[Paper Review] Aeolian dune modelling from airborne LiDAR, terrestrial LiDAR and Structure from Motion-Multi View Stereo

Carlos Henrique Grohmann, Guilherme Pereira Bento Garcia+2 more|arXiv (Cornell University)|Oct 14, 2019

Remote Sensing and LiDAR Applications129 references38 citations

TL;DR

This study evaluates Structure from Motion–Multi View Stereo (SfM-MVS) using Remotely Piloted Aircraft (RPA) imagery for high-resolution 3D modeling of coastal dunes in southern Brazil. Validated against terrestrial LiDAR (TLS) and airborne LiDAR (ALS), SfM-MVS achieved 0.08 m RMSE and 0.06 m MAE, with a dune migration rate of ~5 m/year from 2010 to 2019 and only 0.2% volume change, demonstrating its cost-effectiveness and reliability for continuous dune monitoring.

ABSTRACT

A DEM of a dunefield in Southern Brazil was generated from 810 photos captured by an RPA in February 2019. Altimetric accuracy of the SfM-MVS DEM was validated by comparison with Terrestrial LiDAR (TLS) data collected during the same fieldwork campaign of the RPA flights. The SfM-MVS DEM was then compared to an Airborne LiDAR (ALS) DEM from October 2010. While the SfM-MVS and TLS DEMs are very similar, the SfM-MVS DEM presents a small scale surface roughness not visible in the TLS DEM. The Feature Preserving DEM Smoothing (FPD) algorithm was applied to the SfM-MVS DEM with good results in terms of surface smoothing, but without any significant changes in descriptive statistics and error metrics, with an RMSE of 0.08m and MAE of 0.06m for both the original and the FPD-filtered DEM. Displacement of dune crest lines from the ALS and SfM-MVS DEMs resulted in a migration rate of ~5m/year between 2010 and 2019, in good agreement with rates derived from satellite images and historical aerial photographs of the same area. Sand volume change in the same period showed a decrease of only 0.2%, which can be related to the installation of sand fences to promote dune stabilization and sand removal from the front of the dune field to keep a road open to vehicles. ALS can cover large areas in little time but its high cost still remains a barrier to wider usage, especially by researchers in developing countries. TLS has an intermediate cost but demands more fieldwork and more processing time. In our case we needed three days for the TLS survey and around three weeks to produce a DEM of ~80400m2. On the other hand, we were able to cover ~740900m2 with six flight missions in under three hours, with ~13 hours processing time in a medium-range workstation. This makes SfM-MVS a low-cost solution with fast and reliable results for 3D modelling and continuous monitoring of coastal dunes.

Motivation & Objective

• To assess the accuracy and feasibility of SfM-MVS using RPA-collected imagery for high-resolution 3D modeling of aeolian dunes.
• To validate SfM-MVS-derived Digital Elevation Models (DEMs) against terrestrial LiDAR (TLS) and airborne LiDAR (ALS) data.
• To quantify dune migration and volume change between 2010 and 2019 using multi-source DEMs.
• To evaluate the trade-offs between ALS, TLS, and SfM-MVS in terms of cost, time, and data quality for coastal dune monitoring.
• To demonstrate that SfM-MVS is a low-cost, fast, and reliable alternative to traditional LiDAR methods for geomorphometric analysis of dynamic dune systems.

Proposed method

• Acquired 810 RPA-mounted images at 100 m altitude in February 2019 to generate an SfM-MVS DEM of a dune field in southern Brazil.
• Processed RPA images using SfM-MVS algorithms to reconstruct a dense 3D point cloud and generate a high-resolution DEM.
• Collected TLS data during the same field campaign to serve as a high-accuracy reference for validating SfM-MVS DEM accuracy.
• Compared the SfM-MVS DEM with an ALS DEM from October 2010 to assess dune migration and volume change over time.
• Applied the Feature Preserving DEM Smoothing (FPD) algorithm to reduce surface roughness in the SfM-MVS DEM without altering key error metrics.
• Used statistical metrics (RMSE, MAE) and morphometric analysis (dune crest line displacement) to compare DEMs and quantify dune dynamics.

Experimental results

Research questions

• RQ1 How accurate is SfM-MVS-derived DEM from RPA imagery when validated against terrestrial LiDAR in a coastal dune environment?
• RQ2 What is the dune migration rate between 2010 and 2019 as derived from SfM-MVS and ALS DEMs, and how does it compare to rates from satellite and aerial imagery?
• RQ3 To what extent does the SfM-MVS DEM exhibit surface roughness not present in TLS-derived DEMs, and can this be effectively reduced without compromising accuracy?
• RQ4 How do the processing time, fieldwork duration, and cost of SfM-MVS compare to those of ALS and TLS for dune monitoring applications?
• RQ5 What factors enable successful SfM-MVS reconstruction in homogeneous, low-contrast dune environments with minimal natural texture?

Key findings

• The SfM-MVS DEM achieved an RMSE of 0.08 m and MAE of 0.06 m when compared to the TLS-derived DEM, indicating high altimetric accuracy.
• Despite high similarity between SfM-MVS and TLS DEMs, the SfM-MVS DEM exhibited small-scale surface roughness not present in the TLS data, potentially affecting fine-scale geomorphometric analysis.
• Application of the FPD de-noising algorithm effectively reduced surface roughness without altering the RMSE or MAE, preserving the DEM's accuracy.
• Dune crest line displacement between the ALS (2010) and SfM-MVS (2019) DEMs indicated a migration rate of approximately 5 m/year, consistent with rates derived from satellite and historical aerial imagery.
• Volume change between 2010 and 2019 was only 0.2%, suggesting minimal net sediment loss, possibly due to sand fences and active sand removal to maintain a road.
• SfM-MVS enabled coverage of approximately 740,900 m² in under three hours using six RPA missions, with 13 hours of processing time on a medium-range workstation, significantly outperforming TLS in speed and scalability.