Quantifying Wind Turbine Blade Surface Roughness using Sandpaper

Grit Sizes: An Initial Exploration

Ivan Nikolov and Claus Madsen

Department of Architecture, Design and Media Technology, Aalborg University, Rendsburggade 14, Aalborg, Denmark

Keywords:

3D Reconstruction, Surface Inspection, Sandpaper Roughness, Classiﬁcation, Random Forests, Geometrical

Properties.

Abstract:

Surface inspection of wind turbine blades is a necessary step, to ensure longevity and sustained high energy

output. The detection of accumulation of damages and increased surface roughness of in-use blades, is one

of the main objectives of inspections in the wind energy industry. Creating 3D scans of the leading edges of

blade surfaces has been more and more used for capturing the roughness proﬁle of blades. An important part

in analysing these surface 3D scans is the standardization of the captured data across different blade surfaces,

types and sizes. In this paper we propose an initial exploration of using sandpaper grit sizes to provide this

standardization. Sandpaper has been widely used for approximating different levels of blade surface roughness

and its standardized nature can be used to easily describe and compare blade surfaces. We reconstruct a number

of different sandpaper grit sizes - from coarser P40 to a ﬁner P180. We extract a number of 3D surface features

from them and use them to train a random forest classiﬁcation method. This method is then used to segment

the surfaces of wind turbine blades in areas of different surface roughness. We test our proposed solution on

a variety of blade surfaces - from smooth to course and damaged and show that it manages to classify them

depending on their roughness.

1 INTRODUCTION

Surface inspection is a required part of ensuring

the proper working condition of machinery and in-

frastructure in industries like agriculture (El-Mesery

et al., 2019), manufacturing (Dastoorian et al., 2018)

and energy production (Zhang et al., 2017b) among

other. The wind energy production industry is partic-

ularly susceptible to the effects of infrastructure dam-

ages and degradation. For achieving maximum wind

turbine performance, blades need to be inspected reg-

ularly and potential damages caused by weather ero-

sion and imperfections in the manufacturing process

(Martin et al., 2018), (Du et al., 2020), need to be de-

tected as soon as possible. It has been shown that the

presence of even small surface roughness deviations

and damages can cause 2% and 5% loss in energy pro-

duction (Langel et al., 2015), with numbers as high

as 25% for larger damages and surface imperfections

(Schramm et al., 2017).

Wind turbine blade inspection is normally focused

on the leading edge of the blades, as it is shown

that damages there affect performance the most (Bak

et al., 2016). Inspection can be done manually by ex-

perts on site or in a laboratory setting through contact

measurements and microscopy analysis (Chen, 2018)

or by using machine vision methods (Shihavuddin

et al., 2019). This normally requires capturing images

of the surface, which are used for detecting potential

damages and imperfection (Shihavuddin et al., 2019),

(Moreno et al., 2018), (Al-Kaff et al., 2017). These

algorithms require a lot of image data of both clean

and damaged surfaces of wind turbine blades, which

is not easily accessible. These systems also cannot

normally quantify the micro surface roughness of the

blade, as they lack depth information. This depth in-

formation can be captured by reconstructing the 3D

surface, using techniques like Structure from Motion

(Zhang et al., 2020), (Nielsen et al., 2020).

To quantify this information and be able to com-

pare it between blade surfaces, a standardized model

of measurement of surface roughness is required. Pro-

ﬁle roughness metrics part of the ISO 4287 (ISO4287,

1997) standard are used for estimation of the details

of line scans of surfaces. These proﬁle metrics can

be extended to a 3D area, by using the ISO 25178

(ISO25178, 2012) standard. For 3D, a plane is ﬁt-

ted to the surface patches and the metrics are cal-

Nikolov, I. and Madsen, C.

Quantifying Wind Turbine Blade Surface Roughness using Sandpaper Grit Sizes: An Initial Exploration.

DOI: 10.5220/0010283908010808

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 5: VISAPP, pages

801-808

ISBN: 978-989-758-488-6

801

culated from their values. In this paper we explore

extending these metrics representations, by classify-

ing the surface based on sandpaper roughness that

best describes it. Sandpapers with different grit sizes

have been widely used in the literature for model-

ing wind turbine blade roughness in wind tunnels

(Marzuki et al., 2018), (Genc et al., 2019). It has been

shown that surfaces with attached sandpaper patches

exhibit the same behaviour as damaged ones. By

classifying wind turbine blades, into different sandpa-

per roughness proﬁles, their degradation can be more

easily communicated and compared between blades.

In addition, sandpaper surface grits are standardized

(ISO6344, 1998) and easily accessible, making repro-

duction of results easy and straightforward. Finally,

3D data of sandpaper surfaces can be much more eas-

ily obtained, than real wind turbine blade surface data

with different levels of surface damages.

2 STATE OF THE ART

Wind turbine blade analysis can be used to calculate

the predicted energy production (Han et al., 2018)

and the utilization coefﬁcients of turbines (Wang and

Zhang, 2017). It can be also used for introducing a

more granular control on the ﬂow control and aerody-

namic properties of blades (Langel et al., 2017).

Capturing of 3D data from the surface of wind tur-

bine blades is a widely researched topic. Two main

approaches to capturing 3D can be seen in the state

of the art - for on-site inspections (Xu et al., 2019),

when the wind turbine has been just stopped and off-

line laboratory inspections, where decommissioned

blades are normally inspected (Chen, 2018). The ﬁrst

type of inspection is performed on a more regular ba-

sis and aims to keep the blades in as close as possible

optimal conditions, while the second type is focus on

understanding why and how a blade failed or it was

decided that the damages are too severe. For on-site

inspections cameras and 3D sensors like LiDARs and

stereo cameras are mounted on unmanned aerial ve-

hicles (Zhang et al., 2017a), (Peng and Liu, 2018).

These aim to produce less high detailed reconstruc-

tions, which could under optimal capturing conditions

capture enough information to give an estimate of

the current condition of the blades. Laboratory in-

spections normally use more detailed surface analy-

sis technique, employing electronic microscopes or

surface probes (Chan et al., 2019), (Amenabar et al.,

2011). These methods can capture very high reso-

lution sub-millimeter accuracy reconstructions of the

blade surfaces.

In this paper we try to classify wind turbine blade

surfaces using sandpapers with different grit sizes. As

sandpapers have known surface properties, we can

then infer the same information about the blade sur-

faces. In addition, sandpaper data is widely used for

simulating how different surface roughness affects the

aerodynamic characteristics of wind turbine blades

and their energy production potential. We base our

research on traditional supervised learning methods,

using hand crafted features like the ones presented in

the work of (Weinmann et al., 2015), (Hackel et al.,

2016), (Dittrich et al., 2017), for use on large scale Li-

DAR point cloud segmentation, focusing on extract-

ing geometrical features, which describe the under-

lying surface in a robust way. Our proposed solu-

tion uses 3D sandpaper data to train a Random Forest

classiﬁer, which is later used to segment wind turbine

blade surfaces into ares best described by the differ-

ent sandpaper grit sizes This information can be later

used to model test setups or to better evaluate how

the roughness would affect the blade, based on sand-

paper wind tunnel tests. We show that the proposed

approach gives promising results.

3 METHODOLOGY

3.1 SfM Overview

As both the sandpaper patches and the testing wind

turbine data are captured using Structure from Motion

(SfM), we will ﬁrst give an overview of the method,

for easier reproducibility. SfM is a part of the multi-

view stereo algorithms, that uses only 2D image data

from different positions and view directions to recon-

struct the full 3D surface of an object (Ozyesil et al.,

2017). A number of images with a certain overlap

are used as input to the algorithm. Features are ex-

tracted from each image and matched between im-

ages. This can be done using algorithms like SIFT

(Lowe, 2004) or ORB (Rublee et al., 2011). These

matched features, together with information about the

intrinsic parameters of the camera used to take the im-

ages, are used to iteratively triangulate the camera po-

sitions and create a sparse point cloud of the scanned

object. The sparse point cloud is then reﬁned using

bundle adjustment (Triggs et al., 1999) and a dense

point cloud is created from the 3D sparse points and

camera positions.

3.2 Sandpaper 3D Reconstruction

A number of sandpaper grit sizes are chosen for devel-

oping the training examples and given in Table 1, to-

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

802

Table 1: Sandpaper grit size and the nominal average parti-

cle diameter in mm (FEPA, 1955).

Grit size Nom. av. diam. (mm)

P40 0.425

P60 0.269

P80 0.201

P100 0.162

P120 0.125

P180 0.082

gether with their average particle size. This standard

is proposed by (FEPA, 1955) and widely used. The

”P” notation of the grit sizes is inversely related to the

coarseness of the sandpaper material and represents

the size of the particles, embedded in the material.

These grit sizes are chosen, because they are widely

used for blade roughness approximation (Genc et al.,

2019) and the nominal average diameter of their grit

structure is representative of blade roughness values,

at which energy production loss starts to be observed

(Bak et al., 2016). To more closely mimic how rough-

ness on a wind turbine blade edge would behave,

the sandpaper patches are mounted on blade replicas

made from styrofoam. The replicas are modeled after

the NACA 633418 blade and can be seen in Figure

1(a).

For 3D reconstructing the blade replicas a Canon

5Ds camera is used. As proposed by (Nikolov and

Madsen, 2016), a semi-circular pattern with a 1.5m

radius is used and 18 images are taken in three dif-

ferent heights of the leading edge with the sandpaper

attached. The thus created 54 images are then used as

an input to Metashape (Agisoft, 2010), a commercial

SfM solution, shown to produce high detailed and ro-

bust reconstructions. An example of the reconstructed

point clouds for grit size P40, can be seen in Figure

1(b).

(a) (b)

Figure 1: Example of styrofoam blade replica with sand-

paper of P40 installed on it 1(a), together with the recon-

structed surface 1(b).

3.3 Sandpaper Feature Extraction

We extract a number of features from the sandpaper

point clouds, as presented in the work by (Blomley

et al., 2016). To calculate these features ﬁrst the lo-

cal neighbourhood around each point needs to be ex-

tracted, using a KDtree (Bentley, 1975), using the im-

plementation from (Zhou et al., 2018). In their work

(Blomley et al., 2016) and (Weinmann et al., 2015)

test features extracted from different types of neigh-

bourhood selection for each point. From the results of

their work it can be seen that a combination of mul-

tiple neighbourhoods with different shapes and sizes

gives the best results. We also choose this approach,

but focus on neighbourhoods of the same shape, as the

classes we are trying to detect, have uniform sandpa-

per roughness structures. We choose the neighbour-

hood scales, in a way that they can encompass the

average particle sizes of the selected sandpaper grits.

We choose spherical neighbourhoods with radii in the

interval [1 . . . 0.1] mm and a delta change of 0.1 mm.

This creates 10 progressively smaller scales of neigh-

bourhoods. If not enough points are present in the

neighbourhood for calculating the features, it is ze-

roed out.

Three types of features are selected, as described

in (Blomley et al., 2016) and (Weinmann et al., 2015),

(Dittrich et al., 2017) - fundamental geometrical prop-

erties of the point clouds, local shape covariance fea-

tures and local statistical shape distribution features.

These are calculated for each neighbourhood scale

and combined. As described by (Blomley et al.,

2016), the fundamental geometrical properties are as

follow:

• Local point density of the neighbourhood around

a given point (Figure 2(a)),

• The farthest distance between points in the neigh-

bourhood (Figure 2(b)),

• The maximum height of difference between

points in the neighbourhood, where height is ex-

pressed in the direction of the average normal of

the neighbourhood (Figure 2(c)),

• The standard deviation of the height differences

between points in the neighbourhood, where again

the height is expressed in the direction of the av-

erage normal (Figure 2(d)).

A visualization of these features for a neighbour-

hood size of 1mm on the P40 sandpaper are shown in

Figure 2, where just a small area of the whole sand-

paper leading edge is visualized.

The local shape features proposed by (Blomley

et al., 2016) and (Weinmann et al., 2015) are based

on the use of covariance features and require the cal-

culation of the eigenvalues λ

, λ

of each neigh-

bourhood scale. These features are as follows:

Quantifying Wind Turbine Blade Surface Roughness using Sandpaper Grit Sizes: An Initial Exploration

803

(a) Local Density (b) Farthest Distance

Figure 2: Example local fundamental geometrical proper-

ties, extracted from the P40 sandpaper grit size. Just a small

part of the sandpaper leading edge is shown for easier visu-

alization.

Linearity: L

− λ

(1)

Planarity: P

− λ

(2)

Sphericity: S

(3)

Omnivariance: O

∏

i=1

(4)

Anisotropy: A

− λ

(5)

Eigenentropy: E

= −

∑

i=1

lnλ

(6)

Sum of Eigenvalues: Σ

∑

i=1

(7)

Local surface variation: C

∑

i=1

(8)

To calculate the eigenvalues, we ﬁrst calculate the

covariance matrix, of the points inside of the neigh-

bourhood and extract the eigenvalues from it, which

are then sorted in descending order. Example of how

these features look on a P40 sandpaper, calculated

from a neighbourhood size of 1mm is given in Figure

The third type of features are the statistical shape

distributions. These distributions are derived from the

work by (Osada et al., 2002), for parameterization of

the whole object’s shape, but for use as local features,

they can be used on neighbourhoods of points. These

distributions are calculated as histograms of randomly

sampled shape values. These shape values are based

on ﬁve metrics:

• D1 - distance from the neighbourhood centroid,

to a random point from the same neighbourhood

4(a),

• D2 - distance between two random points from

the neighbourhood 4(b),

• D3 - the square root of the area between three ran-

dom points from the neighbourhood 4(c),

• D4 - the cubic root of the volume of a tetrahedron

made from four random points from the neigh-

bourhood 4(d),

• A3 - the angle between three random points from

the neighbourhood 4(e).

We follow the suggestions by (Weinmann et al.,

2015) and calculate the distributions as histograms

with 10 bins and 255 random samplings from each

neighbourhood scale. These results in 100 statistical

distribution features per point. The visualization of

the ﬁve used metrics, can be seen in Figure 4.

These features are extracted from each of the re-

constructed sandpaper patches and used to train a

Random Forest classiﬁer. The same classiﬁer is used

in the work of (Weinmann et al., 2015) and (Nikolov

and Madsen, 2020) and is proven to provide good re-

sults, when used with 3D surface features.

4 EXPERIMENTAL SCENARIOS

AND RESULTS

Initially a simple classiﬁcation scenario is test, us-

ing control reconstructed sandpaper patches. These

patches are then classiﬁed by the trained Random For-

est algorithms, to prove that the selected features can

describe the surface of sandpaper patches of differ-

ent grit size. After that a second testing scenario is

proposed, using decommissioned wind turbine blades

and classifying their surface depending on how close

it resembles different sand paper grits.

The control sandpaper patches are from a differ-

ent brand and represent again three different grit sizes

- a rough large grit size P40, a medium grit size P80

and a ﬁne small grit size P180. They are reconstructed

using the same capturing protocol given in Subsection

3.2. The reconstructions are shown in Figure 5. These

reconstructions are then used as testing data, to deter-

mine if the chosen features could describe unknown

sandpaper surfaces, before transferring the knowledge

to blade surfaces. The prediction results are given in

Table 2. It can be seen that the highest percentage

prediction for each of the testing sandpaper patches,

showed in bold, corresponds to the real grit size of

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

804

(a) Linearity (b) Planarity (c) Sphericity (d) Omnivariance

(e) Anisotropy (f) Eigentropy (g) Sum of Eigenvalues (h) Local Surface Variation

Figure 3: Example local covariance features, extracted from the P40 sandpaper grit size. Just a small part of the sandpaper

leading edge is shown for easier visualization.

(a) D1 (b) D2 (c) D3 (d) D4 (e) A3

Figure 4: The ﬁve density metrics, used to calculate the statistical shape distributions, as proposed by (Osada et al., 2002).

(a) P40 (b) P80

Figure 5: The three sandpaper patches, used as an initial

test for how good are the used features for describing the

sandpaper surface.

that patch. Some of the other grit sizes are also de-

tected, as each of the patches contains smoother and

rougher areas. For the P180, some problems come

from the possible present reconstruction noise and

low frequency surface roughness (Figure 5(c)), which

can be classiﬁed as higher grit sizes.

For the second testing scenario three wind turbine

blades, with varying surface roughness are selected.

All the selected blades have been decommissioned

and contain both smooth and very rough and damaged

surface patches. One is a full blade, from which a

number of patches are selected for reconstruction and

the other two are smaller blade segments, which are

Table 2: Results from the initial sandpaper data test, show-

ing predictions for the three tested grit sizes P40, P80, P180

(horizontally). All three have been correctly predicted.

Grit Size P40 P80 P180

P40 0.761 0.013 0.002

P60 0.193 0.127 0.101

P80 0.029 0.696 0.010

P100 0.007 0.055 0.059

P120 0.003 0.104 0.253

P180 0.007 0.005 0.575

used whole for the experiments. Four patches are se-

lected from the full blade - two representing damaged

and very rough areas and two representing relatively

clean areas, with small amounts of surface roughness.

The large blade, together with the two blade segments

are shown in Figure 6.

(a) (b) (c)

Figure 6: The three wind turbine blades used for the second

experiment. Four patches with varying degree of rough-

ness are selected from the ﬁrst large blade (Figure 6(a)),

while the other two segments (Figure 6(b) and 6(c)) are used

whole.

Quantifying Wind Turbine Blade Surface Roughness using Sandpaper Grit Sizes: An Initial Exploration

805

(a) Patch 1 (b) Patch 1 (c) Patch 2 (d) Patch 2

(e) Patch 3 (f) Patch 3 (g) Patch 4 (h) Patch 4

(i) Segment 1 (j) Segment 1 (k) Segment 2 (l) Segment 2

Figure 7: Results from segmentation of the wind turbine blades, visualized in pseudocolor, where P40 grit is shown as red

and P180 is dark blue, while the other grit sizes are the colors between them. Just the point cloud is also given for each, for

easier visualization of the roughness.

The four patches and two blade segments are

reconstructed using Metashape, following the same

capturing protocol presented in Subsection 3.2. The

selected features, presented in 3.3, are extracted from

the blade reconstructions and used as testing data for

the Random Forest classiﬁer. The result point clouds

with pseudocolor information, on which sandpaper

roughness best describes each point is given in Fig-

ure 7. The P40 grit size is shown as red color and

the P180 as dark blue color, with all other grit sizes

represented with the in-between colors.

Because there is no ground truth for calculating

the accuracy of the surface classiﬁcation, the rough-

ness of each of the blade segments is ﬁrst calcu-

lated as the distance to a best ﬁtting plane. Each

plane is ﬁt to a spherical neighbourhood, which has

been heuristically selected to best describe all parts

of the point cloud. This is done through the use of

the roughness calculation functionality of CloudCom-

pare (Girardeau-Montaut, 2011). The distance be-

tween this roughness and our proposed method is then

computed for each point. The nominal average di-

ameter of the sandpaper grains given in Table 3.2 is

used. We then calculate their root mean square error

(RMSE) and standard deviation, which would give an

overview of how good the sandpaper roughness is ﬁt

to the blade surface. These values are presented in

Table 3. It can be seen that the damaged surface parts

are mostly classiﬁed as a P40 grit, with the parts that

have varying degrees of roughness classiﬁed as the

smaller grit sizes. The P60 and P120 have the least

amount of point classiﬁed as them. The smooth and

clean surfaces, especially on the two patches (Figure

7(f) and 7(h)) and on the blade segments are repre-

sented with P180 grit size. From Table 3 it can be seen

that on average the rougher patches exhibit a higher

RMSE between the sandpaper representation and the

real roughness. The blade segments on Figure 7(j)

has a larger RMSE value, because the damages on it

exhibit roughness values, which far exceed the values

of the sandpaper grits. This testing scenario can be

very dependent on the selected parameters and addi-

tional testing is required, with examples presented in

the Future Work section 5.

Table 3: RMSE and standard deviation between the best ﬁt-

ting plane roughness calculation and our proposed solution.

RMSE [mm] Std. Dev. [mm]

Patch 1 0.179 0.113

Patch 2 0.125 0.069

Patch 3 0.070 0.009

Patch 4 0.101 0.060

Segment 1 0.162 0.110

Segment 2 0.288 0.122

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

806

5 CONCLUSION AND FUTURE

WORK

In this paper we presented an idea for segmenting

wind turbine blade surfaces depending on the sandpa-

per grit size that best represents their roughness. This

solution aims to provide a standardized method clas-

sifying surface roughness of wind turbine blades that

can be used for calculating their energy output and

performance, as well as more easily modeling them

for tests in wind tunnels.

We selected six different sandpaper patches with

varying grit sizes and 3D reconstructed them using

SfM. We then extracted a number of geometrical, co-

variance and statistical features from neighbourhoods

with progressively smaller sizes. We used these fea-

tures to train a Random Forest classiﬁer.

To test the proposed solution we ﬁrst evaluated the

classier on a testing set of sandpaper patches. We

veriﬁed that the extracted features could be used to

identify each grit size of the training set. We then in-

troduced surface data from three wind turbine blades.

The data represented surfaces with varying degrees

of surface roughness and damages. These surfaces

were also 3D reconstructed and then used as a test-

ing dataset. We demonstrated that we Random For-

est classiﬁer managed to sufﬁciently segment the sur-

faces and to represent their roughness as sandpaper

grit sizes.

The current research lacks an in-depth veriﬁcation

of the robustness of the proposed methods. The next

steps for the current research can be divided in two

directions. The ﬁrst direction requires the creation of

ground truth 3D surface scans of the test blades and

extracting the roughness from patches using a pro-

ﬁlometer. These surfaces can then be used for more

clear comparison to the proposed solution. The sec-

ond direction would require the use of a wind tun-

nel, where the performance of a damaged blade and a

blade with sandpaper of different grit sizes mounted

on surfaces deﬁned by the proposed algorithm can be

compared.

REFERENCES

Agisoft (2010). Metashape. http://www.agisoft.com/. Ac-

cessed: 2019-09-20.

Al-Kaff, A., Moreno, F. M., San Jos

e, L. J., Garc

ıa, F.,

Mart

ın, D., de la Escalera, A., Nieva, A., and Garc

ea,

J. L. M. (2017). Vbii-uav: Vision-based infrastructure

inspection-uav. In World Conference on Information

Systems and Technologies, pages 221–231. Springer.

Amenabar, I., Mendikute, A., L

opez-Arraiza, A.,

Lizaranzu, M., and Aurrekoetxea, J. (2011). Com-

parison and analysis of non-destructive testing

techniques suitable for delamination inspection in

wind turbine blades. Composites Part B: Engineering,

42(5):1298–1305.

Bak, C., Gaunaa, M., Olsen, A. S., and Kruse, E. K. (2016).

What is the critical height of leading edge roughness

for aerodynamics. In Journal of Physics: Conference

Series, volume 753, page 022023.

Bentley, J. L. (1975). Multidimensional binary search trees

used for associative searching. Communications of the

ACM, 18(9):509–517.

Blomley, R., Jutzi, B., and Weinmann, M. (2016). Clas-

siﬁcation of airborne laser scanning data using geo-

metric multi-scale features and different neighbour-

hood types. ISPRS Annals of Photogrammetry, Re-

mote Sensing & Spatial Information Sciences, 3(3).

Chan, C. H. L., Wang, Q., Holden, R., Huang, S., and

Zhao, W. (2019). Optimal number of control points

for ﬁtting b-splines in wind turbine blade measure-

ment. International Journal of Precision Engineering

and Manufacturing, 20(9):1507–1517.

Chen, X. (2018). Fracture of wind turbine blades in opera-

tion—part i: A comprehensive forensic investigation.

Wind Energy, 21(11):1046–1063.

Dastoorian, R., Elhabashy, A. E., Tian, W., Wells, L. J.,

and Camelio, J. A. (2018). Automated surface in-

spection using 3d point cloud data in manufactur-

ing: A case study. In International Manufacturing

Science and Engineering Conference, volume 51371,

page V003T02A036. American Society of Mechani-

cal Engineers.

Dittrich, A., Weinmann, M., and Hinz, S. (2017). Analytical

and numerical investigations on the accuracy and ro-

bustness of geometric features extracted from 3d point

cloud data. ISPRS Journal of Photogrammetry and

Remote Sensing, 126:195–208.

Du, Y., Zhou, S., Jing, X., Peng, Y., Wu, H., and Kwok,

N. (2020). Damage detection techniques for wind tur-

bine blades: A review. Mechanical Systems and Sig-

nal Processing, 141:106445.

El-Mesery, H. S., Mao, H., and Abomohra, A. E.-F. (2019).

Applications of non-destructive technologies for agri-

cultural and food products quality inspection. Sensors,

19(4):846.

FEPA (1955). Fepa - federation of european produc-

ers of abrasives. https://www.fepa-abrasives.com/

abrasive-products/grains. Accessed: 2019-03-06.

Genc, M. S., Kemal, K., and Ac¸ikel, H. H. (2019). Investi-

gation of pre-stall ﬂow control on wind turbine blade

airfoil using roughness element. Energy, 176:320–

334.

Girardeau-Montaut, D. (2011). Cloudcompare-open source

project. OpenSource Project.

Hackel, T., Wegner, J. D., and Schindler, K. (2016). Fast se-

mantic segmentation of 3d point clouds with strongly

varying density. ISPRS annals of the photogramme-

try, remote sensing and spatial information sciences,

3:177–184.

Han, W., Kim, J., and Kim, B. (2018). Effects of contam-

ination and erosion at the leading edge of blade tip

Quantifying Wind Turbine Blade Surface Roughness using Sandpaper Grit Sizes: An Initial Exploration

807

airfoils on the annual energy production of wind tur-

bines. Renewable energy, 115:817–823.

ISO25178 (2012). Iso25178(2012) geometrical product

speciﬁcations (gps) — surface texture: Areal — part

2: Terms, deﬁnitions and surface texture parame-

ters. https://www.iso.org/obp/ui/#iso:std:iso:25178:

-2:ed-1:v1:en. Accessed: 2020-08-03.

ISO4287 (1997). Iso4287(1997) geometrical product spec-

iﬁcations (gps) — surface texture: Proﬁle method

— terms, deﬁnitions and surface texture parameters.

https://www.iso.org/standard/10132.html. Accessed:

2020-08-03.

ISO6344 (1998). Iso6344(1998) coated abrasives — grain

size analysis parts 1-3. https://www.iso.org/standard/

12643.html. Accessed: 2020-08-03.

Langel, C. M., Chow, R., Hurley, O. F., Van Dam, C.

C. P., Maniaci, D. C., Ehrmann, R. S., and White,

E. B. (2015). Analysis of the Impact of Leading Edge

Surface Degradation on Wind Turbine Performance.

In 33rd Wind Energy Symposium, Reston, Virginia.

American Institute of Aeronautics and Astronautics.

Langel, C. M., Chow, R., Van Dam, C., and Maniaci, D. C.

(2017). Rans based methodology for predicting the

inﬂuence of leading edge erosion on airfoil perfor-

mance. Technical report, Sandia National Lab.(SNL-

NM), Albuquerque, NM (United States).

Lowe, D. G. (2004). Distinctive image features from scale-

invariant keypoints. International journal of computer

vision, 60(2):91–110.

Martin, R. W., Sabato, A., Schoenberg, A., Giles, R. H., and

Niezrecki, C. (2018). Comparison of nondestructive

testing techniques for the inspection of wind turbine

blades’ spar caps. Wind Energy, 21(11):980–996.

Marzuki, O. F., Raﬁe, A. S. M., Romli, F. I., and Ahmad,

K. A. (2018). Magnus wind turbine: the effect of

sandpaper surface roughness on cylinder blades. Acta

Mechanica, 229(1):71–85.

Moreno, S., Pe

na, M., Toledo, A., Trevi

no, R., and Ponce,

H. (2018). A new vision-based method using deep

learning for damage inspection in wind turbine blades.

In 2018 15th International Conference on Electrical

Engineering, Computing Science and Automatic Con-

trol (CCE), pages 1–5. IEEE.

Nielsen, M., Nikolov, I., Kruse, E., Garnæs, J., and Madsen,

C. (2020). High-resolution structure-from-motion for

quantitative measurement of leading-edge roughness.

Energies, 13(15).

Nikolov, I. and Madsen, C. (2016). Benchmarking close-

range structure from motion 3d reconstruction soft-

ware under varying capturing conditions. In Euro-

Mediterranean Conference, pages 15–26. Springer.

Nikolov, I. and Madsen, C. (2020). Rough or noisy? metrics

for noise estimation in sfm reconstructions. Sensors,

20(19):5725.

Osada, R., Funkhouser, T., Chazelle, B., and Dobkin, D.

(2002). Shape distributions. ACM Transactions on

Graphics (TOG), 21(4):807–832.

Ozyesil, O., Voroninski, V., Basri, R., and Singer, A.

(2017). A survey of structure from motion. arXiv

preprint arXiv:1701.08493.

Peng, L. and Liu, J. (2018). Detection and analysis of large-

scale wt blade surface cracks based on uav-taken im-

ages. IET Image Processing, 12(11):2059–2064.

Rublee, E., Rabaud, V., Konolige, K., and Bradski, G.

(2011). Orb: An efﬁcient alternative to sift or surf.

In 2011 International conference on computer vision,

pages 2564–2571. Ieee.

Schramm, M., Rahimi, H., Stoevesandt, B., and Tangager,

K. (2017). The Inﬂuence of Eroded Blades on Wind

Turbine Performance Using Numerical Simulations.

Energies, 10(9):1420.

Shihavuddin, A., Chen, X., Fedorov, V., Nymark Chris-

tensen, A., Andre Brogaard Riis, N., Branner, K.,

Bjorholm Dahl, A., and Reinhold Paulsen, R. (2019).

Wind turbine surface damage detection by deep

learning aided drone inspection analysis. Energies,

12(4):676.

Triggs, B., McLauchlan, P. F., Hartley, R. I., and Fitzgibbon,

A. W. (1999). Bundle adjustment—a modern synthe-

sis. In International workshop on vision algorithms,

pages 298–372. Springer.

Wang, L. and Zhang, Z. (2017). Automatic Detection of

Wind Turbine Blade Surface Cracks Based on UAV-

Taken Images. IEEE Transactions on Industrial Elec-

tronics, 64(9):7293–7303.

Weinmann, M., Jutzi, B., Hinz, S., and Mallet, C. (2015).

Semantic point cloud interpretation based on optimal

neighborhoods, relevant features and efﬁcient classi-

ﬁers. ISPRS Journal of Photogrammetry and Remote

Sensing, 105:286–304.

Xu, D., Wen, C., and Liu, J. (2019). Wind turbine blade sur-

face inspection based on deep learning and uav-taken

images. Journal of Renewable and Sustainable En-

ergy, 11(5):053305.

Zhang, D., Burnham, K., Mcdonald, L., Macleod, C., Do-

bie, G., Summan, R., and Pierce, G. (2017a). Remote

inspection of wind turbine blades using uav with pho-

togrammetry payload. In 56th Annual British Confer-

ence of Non-Destructive Testing-NDT 2017.

Zhang, D., Watson, R., Dobie, G., MacLeod, C., Khan,

A., and Pierce, G. (2020). Quantifying impacts on

remote photogrammetric inspection using unmanned

aerial vehicles. Engineering Structures, page 109940.

Zhang, Y., Yuan, X., Fang, Y., and Chen, S. (2017b).

Uav low altitude photogrammetry for power line

inspection. ISPRS International Journal of Geo-

Information, 6(1):14.

Zhou, Q.-Y., Park, J., and Koltun, V. (2018). Open3D:

A modern library for 3D data processing.

arXiv:1801.09847.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

808