11
« on: January 06, 2021, 02:21:31 PM »
Hello together,
For my master thesis I try to produce some digital surface models of large (400km²) high mountains and snow-covered areas around Davos.
I am using around 400 aerial images from the camera Vexcel UltraCam Eagle M3. I have the calibration report with a focal length of 122.7mm and a pixel size of 4*4µm (see attachement 1). I have the data from three different years (March 2018, April 2019 and March 2020) amd furthermore the IMU-coordinates from the airplane and the angles (Omega, Phi, Kappa).
For the last time in 2020 I also have 39 GCPs from GNSS with a high accuracy.
The coordinate system, which I am using is the CH1903+ / LV95 (EPSG:2056) with the Swiss 2004 geoid. I work with Agisoft Metashape Professional 1.6.5.
The aim of this research is to produce a DSM with a high resolution and a high accuracy (RMSE Z 20cm), but without control points. The reason for such is, that the effort for measuring many GCPs is too high and the research area is also often endangered by avalanches. The problem is that offsets are occurring in the elevation depending on the settings. I can see the difference on the 39 GCPs, which are used as check points.
If I produce a DSM with these settings: No precalibration, with IMU and with the angles, I will have an offset of around 1 meter (see attachement 2).
When using precalibration with a f-value of 30675, the focal length of 122.7mm and the pixel size of 4*4µm and the other values of zero, I have an offset of around 2,2 meters (see attachement 3).
After that, I optimized the first model without precalibration, IMU and angles with the 39 GCPs as control points. The accuracy of this model was very high (RMSE Z 15cm). It was noticeable, that the f-value changed to 30658. So, there is a difference of the f-value to the calibration report of 17 pixels.
In the next step I produced a new model with a fixed f-value of 30658, the IMU-coordinates and the angles, but without some control points. The results were quite perfect with a Z-RMSE of 9cm and a total RMSE of 29cm (see attachement 4).
So I worked with the same settings (f-value 30658) for the year 2019. There it didn´t work and I had an offset of the Z-value of 35 cm. There I do not have many GCPs, so I can not optimize the model to figure out the right value for f. Though my supervisor produced the 2019 model without precalibration, but with IMU and the angles and there it worked better with a Z-RMSE of 15cm. We checked the accuracy on snow-free areas like streets with a DSM from a UAV and an airborne laserscan.
With the 2018 model it did not work again. When working only with the IMU and the angles, there was an offset of some meters as well. With a precalibration like it is presented in the calibration report, the difference increased. There we had some GCPs, which we could use as control points to optimize the model.
In summary, we worked with three different procedure for the same process. Normally, it should work, that we only need the IMU, the angles and maybe the precalibration for results with a high accuracy. One solution for the next year could be that only 3 GCPs will be used as control points to optimize the model. But it would be even better, if no GCPs were necessary.
I guess, that the focal length changes during the flight because of some differences in the temperature. I think Metashape works with a constant focal length, which never changes So maybe these little differences in the focal length in combination with the high fly altitude are the reasons for this mistake.
Or do you have some other ideas?
Best regards,
Leon