Analysis of UAS Data Processing Results

Center for Geospatial Analytics at North Carolina State University

Corey White adapted from Justyna Jeziorska and Helena Mitasova

Objectives

• Interpret a processing report section by section
• Compare survey data and errors from different flights of the same site
• Recognize differences in the products generated by different software
• Explain what GCPs change in the processing and in the final products
• Identify which processing step to revisit to fix a particular error
• Explain what causes the bowl effect
• Execute profile, contour, and volume measurements in Agisoft Metashape

• Produced the products; Measured their accuracy against checkpoints
• Today: read the report the software wrote and compare products across flights, software, and control
• The lab (4A) has them measure on those products and difference DSMs in GRASS
• Screenshots are from PhotoScan 1.x; the dialogs have the same content in Metashape 2.x

Two flights

• Section 1: the dataset behind every comparison in this deck
• Two flights on the same day over the same field, plus a later winter flight

Flights, Oct 6 2015

Two consecutive flights with the Trimble UX5 over the Lake Wheeler fields.

• Same site, same aircraft, same camera, one hour apart
• Any difference between the two DSMs is error, not change
• That makes the pair a measurement of repeatability

Flight, Feb 6 2018

A winter flight of the same fields, used for the measurement examples at the end of the deck.

• Different season, bare fields, leafless trees
• The contours and volumes later in the deck come from this flight

The two October flights side by side

	Flight 1	Flight 2
Number of images	502	659
Flying altitude	110.1 m	112.4 m
Ground resolution	3.06 cm/pix	3.13 cm/pix
Coverage area	0.56 km2	0.70 km2
DSM resolution	12.2 cm/pix	12.5 cm/pix

The second flight covered more ground with more images at almost the same altitude, so the GSD and DSM resolution match within a tenth of a centimeter.

• All five numbers come from the first page of the processing report
• Altitude within 2 m, GSD within 0.07 cm: the flights are comparable
• DSM resolution is about 4 x GSD; that is Metashape’s default
• Ask: with these two reports, what else would you check before differencing the DSMs?

Reading the processing report

• Section 2: what each page of the report shows and which numbers matter
• 3B walked the four pages at a glance; here we go page by page with the two October flights

What the report contains

Report section	What to look for
Survey data	Number of images, flying altitude, ground resolution, coverage area, camera model
Camera locations and image overlap	Where the photos were taken and how many see each spot
Camera calibration	The adjusted lens model and the image residual plot
Camera locations error	How far the solved camera positions sit from the geotags
Ground control points	Control (and checkpoint) errors in meters and in pixels
Digital elevation model	The DSM sketch, its resolution, and point density
Processing parameters	Every setting used, for reproducing the run

• The report is the only record of how the products were made
• Two of these sections are accuracy statements, and only if checkpoints were declared (3B)
• The rest is provenance: enough to repeat the processing

Orthomosaic

• First page of the report: the orthomosaic sketch
• Ragged edge is where overlap ran out; the DSM will be weakest there too
• Trees along the edges will matter for the DSM comparison

Camera locations and image overlap

• Dark blue: more than nine images see each point in the interior
• One to three images at the edges: expect noise and holes there

• The planned 80 percent overlap shows up as the deep blue interior
• Anything measured near the edge is measured from one to three photos
• Recall 3A: overlap is what you bought with batteries and flight time

Flight plan

• The plan behind both flights: NEX-5 camera, 100 m, 80 percent overlap, 3.2 cm GSD
• Twelve surveyed targets spread over the polygon, edges and interior (3B’s layout rule)
• The report’s 3.06 cm GSD is the as-flown value against the 3.20 cm planned

Flight paths as flown

Flight 1

Flight 2

• Flight 2 extended the block, which is why it has 659 images and 0.70 square km
• Compare the two overlap figures when you compare the two DSMs

Ground control point errors

• Both flights fit their control to about 2 cm and an eighth of a pixel
• Same markers, same surveyed coordinates, so the tables are directly comparable
• These are control errors: they say how well the model was forced onto the targets
• 3B: control error overstates accuracy 3 to 8 x; checkpoints give the honest number

Reading one row of the GCP table

• Error (m): distance between the surveyed and the estimated marker position
• Projections: how many photos the marker was placed on
• Error (pix): RMS reprojection error of the marker over those photos

• Meters tells you about the georeferencing; pixels tells you about the marker placement
• A marker with large pixel error and few projections was probably misplaced on a photo
• Fix: go back to Stage 2 of 2B and re-place that marker

Camera calibration

During alignment the software estimates both the exterior orientation of every photo and the interior orientation of the camera, including nonlinear lens distortion. The adjusted values are in Tools > Camera Calibration and on the report’s calibration page.

• Interior orientation from 2A: focal length, principal point, distortion
• Self-calibration: these are solved from the photos, not measured in a lab
• Camera type must be right first: Frame for the NEX-5, Fisheye for action cameras

Reading the calibration parameters

Parameter	Meaning
Fx, Fy	Focal length in pixels along each axis
Cx, Cy	Principal point: where the optical axis meets the sensor
Skew	Non-orthogonality of the pixel axes (near zero)
K1 to K4	Radial distortion coefficients
P1, P2	Tangential distortion coefficients

• Cx, Cy should sit near the sensor centre: 2456 by 1632 for this 4912 by 3264 sensor
• K1 dominates radial distortion; K3 and K4 absorb what is left and can overfit
• Large or unstable values between flights of the same camera are a warning sign
• 3B: a bad lens model is the mechanism behind doming

Image residuals

• The calibration page’s second figure: residual vectors over the sensor
• Each vector is averaged over all photos for that part of the frame
• Pattern to look for: small and random in the centre, growing toward the corners
• Systematic swirls or a one-sided pattern mean the lens model is wrong

What a residual is

Image residual: the distance in pixels between where the calibration model projects a 3D point onto the photo and where that point was actually measured in the photo

• Two positions for the same point on the sensor: predicted and observed
• The vector between them is the residual; the plot averages them per sensor cell
• 1 pixel scale bar: these are sub-pixel quantities

Reprojection error

Reprojection error: the RMS image residual over all tie points, the number the report quotes in pixels; it is the same quantity the photogrammetric literature calls the RMS image residual (Agisoft LLC 2025)

• The report prints one number for the whole project; the plot shows its spatial structure
• Low reprojection error says the model is internally consistent
• It says nothing about where the model sits on the ground: that is the GCP and checkpoint tables
• Nouwakpo 2016 is the reading on how residuals and calibration drive accuracy (Nouwakpo et al. 2016)

Flight path

• The flight log drawn as a line, from the camera positions in the report
• Turns are outside the block; the camera only fires on the straight lines

Flight path in 3D

• Same track over the DSM in the GRASS 3D view
• The aircraft held altitude while the ground rose and fell under it: GSD varies with terrain (3A)
• Tree spikes at the edges are the noisy, low-overlap zone from the overlap figure

Comparing surface models

• Section 3: what changes between flights, between software, and with or without control
• Each comparison isolates one variable

Digital surface models

Flight 1

Flight 2

• Same color ramp, 90 to 130 m, so the eye can compare them
• Looking alike is not the same as agreeing: subtract them

Difference between the two flights

Difference in meters between the DSMs from the two consecutive flights. Bare fields agree to within centimeters; the edges and the trees do not.

• The honest repeatability test: nothing changed on the ground in that hour
• Fields near zero: the repeatable part
• Trees and edges at plus or minus 1 to 2 m: wind in the canopy and low overlap, not change
• 3B’s level of detection comes straight from a map like this

The same flight in three programs

• With GCPs, three programs agree on the mean elevation to within a meter
• Without GCPs, the two programs disagree with each other by 35 m and with the truth by 36 to 71 m

• With control, software choice is a second-order effect: the means differ by less than a meter
• Without control, every program lands wherever the navigation-grade geotags put it
• The 3 to 7 m standard deviation is terrain and trees, similar in all five columns
• The lab compares these DSMs in GRASS

With and without ground control

The same photos processed with no GCPs and with 11 GCPs. The shape is identical; the elevation is off by about 30 m without control.

• A pure vertical offset: the relative model is fine, the absolute position is not
• This is the geotag altitude error from 2A and 3B, usually tens of meters on consumer GNSS
• One GCP, or an RTK aircraft, removes most of it (3B, Zhang 2019)
• Not every GCP-free error is a shift: the next section is the one that bends

The bowl effect

• Section 4: the systematic error that control fixes and differencing cannot forgive
• 3B introduced it as doming; here is what it looks like in a real project

The bowl effect

Bowl (or doming) effect: a smooth, broad bending of the whole model introduced during alignment when the camera calibration is estimated inaccurately; the self-calibration trades lens distortion against a curved surface (James and Robson 2014)

• The difference between a DSM processed without GCPs and one with: a dome tens of meters high
• Nothing on the ground is shaped like this; it is the lens model gone wrong
• A mean error far from zero in a report is the detector (3B’s bias rule)
• Parallel nadir photos of flat terrain are the classic trigger: James and Robson 2014

Avoiding the bowl

• Use GCPs spread across the block, or RTK/PPK camera positions, and optimize on them
• Improve the camera calibration: a known camera model, or convergent and oblique photos so the lens can be solved
• Increase the overlap and add cross strips so every point is seen from more angles
• Post-process: fit and remove the dome against control, a last resort that treats the symptom

• Order of power from 3B: known calibration, oblique imagery, two heights, then control
• Overlap alone does not fix doming (James 2020); it helps only when it adds view angles
• The field recipe: tilt the camera 20 to 35 degrees for part of the flight and fly an orbit
• Post-processing is what you do when the flight is over and cannot be reflown

Optimization

Optimize on the camera geotags and the markers sit 36 m off. Optimize on the markers and the cameras sit 37 m off. The ground control wins.

• The same project optimized two ways; the 36 m is the geotag altitude bias
• Whichever reference you weight, the other reports the disagreement
• 2B Stage 2: first on cameras to get close, then on markers to get right
• 3B: the Reference Settings accuracies decide how much each reference is trusted

What optimization does

• A full photogrammetric bundle adjustment with the ground control as additional constraints
• Exterior orientation of every camera and the interior orientation of the lens are re-solved together
• Replaces the simple seven-parameter transform (shift, rotation, scale) that georeferencing uses by default
• The reported errors drop by orders of magnitude when the control is good

Marker accuracy in the Reference Settings is the weight: 5 mm says the targets are trusted to 5 mm. Wrong weights give a wrong adjustment (Agisoft LLC 2025)

• Without optimization the model is only moved, rotated, and scaled onto the GCPs; the bowl survives
• With it, the lens model is refit under the constraint that the GCPs are where they were surveyed
• That is why the bowl disappears with control and not with a plain georeference
• Lab: the DSM without GCPs is exactly the unoptimized case

Measuring on the products

• Section 5: what the lab does with the products
• Simple measurements in Metashape, terrain analysis in GRASS

Profiles in Metashape

Draw a polyline on the DSM or model and read the elevation profile along it.

• Measure Shape: Planar, Profile, Volume tabs
• Profile over a brush pile: 2 to 3 m above the field in the DSM
• Compare the DSM profile with the same line in GRASS (r.profile) in the lab

Contours in Metashape

Tools > DEM > Generate Contours. Contours are shapes, so they export as vector files alongside the DEM.

• February 2018 project: 507 cameras, 7 markers, 15 cm DEM, 3.75 cm orthomosaic
• Contours on a DSM follow the canopy as well as the ground; look at the tree lines
• Bare winter fields make this flight the better one for terrain contours

Volumes and areas in Metashape

Close a polygon around the pile, pick the base plane, and read the volume above and below it. The model must be watertight under the polygon.

• Volume needs a base: best fit plane, mean level, or a custom plane
• Holes under the polygon make the number wrong; this is where Close Holes (2B) matters
• Lab: measure the same pile twice with different bases and report the difference
• 3B: the volume uncertainty follows from sigma-z times the footprint area

Terrain analysis in GRASS

Slope

Aspect

r.slope.aspect computes slope, aspect, and curvatures from the DSM in one run.

• Same DSM, GRASS instead of Metashape
• Slope shows the terraces and every tree edge; aspect shows which way they face
• The noise from the DSM comparison shows up here as speckle

Curvature and shaded relief

Curvature

Shaded relief

r.relief for the hillshade; curvatures from r.slope.aspect or from the interpolation itself (v.surf.rst).

• Curvature amplifies the smallest surface detail, including the error
• Shaded relief is the quickest visual check of a DSM: artifacts jump out
• The DSM without GCPs would show the bowl as a slow gradient in slope and aspect

Advanced terrain analysis

Local relief

Sky-view factor

Addons r.local.relief and r.skyview; install with g.extension.

• Local relief removes the broad trend and leaves the micro-topography: furrows, rows, tracks
• Skyview factor is independent of light direction, unlike a hillshade
• Both are tools for subtle features on bare ground, which is where UAS DSMs excel
• 4B does the same analysis on lidar surfaces

Key terms

Processing report: the record of images, overlap, calibration, and errors from a run;

Image residual: pixel distance between a point projected through the camera model and its measured position

Reprojection error: the RMS of the image residuals over all tie points

Bowl effect: broad systematic bending of the model from an inaccurate self-calibration; also called doming

Optimization: the bundle adjustment re-run with ground control as constraints

DSM difference: between repeat flights it measures error; between epochs, change plus error

• Midterm vocabulary, and the words the lab report has to use correctly
• Control error versus camera location error: neither one is the map accuracy

Wrap-up

Assignment 4A: Analysis of UAS data processing results: measure profiles, contours, and volumes in Metashape, then compare DSMs with and without GCPs, across software, and between surveys in GRASS

• Recap: the report records how the products were made and, with checkpoints, how good they are
• Two flights an hour apart show the repeatable part and the noisy part
• Control fixes the offset; optimization fixes the bowl; software choice matters least
• The lab measures on the products and differences them in GRASS

References

Agisoft LLC. 2025. Agisoft Metashape User Manual: Professional Edition, Version 2.3. Agisoft LLC. https://www.agisoft.com/pdf/metashape-pro_2_3_en.pdf.

James, Mike R., and Stuart Robson. 2014. “Mitigating Systematic Error in Topographic Models Derived from UAV and Ground-Based Image Networks.” Earth Surface Processes and Landforms 39 (10): 1413–20. https://doi.org/10.1002/esp.3609.

Nouwakpo, Sayjro K., Mark A. Weltz, and Kenneth McGwire. 2016. “Evaluation of Structure from Motion for Soil Microtopography Measurement.” The Photogrammetric Record 31 (153): 17–35. https://doi.org/10.1111/phor.12072.