Evaluation of SfM against tradional stereophotogrammetry and LiDAR techniques for DSM creation in various land cover areas
Date
2016-12
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
AFRIKAANS OPSOMMING:Struktuur-deur-Beweging (SdB) is moontlik ʼn effektiewe hulpmiddel om digitale oppervlak
modelle (DOMs) en orto-mosaïek beelde te skep. Die toename in kamera-resolusie asook die
ontwikkeling van bekostigbare en kommersiële onbemande vliegtuie (OVs) tegnologie het 'n basis
geskep vir die effektiewe gebruik van SdB om ʼn DOM te skep. Onlangse studies wat verband hou
met SdB het ʼn vertikale akkuraatheid van minder as ʼn meter in spesifieke studieomgewings
voorspel. Daar is ʼn behoefte aan oppervlak modelle wat vir topologiese- en ingenieursopnames
gebruik kan word, maar toegang tot hoë resolusie digitale oppervlak modelle (DOMs) is beperk
vir gebiede van Suid-Afrika en slegs lae resolusie DOMs is beskikbaar vir die res van die land.
SdB kan potensieel die gaping oorbrug indien konsekwente akkuraathede verkry kan word. Om
die effektiwiteit van SdB vir Geografiese Inligting Stelsels (GIS) beter te verstaan het hierdie
studie die ruimtelike akkuraathede van SdB met gevestigde, tegnieke (tradisionele stereofotogrammetrie
en Ligopsporing en afstandskatting of LiDAR) vir DOM-skepping in wisselende
grondgebruik- en grondbedekkingsomgewings. Hierdie assessering is aangepak deur elke tegniek
met onafhanklike globale posisioneringstelsel (GPS) verwysingspunte, asook met mekaar te
vergelyk.
Daar is bevind dal LiDAR die beste metode vir oppervlak herkonstruksie is en die meeste detail
vasgevang het asook algeheel die laagste relatiewe fouttelling oor verskillende landgebruike gehad
(wortel gemiddelde kwadratiese fout, WGKF, 0.57m). SdB het ʼn algehele WGKF van 0.67m
behaal, terwyl stereo-fotogrammetrie die swakste WGKF van 1.22m gelewer het. SdB het beperkte
vermoë getoon om in begroeide omgewings oppervlaktes te herkonstruktureer en slegs 58.2% is
binne 1m van die LiDAR oppervlak. SdB het egter rye in wingerde geïdentifiseer en ʼn lae
gemiddelde fout getoon toe dit met ‘n LiDAR verwysingsoppervlak vergelyk is. Stereofotogrammetrie
nie die wingerdrye geïdentifiseer nie, het begroeide gebiede te veel veralgemeen
en 'n laer vlak van insluiting getoon. SdB het beter resultate gelewer in beboude gebiede met 73.4-
81.4% wat binne 1 m van die LiDAR oppervlak val en het ʼn WGKF van 028-0.62m van GPS
verwysingspunte af gekry. Die inter-DOM ooreenkoms van SdB in beboude omgewings wys
daarop dat SdB moontlik beter kan wees vir data smelting as stereo-fotogrammetrie. Die
herkonstruksie van beboude gebiede was voldoende, maar is deur baie dig beboude gebiede wat
verdere ongewenste veralgemening van oppervlakmodelle veroorsaak.
LiDAR is steeds die beste tegniek vir DOM-skepping, maar beperkings ontstaan as gevolg van die
beskikbaarheid, koste, gespesialiseerde toerusting en kundigheid wat benodig word vir die
implementering van hierdie tegniek. Vir beeld-gebaseerde tegnieke lewer SDB beter resultate as
stereo-fototogrammetie maar dit beteken nie noodwendig dat stereo-fotogrammetie ʼn oorbodige
of verouderde tegniek is nie. SdB moet eerder beskou word as die natuurlike vertakking van
fotogrammetrie wat in kleiner, hoogs gedetailleerde omgewings wat moontlik hoë temporale
veranderlikheid toon gebruik kan word
ENGLISH ABSTRACT: Structure from Motion (SfM) is potentially an effective tool for the creation of digital surface models (DSMs) and ortho-mosaiced imagery. The increase in camera resolution and the emergence of affordable and commercial unmanned aerial vehicle (UAV) technology has provided a basis from which SfM can be deployed effectively for DSM creation. Recent studies with SfM have predicted a vertical accuracy of less than a meter in specific study environments. A need exists for surface models for topographical and engineering surveying but there is limited access to high resolution digital elevation models (DEMs) for parts of South Africa only and low resolution DEMs for the remainder. Potentially SfM could help to bridge this gap, if consistent accuracy can be proven. To better understand the effectiveness of SfM for Geographic Information Systems (GIS) application, this study compared spatial accuracies of SfM to well understood commercial techniques (traditional stereophotogrammetry and Light Detection and Ranging or LiDAR) in DSM creation in varying land use and land cover environments. This assessment was made by comparing each technique to an independent Global Positioning System (GPS) reference as well as relative to each other. It was found that LiDAR was the best method for surface reconstruction picking up the highest level of detail and overall the lowest margin of error over different land uses (Root Mean Square Error, RMSE, 0.57m). SfM in the same assessment found an overall RMSE of 0.67m with stereophotogrammetry performing least well with and RMSE of 1.22m. SfM showed limited reconstruction ability in vegetated environments with only 58.2% included within 1m from the LiDAR surface. SfM did, however, pick up formal crop rows and demonstrated a low mean error from a LiDAR surface reference in the vineyard land cover class. Stereophotogrammetry failed to pick up vine rows and displayed higher levels of overgeneralization in vegetated areas with a lower level of inclusion. SfM performed better in built environments with 73.4-81.4% inclusion within 1m from the LiDAR surface and a RMSE range of 028-0.62m from a GPS reference. The interDSM agreement of SFM in built environments suggested better propensity for data fusion than stereophotogrammetry. The ability to reconstruct built features was adequate but limited by highly clustered buildings resulting in overgeneralized surfaces. LiDAR is still the best technique for DSM construction but experiences limitations related to availability, cost, specialist equipment and knowledge. For image-based techniques, SfM out performed stereophotogrammetry but this does not necessarily imply stereophotoggrammetry is redundant as a technique. Instead SfM should be considered the natural extension of photogrammetry for use in smaller, highly detailed and potentially highly temporal environments
ENGLISH ABSTRACT: Structure from Motion (SfM) is potentially an effective tool for the creation of digital surface models (DSMs) and ortho-mosaiced imagery. The increase in camera resolution and the emergence of affordable and commercial unmanned aerial vehicle (UAV) technology has provided a basis from which SfM can be deployed effectively for DSM creation. Recent studies with SfM have predicted a vertical accuracy of less than a meter in specific study environments. A need exists for surface models for topographical and engineering surveying but there is limited access to high resolution digital elevation models (DEMs) for parts of South Africa only and low resolution DEMs for the remainder. Potentially SfM could help to bridge this gap, if consistent accuracy can be proven. To better understand the effectiveness of SfM for Geographic Information Systems (GIS) application, this study compared spatial accuracies of SfM to well understood commercial techniques (traditional stereophotogrammetry and Light Detection and Ranging or LiDAR) in DSM creation in varying land use and land cover environments. This assessment was made by comparing each technique to an independent Global Positioning System (GPS) reference as well as relative to each other. It was found that LiDAR was the best method for surface reconstruction picking up the highest level of detail and overall the lowest margin of error over different land uses (Root Mean Square Error, RMSE, 0.57m). SfM in the same assessment found an overall RMSE of 0.67m with stereophotogrammetry performing least well with and RMSE of 1.22m. SfM showed limited reconstruction ability in vegetated environments with only 58.2% included within 1m from the LiDAR surface. SfM did, however, pick up formal crop rows and demonstrated a low mean error from a LiDAR surface reference in the vineyard land cover class. Stereophotogrammetry failed to pick up vine rows and displayed higher levels of overgeneralization in vegetated areas with a lower level of inclusion. SfM performed better in built environments with 73.4-81.4% inclusion within 1m from the LiDAR surface and a RMSE range of 028-0.62m from a GPS reference. The interDSM agreement of SFM in built environments suggested better propensity for data fusion than stereophotogrammetry. The ability to reconstruct built features was adequate but limited by highly clustered buildings resulting in overgeneralized surfaces. LiDAR is still the best technique for DSM construction but experiences limitations related to availability, cost, specialist equipment and knowledge. For image-based techniques, SfM out performed stereophotogrammetry but this does not necessarily imply stereophotoggrammetry is redundant as a technique. Instead SfM should be considered the natural extension of photogrammetry for use in smaller, highly detailed and potentially highly temporal environments
Description
Thesis (MSc)--Stellenbosch University, 2016.
Keywords
LiDAR, Traditional Stereophotogrammetry, Geographic Information Systems, SfM techniques, Structure from motion techniques, Geographic Information Systems, Light Detection and Ranging, UCTD, Digital Surface Models, DCMs, GIS