Why would a real time calculation of a data set yield different results than the same calculation of the data later?
> Why would a real time calculation of a data set yield different results than the same calculation of the data later?
Well, most fundamentally, when a vector is inserted into a network with other conditions, one can test the vector residuals to see whether they are consistent with the a priori estimates of variance or not. The whole problem should be familiar from post-processed GPS solutions where processor estimates of uncertainties tend toward optimism and away from reality.
If you were asking about a PPK solution as opposed to RTK, the differences would be the result of the processing algorhithms.
The tendency toward optimism is found in the flawed application of basic math. Nearly all RTK units report a QC number that continues to improve through an occupation. The problem is they abuse the square root of the number of observations by counting each epoch as a separate observation. The QC numbers continue to improve long after the point of diminishing returns. Nearly every brand out there reports quality expectations that should be 2.5 times what is shown...
> ...no residuals.
>
Yes. That is why an adjustment with no residuals borders on the absurd.
> > ...no residuals.
> >
>
> Yes. That is why an adjustment with no residuals borders on the absurd.
It would if you forget about the fact that error propagation can be done on even a minimally constrained figure using the standard errors of the observations. A least squares adjustment program, such as Star*Net, that can also perform error propagation can provide useful analysis even when there is no redundancy to adjust (assuming that the standard errors are fairly realistic). The way that Star*Net works is that it adjusts the survey measurements (or not if there is no redundancy to adjust) and then performs error propagation on the results using the a priori standard errors assigned to the observations.
Error analysis is an important reason to use least squares survey adjustment software.
Wouldn't each epoch yield it's own vector? The longer it runs the more observations and the closer the mean should be to the point. That is if you are getting clean signal. I've seen observations where the precision doesn't get better as you let it go on, even gets worse and lose lock. Tells me I'm not receiving clean signal (at least in my mind).
> The longer it runs the more observations and the closer the mean should be to the point. That is if you are getting clean signal.
If you were to export the RTK vectors to a least squares adjustment and had sufficient redundancy, you'd be able to see for yourself how unrealistic the processor estimates of uncertainty are and how much they need to be scaled up (meaning: how much larger the uncertainties really were than what the processor estimated).
> Wouldn't each epoch yield it's own vector?
As far as I know, Yehuda Bock's RTD products are the only commercially-available applications that compute individual-epoch vectors. But since the whole RTK concept is a bit fuzzy in my mind, there may be others out there capable of treating each epoch as a distinct vector.
The way you dis RTK I thought you'd think a couple tenths was tight. It's more like 0.02 to 0.07 feet (horz). If it's a couple tenths probably poor observation conditions, probably kill the lock. I like to see it during the observation at 0.02 to 0.03 but 0.04 to 0.06 is all you can get sometimes. When it's more than that you are probably in a poor environment for signal. If doing topo and it's not that critical I'll accept it at 5 epochs unless the vertical is more than I like for the purpose of the topo. I would expect if I upgraded to GNSS I would see improvement.
Some do topo at 1 epoch, I've never done that but it would save time. If you collect on the move 1 epoch is the only choice.
> The way you dis RTK I thought you'd think a couple tenths was tight. It's more like 0.02 to 0.07 feet (horz).
But you're quoting the processor-estimated standard errors of the coordinates, right?
I think the RTD can solve the vector without the requirement of initialization. There still needs to be a communication between the rover and the network (or base) for real time. I looked into it quite a bit years ago. I was convinced it is better than RTK but they just never made the market penetration needed. The big boys like Trimble kept them from going mainstream.
I'm not certain but I think it is RTD that guides the smart weapons and they make most of their income from DoD contracts. I tried to get Utah to go RTD when they set up the state RTN. No go, Trimble got the contract for a lot more money. I was able to get the Utah RTN to be an open system. You could set up a RTD server and go forth. I even had it tested. I just didn't have the money or was able to take the risk that I'd get enough customers to make it pay, so I let it go.
Maybe I'll look at it again, cell phone coverage is much, much greater now than before. But I still do much work where the cell phone doesn't.
Another thing that made me reluctant was all you got in the file after an RTD solution was the coordinate. Maybe they've fixed that.
So what do you need? I can set up a base and a rover antenna on a tripod. I can do as many 3 min RTK solutions over, say a day, as needed on the same point. I'll send them to you and you can run them in StarNet. Maybe do some short (5-15 epoch) observations also to compare. I know I can run static at the base. Not sure if I can collect static at the rover and do all these RTK obsetrvations at the same time (might be RTK and infill). I could collect static the next day for as many hours as you want to compare a static vector. I suppose we could run OPUS on the two points but that introduces the error in the original OPUS station coordinates.
Do you just need the dx, dy and dz (ECEF) for each vector and the covariances. Do you want feet or meters? Or for coordinates do you want lat, long and height (or elevation) or just plain old N, E and Z?
> So what do you need? I can set up a base and a rover antenna on a tripod. I can do as many 3 min RTK solutions over, say a day, as needed on the same point.
Well, you'd want to set up the base and rover in conditions that resemble how you actually use RTK, namely:
- in a setting that does have sources of multipath (I assume),
- without regard to changes in satellite constellations other than avoiding some maximum value of PDOP.
I assume for rural work, you may often have vectors longer than a couple of miles as I do, and so would want:
- rover and base to be 1.5 miles apart or more, but in any case separated by at least a couple of thousand feet.
Collect positions based upon your typical session. If three minutes is typical, collect positions from three minutes worth of data. The important thing is that they not just be back to back, but be distributed throughout the day under different constellations, including sessions with satellites rising or setting, as would presumably happen in your typical field work.
Then, import the vectors into whatever network adjustment software you are using and adjust all of the vectors to nominally the same point. Since you will be only interested in checking the uncertainties that the RTK controller generates, you don't particularly care about the centering errors of the rover antenna (although they are significant in the final uncertainty of the RTK position).
The network adjustment software will either automatically scale your RTK vectors covariances or will prompt you for a value. Either way, note the value that must be used to pass the chi square test. That is a measure of how optimistic the processor estimates were.
It would be nice, but probably not absolutely necessary, to have a couple of excellent static baseline solutions between the same points to get a very good control value to compare the RTK vectors to.
If for some reason you don't have network adjustment software that is up to the task, if you can export the vectors in either Trimble Geomatics Office Format or G-File format (with Std Error and Correlations on the "D" line), then just post them here (along with the latitude and longitude of the base and zone of Utah SPCS) and various posters, including myself, would probably be willing to do the adjustment for you in Star*Net or some other adjustment software.
> Error analysis is an important reason to use least squares survey adjustment software.
I agree completely.
I just don't see myself running least squares adjustments on 2 point control networks, with only one observation, on a frequent basis.:-D
> > Error analysis is an important reason to use least squares survey adjustment software.
>
> I agree completely.
>
> I just don't see myself running least squares adjustments on 2 point control networks, with only one observation, on a frequent basis.
Well, two points connected by one GPS vector is hardly an uncommon situation for error propagation. Typical uses of that feature would include positioning a slew of control points by radial GPS vectors and studying the uncertainties in the positions before connecting them by conventional measurements. It is also very convenient to be able to examine the uncertainties of points in a GPS topo via error propagation.
> I just don't see myself running least squares adjustments on 2 point control networks, with only one observation, on a frequent basis.
This is what I have been describing all along. Two points and one observation.
> Well, two points connected by one GPS vector is hardly an uncommon situation for error propagation. Typical uses of that feature would include positioning a slew of control points by radial GPS vectors and studying the uncertainties in the positions before connecting them by conventional measurements. It is also very convenient to be able to examine the uncertainties of points in a GPS topo via error propagation.
This is what I think you are describing. Clearly more than 2 points. Clearly more than one observation.
To quote your earlier statement:
>Except it's assumed that the reason why least squares adjustments are typically being used in the first place is that there IS redundancy, i.e. is something to adjust.
> This is what I think you are describing. Clearly more than 2 points. Clearly more than one observation.
>
That situation is functionally identical in that each coordinate is computed from exactly one GPS vector = one observation (df=0). The actual number of coordinates doesn't really effect the situation or utility of the analysis.
Let me put the question to you a different way. How do you visualize or study the uncertainty of coordinates determined by any method with minimal constraint (df=0) if not by using the error propagation functions of some adjustment software?
> The important thing is that they not just be back to back, but be distributed throughout the day under different constellations, including sessions with satellites rising or setting, as would presumably happen in your typical field work.
Probably random sampling at 0.5 hr. intervals (= 17 total for 8 hrs.) would work just fine. If the PDOP value is excessive at the time for the session, log the next one 0.5 hours later. What you want is a uniform distribution of sessions across the work day to make an unbiased sample, as opposed to several sessions taken in succession under very similar conditions, all containing similar biases.
> That situation is functionally identical.
But not identical in every sense.
> Let me put the question to you a different way. How do you visualize or study the uncertainty of coordinates determined by any method with minimal constraint (df=0) if not by using the error propagation functions of some adjustment software?
I am asserting that a 2 point survey (read the entire job contains only 2 points) is not the norm for the practice of land surveying. Even establishing the positions of two points typically requires the production of a multitude of related positions. In my experience a job including only a single line (with a single observation) (df=0) is not common in the practice of land surveying.
Even if I was locating the position of two points you could bet your straw hat, that I (or any prudent land surveyor) would measure more than once (df>0).
Maybe 2 points surveys are normal in Texas?
Do you have a hard drive full of those two point jobs? Field books with one page of notes detailing the single observation from point 1 to point 2?
Sir, you are ignoring my position to create a debate.
We agree on this point...
>Except it's assumed that the reason why least squares adjustments are typically being used in the first place is that there IS redundancy, i.e. is something to adjust.
That's what I see on the controller as the horizontal precision. It's just a number. My experience through the years when I've been able to compare to static vectors is the numbers are reasonable but, how good is the static vector? I've had multiple long time static observations on the same point show a spread also at times.