I am a surveyor and have done several small projects and test flights with my P4P with GCPS - processed with Photoscan. I have had really impressive results horizontally but the verticals have never worked. My processed horizontal and vertical residuals have come out well but when you pick points or contour the elevations are way off. I have gone through the standard protocol(s) as far as processing and suspect that there is something big that I am missing. Any advice would be greatly appreciated.
Surveying and Mapping with P4P and Photoscan
- Thread starter surveypilot
- Start date
Can you give some basic information like the number of ground control and check shots that you're using, the size of the site, number of rays per point, and distance between GCPs?
Thanks for this. Not to get too far into the weeds but I mostly wanted to see if the Phantom 4 Pro would produce accurate vertical results and/or if there is some basic setting in photoscan that I'm missing. The sites were approximately 2 to 5 acres and had five or six Ground Control Points. Is there a difference between the vertical and horizontal processing or accuracy? If I am getting great horizontals visually and checking into indeoendent check points, does that indicate that my overall gcps and processing is solid? Thanks!
Did you run a processing report? Do any of your GCPs have large errors in Table 4? Vertical accuracy is generally going to be 2-3x the horizontal accuracy. How far off are you reading? Are we talking feet or inches?
We routinely get RMSE values of less than 0.1' for vertical using our P4P, so it's definitely possible.
We routinely get RMSE values of less than 0.1' for vertical using our P4P, so it's definitely possible.
The verticals are off a few to several feet. The H/V accuracies were ultimately decent - a few to several cms or so and the contours and dtm look relatively correct with themselves. I am looking for low hanging fruit, as far as others experiencing this. I have been away from the aerial processing for a while so I am sort of lost on the lingo. I am getting back into it and will post more info. Thank you.
Since you're looking for low hanging fruit, are you doing the following procedure in order on PhotoScan?
I hope this helps...
- Load photos
- Import Exif from Reference Pane
- Tools - Camera Calibration
- Align Photos from Workflow menu with high accuracy and both generic and reference selected
- Optimize camera alignment by selecting Settings in Reference Pane
- Click the Optimize wand icon in the Reference Pane
- In the Optimize Camera Alignment window, select the top five checkboxes on the left, the top four on the right, and 'Fix rolling shutter'
I hope this helps...
Hey thanks for this. It's strange because I ultimately got good residuals on both horizontal and vertical and it was just the Contours and the spot elevations/check points that were off - i'll try this.
The P4 altitude drifts downwards over a 20 minute flight by 3 metres due to the sensor warming up. If you do not use GCP's or a bolt on RTK GPS then you will end up with slopes that don't exist.
I have the data to support this if anyone doubts. It was a bit of a shock. You can prove this yourself without any additional sesnor if you fly a mission with the mapsmadeeasy mappilot app and have a look at the beginning and end altitude in the log file. most missions end 3m higher than they started! indicating the drone thought it was 3m higher when it hit the ground at the end, i.e it has drifted donwwards, but thought it was at the same altitude.
I have built a cradle and timing capture device for the Emlid Reach RTK/PPK unit. When fitted on my Phantom 4 it captures precise timing and raw GPS data for post processed RTK (can capture RTK but why carry the extra weight?).
With this on my P4 I get sub 50cm agreement between my orthomosiac and my reference points (if I use the reference points as GCP's that is a different story, I am just testing non GCP mapping).
With a single GCP at the launch site and then applying a global shift this comes down to sub 25cm.
These results were obtained using photoscan for processing.
I have the data to support this if anyone doubts. It was a bit of a shock. You can prove this yourself without any additional sesnor if you fly a mission with the mapsmadeeasy mappilot app and have a look at the beginning and end altitude in the log file. most missions end 3m higher than they started! indicating the drone thought it was 3m higher when it hit the ground at the end, i.e it has drifted donwwards, but thought it was at the same altitude.
I have built a cradle and timing capture device for the Emlid Reach RTK/PPK unit. When fitted on my Phantom 4 it captures precise timing and raw GPS data for post processed RTK (can capture RTK but why carry the extra weight?).
With this on my P4 I get sub 50cm agreement between my orthomosiac and my reference points (if I use the reference points as GCP's that is a different story, I am just testing non GCP mapping).
With a single GCP at the launch site and then applying a global shift this comes down to sub 25cm.
These results were obtained using photoscan for processing.
If you're not using GCPs or controlling your site, then you're just taking pretty pictures, you're not surveying.
I agree, however there is a difference between providing control and laboriously adding GCP's to constrain a poorly positioned survey platform. Two GCPs to confirm scale, orientation and position are adequate with a PPK GNSS positioned aerial platform.
We generally over-control because we need to provide check shots for an accuracy report. We do hope to prove to ourselves that our eBee RTK is capable of repeatable <5cm accuracies with a minimum of GCPs, but we need to test more projects to get that warm fuzzy feeling. I like your Reach approach and would love to build a system like that, but we generally buy off the shelf/ turnkey equipment for our needs.
Also, I'm not sure if you saw this post, but you can get some great accuracies with a minimal amount of well placed GCPs. This was all performed using a 4 Pro: GCP Assessment
Thanks for the link. Would be interested in repeating the linked test (or maybe I find a similar quarry) and do the analysis with and without the reach PPK.
I keep running into the accuracy issue. To be clear: As a survey company / drone service providerit is absolutely beholden on you to deliver defensibly accurate positioning. With the projected drone market to be over 1 billion active drones (drone in use in business) by 2030 not all of them will be operated within survey companies or by drone operators. In Australia the rules for commercially operating sub 2kg (4.4lb) drones have been relaxed. Now engineers, irrigators, farmers, fishermen are operating them and making maps. My goal is to develop tools that will allow these sub 2kg operators to deliver information to them that is fit for purpose.
To give an example. In a vineyard with the vines spaced at 1m intervals 30cm RMSE and 2.5cm (or even 5cm) opixel size is enough for the grower to derive insight from maps and actions on the ground at an individual plant level. With a low cost owner operated system and an online processing package like MapsMadeEasy combined with QGIS it is possible (even easy) for the grower to develop a 4D picture of the grape season. Yes, GCP's can be used, but are they necessary in this usecase?
I'm very surprised by how static the DJI elevations remain through that graph. Very interesting!
You're absolutely correct that the number of non-surveyors representing themselves as such is an issue and will only continue to get worse. We have sub-contractors from national firms that have moved into UAV surveying and produce absolute garbage. I mean rookie level mistakes. These are companies with certified photogrammetrists, licensed surveyors, and civil engineers, but they're giving the UAV to the young recreational pilot and having him push buttons in Pix4D and churning out crap.
Fortunately, there are other professionals, like you, that are trying to combat this pandemic of bad data. I hope the advances in RTK/PPK technology will start to correct some of the issues, but it could cause additional levels of trust to be placed in "black box" systems, without the fundamental knowledge of aerial surveying and photogrammetry. I believe that the ASPRS UAV mapping certificate will help vet some of the service providers and legitimize UAV surveyors that have the experience and knowledge to provide accurate projects.
I guess I should define "surveying" as precision mapping for my examples and what I'm trying to achieve. Yes, there are many use cases where having <3cm accuracies are not necessary and where the variability in the final product is acceptable. If you want to take some ballpark measurements and can live with relative dimensions, then you're probably fine. Our clients are design engineers and we're looking at providing as much detail as we can, with a high level of repeatability and accuracy.
You're absolutely correct that the number of non-surveyors representing themselves as such is an issue and will only continue to get worse. We have sub-contractors from national firms that have moved into UAV surveying and produce absolute garbage. I mean rookie level mistakes. These are companies with certified photogrammetrists, licensed surveyors, and civil engineers, but they're giving the UAV to the young recreational pilot and having him push buttons in Pix4D and churning out crap.
Fortunately, there are other professionals, like you, that are trying to combat this pandemic of bad data. I hope the advances in RTK/PPK technology will start to correct some of the issues, but it could cause additional levels of trust to be placed in "black box" systems, without the fundamental knowledge of aerial surveying and photogrammetry. I believe that the ASPRS UAV mapping certificate will help vet some of the service providers and legitimize UAV surveyors that have the experience and knowledge to provide accurate projects.
It's interesting PPK / RTK increases the opportunity for bad data if it is processed without thought or understanding. In my view the issue is around drawing the line between what can be systemised/automated vs where thought is needed. Reliable automatic image recognition for large permanently installed machine readable GCP's will make routine mapping for fixed sites a piece of cake, so much so that surveyors / UAV professionals have to ask themselves, 'do they really want to do that work, or help the operator make their own defensible system?' For thos einstallations the next big improvement is not ppk or rtk, but incident light sensors, which only cost a few dollars, but need to be integrated and processed properly. I am enjoying the challenge set by the Aus 2KG law of seeing just what can go on a sub 2kg drone.
- Joined
- Dec 17, 2017
- Messages
- 22
- Reaction score
- 4
- Age
- 57
I agree absolutely. End of the day each quarry will have standard fixed gcps and will do the flight by themselves. Photos will be sent via dropbox or something else.
The big target will not be the photogrammetry, but to create a normal map out of the photogrammetry. I am quite sure that this will be done by professional surveyors in future.
regards
Martin
I know I'm trying to rekindle an old thread hear but I read that the altitude deviation from the Phantom 4 getting warmer is most pronounced early on in the flight. This graph only shows the deviation over a 10 minute interval? It would be much more helpful in my opinion to see the deviation for two flights using two full batteries over the entire length of the flight. Does the deviation continue indefinitely all the way to the end of the flight or does it slow and eventually flatten out at some point?
Last edited:
