It's a while since I've looked at a fenix file, I thought there was a doppler output. If there's a sat count that can be helpful in gauging accuracy. I know there isn't SDoP but does it have HDoP, (horizontal depletion of precision). That applies to track point accuracy, but can also be an indication of accuracy.

GPSResults will process fenix files and uses filters as I've said, if the file includes that data.

I agree, there's a lot of sus results out there, that's why the GPSTeam Challenge only approves devices that include accuracy info.
Unfortunately Garmin devices don't, they are popular with sailors, who are disappointed they can't use them.

What Decrepit said.

It appears that the Garmin and Suunto GPS Watches use only positional data. What sometimes appears to be Doppler speed data in analysis programs, seems to actually be a smoothed or filtered Positional output.

I thought that in farming, mining they had accuracy down to
cm/10x mm. Sure it's prob slow speed but Motion seems accurate. Trust your device and decipher program (KA72)

Select to expand quote

BSN101 said..
I thought that in farming, mining they had accuracy down to
cm/10x mm. Sure it's prob slow speed but Motion seems accurate. Trust your device and decipher program (KA72)

Dave, that sort of accuracy can be obtained but it's expensive and complicated. I think it needs a fixed base station so random atmospheric fluctuations can be monitored and adjusted for.
Gorgo has a Fenix not a motion, and he's wearing it under his wetsuit, there's a lot of difference.
The motion is a ublox device, runs at 10hz with multiple GNNS and outputs all the accuracy info. The garmin does none of that, well it may run multi GNNS I'm not sure.

KA72 is great and convenient, but it's accuracy filters are set very slack, to let all sorts of devices upload. If you want an accuracy check, use GPSResults for Garmin devices and GPSSpeadreader for .sbp and .ubx files.
KA72 would probably give results for Gorgo's under wetsuit file, which could be absolute nonsense.

Select to expand quote

BSN101 said..
I thought that in farming, mining they had accuracy down to
cm/10x mm. Sure it's prob slow speed but Motion seems accurate. Trust your device and decipher program (KA72)

Farming, mining and surveying use differential GPS. They have local base stations on surveyed points that measures errors in the GPS signal and corrects the data in real time. It's almost a different technology to that used by the general public and it gives false impression of what your basic GPS can do.

My understanding is that Doppler shift speed calculation is somewhat immune to errors in the GPS signal, but the consumer units may or may not use that.

What I don't understand is how someone with 4634 posts raises this as an issue on Seabreeze!?! Hasn't this been explained many, many, many, many times in the approved devices epic? Unless the guy is a kiter....... that would explain the 'you people' comment.

Select to expand quote

Carindale said..
...you people...

Refers to people who actually do speed logging using GPS. ie. people who know what they're doing as opposed to the general public who don't.

The question was about filtering, not devices. I did apologise for not searching through every thread but I didn't want to trawl through the entire internet. I thought I was reasonably polite in asking the question.

20 years windsurfing. 20 years kiteboarding. 6 years kite foiling. 1 year wing foiling. 30 years using GPS for tracking (mostly paragliders) and work (rail track geometry and asset surveying). 0 years doing GPS speed so quite ignorant about that.

Select to expand quote

Gorgo said..
My understanding is that Doppler shift speed calculation is somewhat immune to errors in the GPS signal,

That is true to some extent. The Doppler shift speed calculation is not affected anywhere near as much as the positional computations by atmospheric conditions, but there are still other factors which can dramatically decrease accuracy in adverse circumstances. The Doppler speed error figures in the approved devices give a very good indication of when this may have happed and how much trust we can put in the results. Critically, those figures also enable us to do very controlled evaluation tests between identical, and different devices to verify their accuracy and consistency in all the circumstances relevant to us. (see Decrepits example in his earlier post)

Select to expand quote

Gorgo said..
but the consumer units may or may not use that.

That is very true. Many consumer devices simply use the positional point information to calculate and record speed over ground. Sometimes it is very hard to get reliable information from manufacturers of such devices as to if they truly record the Doppler speed data. In many cases, they don't do it. Some devices appear to record Doppler Speed, but it is fact a smoothed (filtered) positional speed calculation, which is subject to all the same errors as the original positional data.

It does appear there's no advantage in using cm accuracy positional data over good doppler data.
I just need to do a conversion from 0.1 knots at 30kts to cm, I make that 5cm accuracy

Select to expand quote

Gorgo said..

Carindale said..
...you people...

Refers to people who actually do speed logging using GPS. ie. people who know what they're doing as opposed to the general public who don't.

The question was about filtering, not devices. I did apologise for not searching through every thread but I didn't want to trawl through the entire internet. I thought I was reasonably polite in asking the question.

20 years windsurfing. 20 years kiteboarding. 6 years kite foiling. 1 year wing foiling. 30 years using GPS for tracking (mostly paragliders) and work (rail track geometry and asset surveying). 0 years doing GPS speed so quite ignorant about that.

Fair enough. Just get ready for highly technical gps/Doppler/error files jargon discussion.

Select to expand quote

Carindale said..Fair enough. Just get ready for highly technical gps/Doppler/error files jargon discussion.

Why?
Do you want to have one?

Select to expand quote

Gorgo said..
I was wondering how you people analyse your GPS logs to get plausible speed numbers out of it?

I'm all for very reliable GPS units, worn so they have a clear view of the sky. But for foiling, I am often more interested in the "typical" speed than in top speeds. One great way is to get a good idea of typical speeds is to use a phone with software that can announce speeds, like Windsport Tracker. Since foiling is so quiet without all the slapping, it is easy to hear the announcements even through a waterproof armband (and a second zip lock bag for backup).

The phones usually don't have great GPS antennas, so you will often hear spikes. But it's pretty easy to spot numbers that randomly are a few knots higher while you're cruising along, and ignore them (many of the 40 knots peaks are linked to crashes).

On my Slingshot i84, I typically hear around 12-15 knots when cruising, and down to below 10 knots is lulls. The i76 may be a knot faster, the i99 a knot slower. The racier Starboard GT-R is very different: it needs about 12-13 knots for stable flight, and typical speeds are often around 17 knots. I've had only a few sessions on the GT-R, and tons of sessions on the Slingshots, so the difference may get larger over time.

Select to expand quote

sailquik said..

Carindale said..Fair enough. Just get ready for highly technical gps/Doppler/error files jargon discussion.

Why?
Do you want to have one?

You cheeky monkey Andrew. Those conversations quickly elevate above my level of intellect or patience. I have or have had a Canmore102, GT31, GW60 and now a Motion. As long as I have an approved device I don't care

You have it sussed.

Going though and cleaning up a backup hard drive and came access this image. I knew I had it somewhere!

For me, this is a very clear example of how the overhand/underhand grip can affect speed accuracy in the GW-60 GPS, and all 'watch' type GPS devices in general.

The following is a speed graph of a session where I was just back and forth-ing on alternate tacks. On Starboard tack the GW-60 watch was in underhand orientation (antenna facing down at the water) and on Port tack it was overhand (antenna facing up to the sky). When the antenna is facing the water it gets a high higher proportion of reflected GPS signals (multi path) and often a lower number if satellites tracked. This is what causes higher error in the calculations.

The top section of the graph is the speed trace. the bottom section is the Doppler error trace. There is a very strong and obvious pattern of double or higher error visible when the watch is in under grip. On the good tack the error values are mostly under 1. On the bad tack, the errors are regularly over 1.5 and even reaching almost 2 at times. In this particular case it didn't significantly affect the results, but it does illustrate how it is an issue we need to know about and keep a close eye on.

This is one of the many reasons that we recommend wearing your GPS on the upper arm or helmet, and why we are not particularly keen on GPS's in the watch format which are worn on the wrist.

In comparison to the GW-60 data above, here is the graph of a test session I did today on my motorbike (There was no wind. )

The data below is from two Motion LCD units fixed side by side on the tank bag. You can see that the sAcc (Doppler error) did not get above 0.5kt Per point at any time during the ride. The two units were seeing and using 16-18 satellites during this test. The data from both agrees exceptional closely both in the speed graph and the actual calculated results.

The point I am trying to illustrate is that, although the GW-60 was quite good, the Motion is very much better, even at double the logging rate (10Hz v's 5Hz). Even when the GW-60 watch antenna is facing towards the sky, the Motion error values are considerably lower., despite being recorded at double the Hz rate.

In case it's not clear, the error graph is the bottom one. The line with at the top of the window is the 1Kt error line. Most of the points are well below the 0.5kt line.



And from the same test, here is a zoomed in speed graph section, from two Mini' Motions, where the gear changes, which are sub second, are clearly visible and correlate extremely closely, supporting my observations that the Doppler speeds recorded are very accurate indeed. (The yellow is the 2 sec peak of that run)


Here is a pic of the location of the 2 pairs of Motions during this test.

Select to expand quote

Gorgo said..
I am using a Garmin Fenix 3..

That's your main problem. I used a Fenix 3 when mountainbiking and found it highly inaccurate when compared to my old bike computer with wheelsensor and a GT-31. Wore the watch on my left wrist facing upwards. My local track is 27,8 km and a mix of single track and dirt roads through woods and open fields. The bike computer and GT-31 gave the same distance, avg speed and max speed each time I rode the track.

The Fenix 3 would go as low as 26,5 km in distance and as high as 28,2 km and everything in between. Elevation reading was also of. I've seen results anywhere between 130 m and 380 m height difference. Always started and ended at my home, 20 m above sealevel, the start was nearly always accurate but when returing home the elevation would be anywhere between 5 m below sealevel to 60 m above sealevel. The GT-31 gave 185 m height difference each time and no difference in elevation between start and end.

When analyzing the GPS track the Fenix would cut corners and be of the track by 20+ meters in open fields. According to the Fenix 3 I've even been biking through the middle of a pond, through gardens and even through buildings at the side of the road. Max speed wasn't accurate either. I've seen spikes as high as 65 km/h while my bike computer and GT-31 never exceded 40.

The Fenix 3 is a nice sports watch and has some good features but I'm not impressed by the data quality and accuracy. When only used for walking or running it's a good watch. But for everyhing else which requires accurate GPS data you'd be better of with a Motion or one of the Locosys devices.

Select to expand quote

sailquik said..
Even when the GW-60 watch antenna is facing towards the sky, the Motion error values are considerably lower., despite being recorded at double the Hz rate.

You are comparing apples and oranges.

GW-60 reports SDoP values, and the Motion reports sAcc values. Both represent some estimate of how accurate the speed is, but they are calculated differently, and therefore can NOT be directly compared. The algorithms how the values are compared have not been published by the two companies (Locosys for the GW-60, u-blox for the Motion). Locosys error estimates are generally higher than u-blox error estimates. Here's one example:


The error values (lower graphs) are higher for Locosys units (green and purple), but the device with the worst accuracy in this test was a u-blox based GPS (red curve) that has significantly lower error estimates.

The GW-60 has a pretty small GPS antenna, which is probably a lot smaller than what the Motion uses. The Motion also seems to use the metal housing as a "ground plate" to block scattered signals, which increases accuracy, and usually tracks more satellites. That means that the Motion will often be more accurate than the GW-60, especially when used for windsurfing. But the difference is much smaller than the direct comparison of accuracy estimates suggests.

You are not comparing apples with apples.

Tell me if I am wrong, but in the example you gave, I am assuming that the GW-60 watch was worn on the wrist and was subject to the usual poor orientation to the Sky as they are almost always are when worn this way.

Select to expand quote

boardsurfr said..GW-60 reports SDoP values, and the Motion reports sAcc values. Both represent some estimate of how accurate the speed is, but they are calculated differently, and therefore can NOT be directly compared.

You don't know that. That is an assumption.

SDOP figures I recorded with the GW-52, in my helmet (and on my upper arm) were much lower than the GW-60 overall, and also much more consistent. They use the same chipset and essentially the same software as the GW-60, but a larger antenna and are not wrist worn.

You may be right that they are not as low as the Ublox/Motions, but from what I remember thay were very comparable when I was comparing GW-52's with a Ublox logger.

BUT, I will do some controlled testing again to compare and report my results.

Select to expand quote

boardsurfr said..
....the Motion will often be more accurate than the GW-60...

I agree. That was my point.

Select to expand quote

boardsurfr said.. But the difference is much smaller than the direct comparison of accuracy estimates suggests.

That remains to be seen, but I am not sure how relevant it really is in practice.

You make an interesting point about the Ublox red line results and comparison though. There could be other factors at Play??

While my GW-52's are charging up I found this test session from 2019.

Here are my two GW-52's compared with my GW-60. You can clearly see the difference in the reported SDOP error, and that confirms my memory that the GW-52 error values are right in the same ball park as what I am getting from the Motions.

Red and Blue line are the GW-52's. The Green is the GW-60.

I found that in the same session I was testing and comparing the original LCD Motion.

Very interesting results. The Reported error numbers are very, very similar. In the fastest run the Motion numbers were ever so slightly higher!!

Blue line is Motion, Red and Green are GW-52's.


Note that the GW-52's are seeing 10 sats, which is very good for just the GPS system. And in this test I would have been wearing the Motion in it's supplied armband on the Bicep so, as you see, I would not expect the micro variations in speed to correlate as well with the two side by side GW-52's in the helmet.

But in the other top 5 runs the error numbers were were almost identical. This is run#2.

My anecdotal evidence, comparing results from my GT31, GW52 & GW60 for the same session, was that the highest speed more often than not came from the GW52. There was never more than 0.1 kt in it, but the GW52 was generally faster. I have no idea which had the lowest error reading. I wore the watch on my upper forearm and the other two in a bag on my upper arm. I only still have the GT31, the other two broke. Both my LCD and Mini Motions provide similar results as the GW52 in comparison with the GT31.

Thanks John. It would be normal that the 5Hz device (GW-52), would somtimes give you a slightly (maybe 0.05-0.1Kts) faster run result than a GT-31, especially in the 2 sec category, because it would have more chance to select the optimum part of the speed curve that a 1Hz device when you have a typical "hill" shaped speed peak. To a lesser extent this is true also of the 10 second category. This will also be true in theory of a 10Hz device compared with a 5Hz, but the difference would often be un-noticeable.

My testing comparisons have often found slightly higher 2 sec peaks from the GW-60, just because it often has more error, and the software obviously automatically selects for that highest peak (within the error filter margin obviously)

Select to expand quote

sailquik said..
Tell me if I am wrong, but in the example you gave, I am assuming that the GW-60 watch was worn on the wrist and was subject to the usual poor orientation to the Sky as they are almost always are when worn this way.

Yes indeed, you are wrong. The example I gave was from driving test, with all antennas oriented towards the sky. One of the Locosys units was a GW-60, the other one a GW-52. Both show the same characteristics.

Select to expand quote

sailquik said..
I found that in the same session I was testing and comparing the original LCD Motion.

Excellent. The data you show, limited as they are, actually nicely illustrate some differences between Motion and u-blox error estimates.

Select to expand quote

sailquik said..

If you look at the error estimates for the single data points, they are around 0.136 for the Locosys units, and around 0.194 for the Motion, so they are almost 50% higher for the Motion.

Comparing the +/- numbers for the speed results, the reported accuracy looks very similar. But the Motion data are 10 Hz, and the Locosys data are 5 Hz, and these numbers are based on Gaussian error propagation. That means that, if the error estimates were indeed nearly identical, the numbers in the results table for the Locosys data should be about 1.4 fold higher, not roughly the same.

You data also show nicely that the u-blox error estimates vary a lot less than the Locosys error estimates. For example, the error estimates for the 10 second runs vary between 0.35 and 0.40 for the Motion, but between 0.33 and 0.56 resp. 0.29 and 0.61 for the Locosys units. Motions show about 15 percent differences, Locosys units closer to 100 percent. The 2 second numbers show a similar pattern.

Bottom line: the error estimates from Motion and Locosys 5 Hz units show clearly different characteristics. In general, u-blox error estimates are constrained to lower ranges, and fluctuate less than Locosys estimates.

Here is another example of comparing errors estimates from Motion GPS units to Locosys units, from the non-stationary portion of a driving test (see boardsurfr.blogspot.com/2021/04/measuring-gps-speed-accuracy.html):
Boxes show the interquartile range (25% - 75%), the whiskers (lines above and below) extend 1.5 times the interquartile range above and below. Circles show outlier values.

Here are the corresponding numbers:
The Locosys units show a much larger spread, and the mean error estimate was almost twice as high as for the Motions. Basically, when the GPS reception quality degraded, for example because of trees next to the road, the Locosys units would shown a large increase in error estimates, while the Motions shown no or only a small increase.

The test was done with 7 GPS units (2 Motion, GW-60, GW-52, and three u-blox protoypes), so it is possible to calculate an approximate measured speed error by calculating the difference to the median speed value. Here's the distribution of errors for all 7 units:

It shows that in this test, the GW60 and GW52 performed had roughly the same speed accuracy as the Motions (both the GW52 and the GW60 did better than one of the Motion units, and worse than the other). The lower plot used a speed minimum of 10 knots, so comparisons to the graph and table above are only approximate. The brown 50% compares to the median, the green 20% bar compared roughly to the 3rd quartile value (which would be 25% in the lower graph). This comparison shows that the error estimates that the GW60 and GW52 units give are more pessimistic than the estimates that the u-blox based Motions give.

There's three devices in this chart: black, green, red. You can barely dissociate green from red but what the hell is wrong with black?


All three were on one of my car's heating duct.

Red/green are finished Mini Motions, they gently warmed up due to their thermal mass, reaching what I'd scientifically call "warmish".
Black is a bare-PCB Mini Motion which reached what I'd scientifically call "scorching hot".

All three devices before heat was on:

Funny hey.

Here's an error estimate comparison from a test drive that Mike and I did almost 3 years ago when we were at Lake George (on top of the car, so no heat ducts in play):


It shows quite nicely that the Locosys (blue) SDoP vary a lot more than the u-blox (red) sAcc values. It's not just the number of satellites, since the Locosys has about 2x lower error values than the Motion some times, and 2x higher error estimates at other times. The variation within sub-second time frames is also a lot higher for the Locosys, which is likely to to pretty strong filtering of the u-blox sAcc data (which becomes evident if you look at how slowly sAcc increases after crashes).