PaperlayerView exact revision on arXiv

Source-linked AI summary

The Impact of Processing Parameters on High-Accuracy Measurements in UAV Photogrammetry

Paweł Ćwiąkała, Edyta Puniach, Elżbieta Pastucha, Wojciech Gruszczyński

arXiv:2610.01438v1cs.CV

TL;DR

UAV photogrammetry processing settings remain insufficiently examined despite the need for high accuracy in surface-change analysis. This study assesses how parameter choices relate to product and deformation accuracy, finding that optimized settings reduce systematic errors and motivating specific workflow recommendations.

  • Problem

    The influence of processing parameters on UAV photogrammetry accuracy remains insufficiently studied, despite its use for high-accuracy surface-change analysis.

    Source: §1 · pp. 1, 3, §4 · p. 22

  • Method

    The study compares baseline and enhanced processing variants and examines systematic error patterns and deformation accuracy.

    Source: §2.4 · p. 8, §3.1 · p. 19, §4 · p. 23

  • Results

    Systematic error magnitude was larger on average for the baseline than the enhanced variant, and parameter optimization affected displacement accuracy more than derived tilt changes and horizontal strains.

    Source: §3.1 · p. 19, §4 · p. 23

  • Takeaways & Limitations

    For high-accuracy measurements, the authors recommend at least five well-distributed, high-quality GCPs, additional corrections, and PPK referenced to a local base station.

    Source: §5 · p. 23

    The conclusions reflect the tested measurement conditions and may differ with study-area size, image resolution, or GCP-coordinate accuracy.

    Source: §1 · p. 4

Abstract

from arXiv · show

Unmanned aerial vehicle (UAV) photogrammetry is increasingly used in applications requiring high accuracy, such as determining ground surface changes caused by landslides, mining, or microrelief transformation. While acquisition strategies have been widely studied, the influence of the processing workflow-particularly Bundle Block Adjustment parameter settings-remains insufficiently explored. This study addresses this gap through a systematic, full-factorial evaluation of 768 processing variants applied to ten UAV datasets collected over 1.5 years in a 220 ha study area. Eight key parameters were analysed. The results show substantial variability in final 3D accuracy: the best performing variant achieved a root mean square error (RMSE) of 16 mm, whereas the weakest reached 303 mm. The most influential factors were the number of ground control points, the application of additional camera calibration corrections, and the use of the Post-Processing Kinematic GNSS method for determining camera projection center coordinates. The study also evaluates how workflow optimization affects the accuracy of displacement, tilt changes, and horizontal strain determination. While random displacement errors remained stable (RMSE of ~6-7 mm), systematic errors were significantly reduced by over half in all axes, with vertical median absolute error decreasing from 14 mm to 7 mm in the optimized configuration compared to the baseline previously used by the authors. This study provides the first large-scale, practice-oriented assessment of how processing parameter selection shapes the accuracy of both photogrammetric products and deformation indices determination. The results offer actionable guidance for developing more robust and repeatable UAV photogrammetry workflows tailored to high-precision monitoring.

1. Introduction

Although UAV photogrammetry is widely used, the influence of processing parameters remains less systematically studied than acquisition strategies. This study evaluates that gap and offers guidance for accuracy-oriented workflows, while noting that its conclusions are bounded by the measurement conditions tested.

1. Introduction

  • Prior research emphasized acquisition strategies, GCP quantity and distribution, and GNSS positioning, while many studies used areas under 10 ha and few measurement campaigns.Research also examined camera self-calibration and software, but published processing-parameter details were often limited.Source: §1 · pp. 2–3
  • The study addresses the lack of systematic evidence on individual processing parameters and their consistency across datasets, including the influence of additional camera-calibration corrections.The authors identify most prior papers as lacking detailed processing-parameter information and note that the additional corrections over the standard Brown model had not been studied.Source: §1 · p. 3
  • 768 processing variants across ten UAV surveys over 1.5 years and 220 ha form the study’s systematic evaluation of parameter effects.The work also assesses how parameter selection relates to photogrammetric-product accuracy and deformation-index determination.Source: §1 · p. 3
  • The conclusions are specific to the tested measurement conditions and may differ with study-area size, image resolution, or GCP-coordinate accuracy.The authors present the work as practical guidance for more robust, repeatable workflows for deformation monitoring.Source: §1 · pp. 3–4

2. Materials and methods

The methods combine repeated UAV surveys and controlled bundle-block-adjustment variants to evaluate photogrammetric accuracy and deformation indices. The evaluation compares processing configurations against check points and assesses displacement-derived products.

Study area and data acquisition

  • Ten surveys across a roughly 220 ha study area used a consistent flight plan, and a permanent, disjoint GCP/CP network supported processing and independent accuracy assessment.The surveys ran from May 2022 to September 2023; the network coordinates were rechecked before each series.Source: §2.1 · p. 4

Processing variants

  • The full-factorial design tested 768 variants per measurement series, including camera-positioning methods and uncertainty settings, GCP configurations, and calibration and tie-point settings.The tested options included GNSS PPK or RTN camera coordinates, estimated or fixed camera-coordinate uncertainties, and optional additional camera-calibration corrections.Source: §2.2 · pp. 6–7

Bundle block adjustment

  • Each variant received an initial and final bundle block adjustment, with gross exterior-orientation errors reviewed before final settings were applied.The initial adjustment used common settings across variants; a seven-parameter similarity transformation was applied after final adjustment when required.Source: §2.2 · pp. 7–8

Accuracy evaluation

  • Variant performance was assessed using mean 3D RMSE at common check points, while t-tests and Bartlett’s tests examined parameter effects on component error means and variances.The component analyses used along-track, across-track, and elevation errors; parameter effects were also examined among the best-performing 10% of variants.Source: §2.3 · p. 8

Deformation indices

  • Deformation assessment compared the baseline settings with the variant having the lowest mean 3D RMSE, using multitemporal rasters and moving registration windows to derive displacements.The selected intervals represented different vegetation seasons; December 2022 data were excluded because snowfall prevented accurate automatic displacement determination.Source: §2.4 · pp. 8–9, §2.4.1 · p. 9
  • Filtered displacement differences between neighboring grid nodes, divided by node spacing, yielded tilt changes and horizontal strains; spline cross-validation separated systematic from random displacement errors.Outliers were excluded from RMSE calculations, and systematic and random errors were summarized using MAE and RMSE, respectively.Source: §2.4.2 · pp. 10–11, §2.4.1 · p. 10, §2.4.3 · p. 11

3. Results

Processing-parameter choices produced wide variation in checkpoint accuracy, with the best-performing variants far more accurate than the worst. The analysis also found that an enhanced configuration reduced systematic displacement errors, while parameter optimization had no practically significant effect on tilt-change or horizontal-strain accuracy in the analyzed cases.

3. Results

  • In the worst-performing variants, using one GCP and omitting additional corrections, particularly together, was associated with poor results; parameter settings affected elevation errors more than horizontal errors.The variant ranking by 3D RMSE was nearly identical to ranking by elevation RMSE.Source: §3.1 · pp. 11–12, Figure 4 · p. 12, Figure 5 · p. 12
  • Every top-decile variant used additional corrections and more than one GCP; over 90% used PPK, and over 70% used transformation and 2-pixel tie-point accuracy.GCP coordinate-accuracy choice and tie-point filtration were not determining factors for achieving the best results.Source: §3.1 · p. 13, Figure 6 · p. 13
  • Mean 3D RMSE ranged from 15.9–19.8 mm among the top 10% of variants to 64.8–303.4 mm among the worst 10%.These ranges are calculated across analyzed intervals for check points.Source: Figure 4 · p. 12, §3.1 · p. 12, Figure 6 · p. 13
  • Additional corrections mattered across error components; PPK and its estimated observation weights benefited AL, transformation mattered for ACR, and PPK plus more GCPs benefited EL.The parameter settings that minimized each component's errors were not mutually exclusive.Source: §3.1 · p. 13, Figure 8 · p. 15
  • For EL errors, using more than one GCP reduced mean error at least fivefold and standard deviation more than twofold; only 23 GCPs brought mean EL error near zero, at 1.7 mm.Practical differences were defined as exceeding 2 mm; the influence of settings on AL and ACR was smaller and mainly reflected in standard deviation.Source: §3.1 · pp. 14, 16, 17
  • The selected enhanced variant used PPK, 23 GCPs, 15 mm GCP-coordinate accuracy, additional corrections, 2-pixel tie-point accuracy, estimated camera-position accuracy, no tie-point filtration, and transformation.Relative to baseline, the listed changes were PPK instead of RTN, estimated rather than 50 mm camera-position accuracy, 15 rather than 25 mm GCP-coordinate accuracy, and filtration off rather than on.Source: §3.1 · p. 18
  • Random displacement-component RMSE was 6–7 mm for both variants; enhanced-variant systematic-error MAE was approximately 2 mm on each horizontal axis and 7 mm vertically, versus baseline values of 4 mm and 6 mm horizontally and 14 mm vertically.Systematic errors varied gradually across neighboring nodes, and the paper attributes them to UAV photogrammetry products rather than the displacement-determination method.Source: §3.1 · p. 19
  • Parameter optimization had no practically significant effect on tilt-change or horizontal-strain accuracy in the analyzed cases, as first differences suppress slowly varying systematic errors.The authors caution that this does not mean parameter selection or product quality has no effect on these derivative indices.Source: §3.1 · p. 21, Figure 13 · p. 21, Table 6 · p. 21

4. Discussion

The discussion identifies processing choices that materially affect UAV photogrammetry accuracy, while noting that parameter optimization matters less for tilt and strain than for displacement. The conclusions are bounded by the study’s single site, ten datasets, and 10 mm GSD.

4. Discussion

  • The study examines processing factors that prior work had not addressed, including Tang et al.’s additional camera corrections, across a large area where comparable studies were more limited.The study also analyzes post-BBA transformation, assumed coordinate accuracies, tie-point accuracy, GCP errors, and tie-point filtration.Source: §4 · p. 22
  • At least five high-quality, well-distributed GCPs are recommended; for areas around 200 ha or larger, a density of about 1 GCP per 40 ha substantially enhances accuracy.Increasing GCP density further had relatively minor effects on vertical accuracy and negligible effects on horizontal accuracy when other processing parameters were appropriate.Source: §4 · p. 22
  • PPK with a local base station outperformed RTN for image projection center determination, especially vertically, under the tested short-baseline and stable-base-station conditions.The authors recommend estimating a priori image projection center accuracies from GNSS post-processing rather than assigning uniform values.Source: §4 · p. 22
  • Additional camera self-calibration corrections improved accuracy over the standard Brown model, with the greatest influence on elevation and a smaller effect on horizontal coordinates.With one GCP, the corrections may be practically significant for horizontal position perpendicular to the flight path (ACR).Source: §4 · pp. 22–23
  • A seven-parameter post-BBA transformation can slightly improve ACR and elevation accuracy, while setting tie-point accuracy to 2 px rather than 1 px also matters in those directions.The authors suggest the transformation may address a small offset between coordinate systems associated with GCPs and image projection centers.Source: §4 · p. 23
  • Under the tested measurement conditions, assigning GCP coordinate accuracy between 15 mm and 25 mm had no practical effect, and tie-point filtration had no significant impact.The authors note that filtration adds time and work despite being recommended in a USGS report.Source: §4 · p. 23
  • Parameter optimization affects displacement accuracy more than tilt-change and horizontal-strain accuracy, as spatially correlated systematic errors vary gradually across the area.Optimization is more critical for vertical than horizontal displacement.Source: §4 · p. 23
  • The findings are based on ten datasets from one 10 mm-GSD study area, so they may not generalize to smaller sensors, coarser GSDs, or substantially different site sizes.The GCPs were well-distributed and determined with high accuracy, so the authors advise caution when applying the findings to other acquisition setups.Source: §4 · p. 23

5. Conclusions

The conclusions recommend carefully selecting processing parameters, particularly GCP number, camera corrections, and PPK positioning, to support accurate UAV photogrammetry. Parameter optimization matters more for displacement accuracy than for tilt-change and horizontal-strain accuracy.

5. Conclusions

  • Careful parameter optimization can preserve high accuracy with fewer GCPs, while a high GCP density can act as a fail-safe when parameter choices are poor.At least five well-distributed, high-quality GCPs are recommended; for areas larger than 200 ha, the recommended density is 1 GCP per 40 ha.Source: §5 · p. 23
  • For similar measurement conditions, the authors recommend additional corrections, PPK tied to a local base station, post-BBA transformation, estimated camera-coordinate accuracy, and 2 px tie-point accuracy.They also recommend at least five GCPs at 1 per 40 ha for areas above 200 ha, GCP coordinate accuracy of 15–25 mm, and optional tie-point filtration.Source: §5 · pp. 23–24
  • Parameter selection affects displacement accuracy more than tilt-change and horizontal-strain accuracy, with systematic errors predominantly affecting vertical displacement.The conclusions attribute the difference to gradual spatial deformations and correlated errors among neighboring points.Source: §5 · p. 24

Declaration of Generative AI and AI-assisted technologies in the writing process; Declaration of competing interest; Data availability

Loading 2610.01438v1…