pith. sign in

arxiv: 2606.20681 · v1 · pith:W6COYSLJnew · submitted 2026-06-13 · 💻 cs.CV · cs.LG

A UAV-Based Multi-Modal Vision System for Automated Sideslope Deformation Monitoring and Hazard Detection

Jingfeng Zhang , Yi Li , Xianchong Liang , Huan Yang This is my paper

Pith reviewed 2026-06-27 03:54 UTC · model grok-4.3

classification 💻 cs.CV cs.LG
keywords UAV LiDARslope deformation monitoringhazard detectionpoint cloud processingRandLA-Netexpressway safetygrid differencingground extraction
0
0 comments X

The pith

UAV LiDAR workflow extracts ground point clouds under vegetation, screens hazards from single scans, and measures centimeter-level slope deformations over time.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops an automated UAV-borne LiDAR workflow for monitoring expressway sideslopes. A shared acquisition and ground-extraction stage feeds two branches: one uses RandLA-Net on single-observation point clouds to flag potential hazard zones, while the other applies grid-wise elevation differencing across multiple epochs to quantify slow surface changes. The approach targets the inefficiency and safety risks of manual inspections on deteriorated slopes. Validation on real expressway sites shows usable ground clouds can be recovered despite vegetation cover and that changes at the centimeter scale become measurable.

Core claim

The workflow consists of a shared data-acquisition and ground-surface extraction stage, a single-observation hazard-screening branch based on RandLA-Net, and a multi-epoch deformation-monitoring branch based on grid-wise elevation differencing. In field tests the system extracts usable ground-surface point clouds under vegetation cover, identifies potential hazard zones from single-observation point clouds, and quantifies centimeter-level elevation changes using multi-epoch grid differencing, thereby establishing an end-to-end UAV-borne LiDAR solution for slope inspection.

What carries the argument

RandLA-Net-based single-observation hazard screening combined with grid-wise elevation differencing on multi-epoch ground point clouds.

If this is right

  • Ground-surface point clouds become extractable even when vegetation is present.
  • Potential hazard zones can be identified from data collected in a single flight.
  • Centimeter-level elevation changes can be quantified by comparing grid elevations across separate observation epochs.
  • The combined pipeline supplies an implementable automated alternative to manual slope inspection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same acquisition flights could support both immediate hazard flagging and long-term trend tracking without additional hardware.
  • Integration with scheduled maintenance windows might allow continuous monitoring of high-risk sections without closing lanes.
  • Similar point-cloud differencing could be tested on other linear infrastructure such as railway embankments or pipeline corridors.

Load-bearing premise

The RandLA-Net model separates ground from vegetation and flags true hazards on varied real-world expressway slopes without excessive false positives or missed events.

What would settle it

A controlled slope with documented vegetation cover and a known centimeter-scale deformation that the workflow either misses or misclassifies as a hazard zone.

Figures

Figures reproduced from arXiv: 2606.20681 by Huan Yang, Jingfeng Zhang, Xianchong Liang, Yi Li.

Figure 1
Figure 1. Figure 1: Overall workflow of the proposed UAV-borne LiDAR-based slope inspection system. 2. Related Work 2.1. Common Slope Hazards and Their Monitoring Requirements Highway slopes are exposed to long-term rainfall erosion, geological weathering, and engineering disturbance, making them susceptible to various hazards. Existing studies generally classify slope hazards into landslides, collapses or rockfalls, erosion-… view at source ↗
Figure 2
Figure 2. Figure 2: Literature organization of UAV-based slope inspection methods. 2.3. Advances in Deep Learning for Three-Dimensional Point Clouds Deep learning has provided new opportunities for automated geological-hazard recognition. Previous studies have applied deep learning to landslides, debris flows, collapses, and related hazards using remote-sensing images, UAV data, and other multisource observations (Ma and Mei,… view at source ↗
Figure 3
Figure 3. Figure 3: Literature organization of deep learning methods for three-dimensional point clouds. 3. Problem Statement 3.1. Engineering Task and Two Observation Scenarios Expressway slope hazards usually evolve through progressive surface deformation rather than instantaneous failure, making timely geometric observation important for early warning and maintenance prioritization. Therefore, this study formulates the cen… view at source ↗
Figure 4
Figure 4. Figure 4: illustrates these two observation scenarios, including obvious morphological hazards that can be judged from a single observation and slow surface changes that require multi-epoch observations for quantitative comparison. Obvious morphological hazards can be identified from a single observation One LiDAR survey Detection of obvious hazard areas Slow and subtle surface changes require multi-epoch quantitati… view at source ↗
Figure 5
Figure 5. Figure 5: Overview of the proposed method, consisting of shared UAV LiDAR data acquisition and preprocessing, single￾epoch semantic anomaly recognition, and multi-epoch differential deformation detection. 4.2. UAV-borne LiDAR Acquisition and Ground-Surface Extraction The shared preprocessing stage is designed to convert raw UAV-borne LiDAR observations into slope ground￾surface point clouds for subsequent analysis. … view at source ↗
Figure 6
Figure 6. Figure 6 [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Detailed pipeline of the multi-epoch deformation monitoring branch, including point-cloud registration, grid construction, common-grid intersection, elevation differencing, and heatmap generation. Page 16 of 29 [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visual comparison of the under-canopy foam-board target before and after CSF filtering. The histogram in [PITH_FULL_IMAGE:figures/full_fig_p018_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Elevation histograms of the foam-board ROI before and after CSF filtering, illustrating the reduction of high￾elevation non-ground returns and the concentration of filtered ground points in the low-elevation range. 5.2. Single-Observation Hazard Screening The single-observation branch was evaluated using simulated slope-hazard point clouds with three semantic classes: normal background, uplift hazard, and … view at source ↗
Figure 10
Figure 10. Figure 10: Qualitative result of the single-observation hazard screening branch, comparing the RGB point-cloud scene, ground-truth labels, and predicted labels for normal background, bulging hazard, and depression hazard. Page 20 of 29 [PITH_FULL_IMAGE:figures/full_fig_p020_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Aerial overview of the controlled known-deformation experiment site, with two manually constructed artificial uplift regions marked for multi-epoch elevation-differencing validation [PITH_FULL_IMAGE:figures/full_fig_p022_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Elevation-difference results of the controlled known-deformation experiment, showing the spatial responses of the two artificial uplift regions in both heatmap and overlay views. The real expressway slope experiment used two UAV-borne LiDAR observations acquired on 15 Dec 2025 and 25 Mar 2026 [PITH_FULL_IMAGE:figures/full_fig_p023_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Aerial overview of the real expressway slope field site. The basic raw point-cloud statistics of the two epochs are listed in [PITH_FULL_IMAGE:figures/full_fig_p024_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Elevation-difference overlay and time-separated UAV images of the real expressway slope case, showing spatial correspondence between negative elevation-change regions and visible erosion or material loss. 5.4. Summary of Experimental Evidence The experiments provide three levels of evidence. First, the under-canopy foam-board experiment confirms that the shared preprocessing stage can extract usable groun… view at source ↗
read the original abstract

Slope hazards constitute a major safety threat to expressway infrastructure, and their evolution is typically manifested as slow surface deformation. Conventional manual inspection suffers from low efficiency and inadequate operational safety, especially on severely deteriorated slopes. Accordingly, there is an urgent need for an automated, high-precision solution capable of large-area slope observation and analysis. This study aims to develop a highly automated workflow for slope hazard detection using Unmanned Aerial Vehicle (UAV)-borne Light Detection and Ranging (LiDAR). The proposed workflow consists of a shared data-acquisition and ground-surface extraction stage, a single-observation hazard-screening branch based on RandLA-Net, and a multi-epoch deformation-monitoring branch based on grid-wise elevation differencing. To validate the effectiveness of the proposed system, we conducted multiple UAV-borne LiDAR data-acquisition flights in real expressway slope environments. The results show that the workflow can extract usable ground-surface point clouds under vegetation cover, identify potential hazard zones from single-observation point clouds, and quantify centimeter-level elevation changes using multi-epoch grid differencing. This study establishes an end-to-end UAV-borne LiDAR-based workflow for slope inspection and demonstrates its feasibility through controlled experiments, field tests, and simulation-based validation, thereby providing an implementable solution for automated slope-hazard monitoring and intelligent early warning.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The paper proposes an end-to-end UAV-LiDAR workflow for automated expressway sideslope hazard detection and deformation monitoring. It comprises a shared acquisition and ground-surface extraction stage, a single-observation branch that applies RandLA-Net to screen for hazards from individual point clouds, and a multi-epoch branch that performs grid-wise elevation differencing to quantify surface changes. The authors report field tests on real expressway slopes and claim that the system extracts usable ground points under vegetation, identifies potential hazard zones, and measures centimeter-level elevation changes.

Significance. A quantitatively validated version of this workflow would address a practical infrastructure-safety need by replacing labor-intensive manual inspections with repeatable, large-area UAV observations. The combination of semantic segmentation for single-pass hazard flagging and direct grid differencing for change detection is a straightforward engineering contribution whose value hinges on demonstrated accuracy under realistic vegetation and slope conditions.

major comments (3)
  1. [Results / Validation] Results section (and abstract): the central claim that RandLA-Net reliably separates ground from vegetation and flags true hazards rests on unquantified field-test outcomes. No precision, recall, IoU, or false-positive rates are reported for the expressway-slope UAV-LiDAR datasets, nor is a test-set construction or training/adaptation procedure described. This absence directly undermines the generalization assertion for the single-observation branch.
  2. [Methods] Methods (RandLA-Net branch): without baseline comparisons (e.g., CSF, PMF, or other ground-filtering algorithms) or an ablation on vegetation density, it is impossible to judge whether the reported qualitative success is attributable to RandLA-Net or to the preceding ground-extraction stage.
  3. [Methods / Results] Multi-epoch branch: the claim of “centimeter-level” elevation change detection via grid differencing lacks an error-propagation analysis or registration-accuracy assessment between epochs; the reported precision cannot be evaluated without these numbers.
minor comments (2)
  1. [Abstract] The abstract states “simulation-based validation” yet the manuscript provides no description of the simulation setup or how it complements the field data.
  2. [Methods] Notation for grid size, epoch alignment, and vegetation-removal thresholds should be defined explicitly in the methods rather than left to figure captions.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which identify key areas where additional rigor will strengthen the manuscript. We address each major comment below and commit to revisions that provide the requested quantitative support and analyses.

read point-by-point responses

  1. Referee: [Results / Validation] Results section (and abstract): the central claim that RandLA-Net reliably separates ground from vegetation and flags true hazards rests on unquantified field-test outcomes. No precision, recall, IoU, or false-positive rates are reported for the expressway-slope UAV-LiDAR datasets, nor is a test-set construction or training/adaptation procedure described. This absence directly undermines the generalization assertion for the single-observation branch.

    Authors: We agree that quantitative metrics are required to substantiate the claims. The present manuscript reports only qualitative outcomes from field tests. In revision we will add a dedicated quantitative evaluation subsection that reports precision, recall, IoU and false-positive rates on held-out expressway-slope test sets, together with explicit descriptions of test-set construction and the RandLA-Net training/adaptation procedure. revision: yes

  2. Referee: [Methods] Methods (RandLA-Net branch): without baseline comparisons (e.g., CSF, PMF, or other ground-filtering algorithms) or an ablation on vegetation density, it is impossible to judge whether the reported qualitative success is attributable to RandLA-Net or to the preceding ground-extraction stage.

    Authors: The observation is correct; the manuscript does not contain such comparisons. We will incorporate baseline experiments against CSF and PMF as well as an ablation study across vegetation-density strata to isolate the contribution of the RandLA-Net stage. revision: yes

  3. Referee: [Methods / Results] Multi-epoch branch: the claim of “centimeter-level” elevation change detection via grid differencing lacks an error-propagation analysis or registration-accuracy assessment between epochs; the reported precision cannot be evaluated without these numbers.

    Authors: We acknowledge the absence of a formal uncertainty analysis. The centimeter-level figures derive from observed field differences, but registration accuracy and error propagation are not quantified. The revision will add inter-epoch registration-error statistics and an error-propagation treatment for the grid-differencing step. revision: yes

Circularity Check

0 steps flagged

No significant circularity; standard methods applied without self-referential reduction

full rationale

The workflow relies on established components (RandLA-Net segmentation and grid-wise differencing) applied to new UAV-LiDAR slope data. No equations, fitted parameters, or predictions are shown to reduce by construction to the inputs. Claims of ground extraction and hazard identification rest on experimental validation rather than definitional equivalence or self-citation chains. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the unverified assumption that RandLA-Net generalizes to vegetated slope data and that grid differencing accurately isolates deformation from noise and registration errors; no free parameters, invented entities, or additional axioms are described in the abstract.

axioms (2)
  • domain assumption RandLA-Net segmentation produces reliable ground-surface extraction and hazard labels on real expressway slopes under vegetation
    Invoked as the basis for the single-observation hazard-screening branch.
  • domain assumption Multi-epoch grid-wise elevation differencing can isolate true centimeter-scale deformation from registration, vegetation, and sensor noise
    Required for the deformation-monitoring branch to deliver usable change measurements.

pith-pipeline@v0.9.1-grok · 5774 in / 1330 out tokens · 22799 ms · 2026-06-27T03:54:12.064361+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

19 extracted references · 18 canonical work pages

  1. [1]

    Bulletin of Engineering Geology and the Environment doi:10.1007/s10064-021-02306-2

    Identification of potential rockfall sources using uav-derived point cloud. Bulletin of Engineering Geology and the Environment doi:10.1007/s10064-021-02306-2. ASPRS,

    work page doi:10.1007/s10064-021-02306-2

  2. [2]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 14, 239–256

    A method for registration of 3-d shapes. IEEE Transactions on Pattern Analysis and Machine Intelligence 14, 239–256. doi:10.1109/34.121791. Chen, B., Maurer, J., Gong, W., 2025a. Applications of uav in landslide research: a review. Landslides 22, 3029–3048. doi:10.1007/ s10346-025-02547-2. Chen,J.,Xie,C.,Zhang,W.,Fu,C.,Shen,J.,Yang,B.,Li,H.,Shi,D.,2025b. ...

    work page doi:10.1109/34.121791

  3. [3]

    Choi, S.K., Ramirez, R.A., Kwon, T.H.,

    doi:10.3390/su17094176. Choi, S.K., Ramirez, R.A., Kwon, T.H.,

    work page doi:10.3390/su17094176

  4. [4]

    Heliyon 9, e20225

    Acquisition of high-resolution topographic information in forest environments using integrated uav-lidar system: System development and field demonstration. Heliyon 9, e20225. doi:10.1016/j.heliyon.2023.e20225. Deng, K., Yang, B., Yu, Z., Pu, Q., He, L.,

    work page doi:10.1016/j.heliyon.2023.e20225 2023

  5. [5]

    Frontiers in Earth Science 13, 1715960

    Automatic extraction algorithm for landslide cracks using insar-uav lidar point cloud coupling. Frontiers in Earth Science 13, 1715960. doi:10.3389/feart.2025.1715960. DJI,

    work page doi:10.3389/feart.2025.1715960 2025

  6. [6]

    Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartog- raphy

    Random sample consensus: A paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM 24, 381–395. doi:10.1145/358669.358692. Froude,M.J.,Petley,D.N.,2018. Globalfatallandslideoccurrencefrom2004to2016. NaturalHazardsandEarthSystemSciences18,2161–2181. doi:10.5194/nhess-18-2161-2018. Galve, J.P., Per...

    work page doi:10.1145/358669.358692 2018

  7. [7]

    doi:10.1007/s10346-024-02449-9

    Applications of uav digital photogrammetryinlandslideemergencyresponseandrecoveryactivities:Thecasestudyofaslopefailureinthea-7highway(sspain).Landslides 22, 1383–1396. doi:10.1007/s10346-024-02449-9. Hu, Q., Yang, B., Xie, L., Rosa, S., Guo, Y., Wang, Z., Trigoni, N., Markham, A.,

    work page doi:10.1007/s10346-024-02449-9

  8. [8]

    11108–11117

    Randla-net: Efficient semantic segmentation of large- scale point clouds, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11108–11117. URL: https://arxiv.org/abs/1911.11236. Jiang, P., Ma, Z., Mei, G.,

    arXiv 1911

  9. [9]

    Natural Hazards and Earth System Sciences 26, 487–529

    Review article: Deep learning for potential landslide identification: data, models, applications, challenges, and opportunities. Natural Hazards and Earth System Sciences 26, 487–529. doi:10.5194/nhess-26-487-2026. Lague, D., Brodu, N., Leroux, J.,

    work page doi:10.5194/nhess-26-487-2026 2026

  10. [10]

    ISPRS Journal of Photogrammetry and Remote Sensing 82, 10–26

    Accurate 3d comparison of complex topography with terrestrial laser scanner: Application to the rangitikei canyon (n-z). ISPRS Journal of Photogrammetry and Remote Sensing 82, 10–26. doi:10.1016/j.isprsjprs.2013.04.009. Liu, K., Wang, Q., Wang, M., Koks, E.E.,

    work page doi:10.1016/j.isprsjprs.2013.04.009 2013

  11. [11]

    Ma, Z., Mei, G.,

    doi:10.1038/s41467-023-38203-3. Ma, Z., Mei, G.,

    work page doi:10.1038/s41467-023-38203-3

  12. [12]

    Earth-Science Reviews 223, 103858

    Deep learning for geological hazards analysis: Data, models, applications, and opportunities. Earth-Science Reviews 223, 103858. doi:10.1016/j.earscirev.2021.103858. Qi, C.R., Su, H., Mo, K., Guibas, L.J., 2017a. Pointnet: Deep learning on point sets for 3d classification and segmentation, in: Proceedings of the IEEE Conference on Computer Vision and Patt...

    work page doi:10.1016/j.earscirev.2021.103858 2021

  13. [13]

    Engineering Geology 346, 107890

    A novel method for landslide deformation monitoring by fusing uav photogrammetry and lidar data based on each sensor’s mapping advantage in regards to terrain feature. Engineering Geology 346, 107890. doi:10.1016/j.enggeo.2024.107890. Sun, J., Yuan, G., Song, L., Zhang, H.,

    work page doi:10.1016/j.enggeo.2024.107890 2024

  14. [14]

    Wheaton, J.M., Brasington, J., Darby, S.E., Sear, D.A.,

    doi:10.3390/drones8010030. Wheaton, J.M., Brasington, J., Darby, S.E., Sear, D.A.,

    work page doi:10.3390/drones8010030

  15. [15]

    Earth Surface Processes and Landforms 35, 136–156

    Accounting for uncertainty in dems from repeat topographic surveys: improved sediment budgets. Earth Surface Processes and Landforms 35, 136–156. doi:10.1002/esp.1886. Wu,G.,etal.,2025. Failuremechanismsandmitigationstrategiesofhighwaysubgradestructuresunderrainwatererosionconditions. CaseStudies in Construction Materials , e05045doi:10.1016/j.cscm.2025.e...

    work page doi:10.1002/esp.1886 2025

  16. [16]

    Zai, D., Liu, J.,

    doi:10.1007/978-981-97-6238-5_43. Zai, D., Liu, J.,

    work page doi:10.1007/978-981-97-6238-5_43

  17. [17]

    Geotechnical and Geological Engineering doi:10.1007/s10706-025-03228-6

    3d slope reconstruction and point cloud analysis using uav multi-source data. Geotechnical and Geological Engineering doi:10.1007/s10706-025-03228-6. Zhang, J., Zhang, S., Ding, Y., et al.,

    work page doi:10.1007/s10706-025-03228-6

  18. [18]

    Zhang,W.,Qi,J.,Wan,P.,Wang,H.,Xie,D.,Wang,X.,Yan,G.,2016

    doi:10.1038/s41598-023-32055-z. Zhang,W.,Qi,J.,Wan,P.,Wang,H.,Xie,D.,Wang,X.,Yan,G.,2016. Aneasy-to-useairbornelidardatafilteringmethodbasedonclothsimulation. Remote Sensing 8,

    work page doi:10.1038/s41598-023-32055-z 2016

  19. [19]

    Page 29 of 29

    doi:10.3390/rs8060501. Page 29 of 29

    work page doi:10.3390/rs8060501