Vegetation Filtering using Colour for Monitoring Applications from

Photogrammetric Data

M. Amparo Núñez-Andrés

, Albert Prades

and Felipe Buill

Civil and Environmental Department, Universitat Politècnica de Catalunya, BarcelonaTECH,

Dr. Marañon 44-50, Barcelona, Spain

Keywords: Vegetation Filter, Colour Space, RGB, Image Classification, Point Cloud Segmentation.

Abstract: Photogrammetry is one of the widest techniques used to monitor terrain changes which occur due to natural

process and geological natural risk zones. In order to carry out terrain monitoring, it is necessary to eliminate

all the non-ground elements. One of the most variable elements in this monitoring is the presence of vegetation,

which obscures the ground and can significantly mislead any multitemporal analysis to detect terrain changes.

Therefore, the focus of this paper is about how best to filter the vegetation to attain an accurate reading of the

terrain. There are several methods to filter it based on colourising an excessive greenness vegetation index or

non-visible channels as the IR in the well-known index NVDI. However, achieving this kind of information

is not always possible because its high cost. Instead this channel we can add new information using the HSV

colour space obtained from the RGB information. In this paper, we propose a double possibility, on one hand

work with RGB+HSV for a supervised segmentation on images. On the other, to use excessive greenness

vegetation indices and RGB+HSV for the segmentation of point clouds. The results shown that the use of

additional channels HSV can significantly improve the segmentation in both studies, and therefore render a

much more accurate assessment of the underlying terrain.

1 INTRODUCTION

Monitoring active geological hazard zones is a

common activity developed by public institutions and

scientific groups. The objective is to improve

understanding of these events, to better predict their

occurrence and therefore minimise any humanitarian

impact. Some of these geological movements are

preceded by slow movements that can be detected,

that are essential to monitor (Jaboyedoff et al., 2012)

(Kamps et al., 2017). Therefore, the repetition of the

surveys to measure, analyse and map any changes

over time is necessary in active zones (Niethammer et

al., 2012) (Travelletti et al., 2012) (Barbarella et al.,

2015) (Núñez-Andrés et al., 2019). Of the types of

hazards classified according to the Varnes

classification (Varnes, 1978) we will limit the current

study to rockfalls. Therefore, our area of study is

quasi-vertical rocky walls, and the unstable blocks are

the monitoring objective.

https://orcid.org/0000-0003-2745-7759

https://orcid.org/0000-0002-0164-1681

https://orcid.org/0000-0002-9222-0072

In order to carry out this task, several geomatics

techniques are commonly used such as surveying

with total station, GNSS (Global Navigation Spatial

System), LiDAR, TLS (Terrestrial Laser Scanning)

and photogrammetry methods.

During last decade digital photogrammetry has

become commonly used for such monitoring

applications. This includes both terrestrial and aerial

methods, with RPAS (Remotely Piloted Aircraft

Systems) becoming increasingly prevalent. However,

in using RPAS systems we have to considerer flight

restrictions in many countries, mainly in protected

natural zones (Dolan & Thompson, 2014)(Stöcker et

al., 2017). In these cases, only the terrestrial option is

possible.

Stereo-photogrammetric methods are the most

widely applied since they allow us to build 3D models

of the terrain from the raw photographic data

(Eisenbeiβ et al., 2005) (Roncella et al., 2014)

(Tziavou et al., 2018). The result of a digital

Núñez-Andrés, M., Prades, A. and Buill, F.

Vegetation Filtering using Colour for Monitoring Applications from Photogrammetric Data.

DOI: 10.5220/0010523300980104

In Proceedings of the 7th International Conference on Geographical Information Systems Theory, Applications and Management (GISTAM 2021), pages 98-104

ISBN: 978-989-758-503-6

 2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reser ved

photogrammetric survey is a 3D point cloud with

similar characteristics to a laser scanning survey but

at a cheaper cost.

However, regardless of the method used, the key

challenge to obtaining accurate results - either for real

time or multitemporal change monitoring via

punctual campaigns - is the existence of elements that

are not bare ground which therefore have to be

eliminated. As a result, before the change analysis can

occur a process of segmentation or classification

ground/non_ground must be done.

One of the main elements that can mislead any

analysis of the ground is vegetation. It suffers gradual

variations due to its own growing, seasonal changes

of colour and morphology, disappearance, etc.

Therefore, it is necessary to detect and eliminate it

before the terrain changes and movements detection

analysis.

Using LiDAR’s sensors mounted on planes or

RPAS allows an easier filtering because they capture

several signal echoes that allows us to classify data

points as either ground or non-ground. However, its

high cost and the difficulty to obtain data with an

aerial LiDAR on a vertical rocky wall, where the

rockfalls happen, invalidate its choice.

The alternative is to work with photogrammetric

process and use the color information to segment the

point cloud. In this case, nowadays there are several

solutions: commercial software’s and free plugins

that deal with this problem but they only achieve a

partial solution, it remains as a difficult issue.

This paper establishes a new approach to perform

the segmentation of ground/vegetation on rocky wall

faces. The vegetation must be excluded from the

monitoring since its changes are not related with

terrain movements. In the next sections, we provide

two options to exclude the vegetation, which work

directly with the original images, before building the

model or with the point clouds.

2 METHODOLOGY

There are a lot of methods to detect vegetation from

satellite and airborne images. Most of them uses

vegetation indices obtained from different channels,

with NDVI (Normalized Difference Vegetation

Index) (Xue & Su, 2017) being one of the most

known. It is calculated from the Red and NIR (Near

Infrared) channels. Therefore, it could be applied in

terrain images from hyperspectral cameras. The main

drawback of this solution is the high cost of this

equipment.

For a point clouds segmentation, there are a

variety of methods for MMS (Mobile Mapping

Systems) urban survey. Most of them are based on

information such as signal intensity in addition to

geometry (Yadav et al., 2016). In natural

environments, these methods are less efficient since

natural features such as vegetation do not fit to regular

shapes.

If the point clouds come from a photogrammetric

process, the algorithm to segment bare ground and

vegetation involves the use of the height information

without taking into account the error in it (Becker et

al., 2017). Moreover, this information is not useful

when we are working on a vertical and irregular rocky

walls, source areas of rockfalls. Therefore, to look for

other options is necessary.

We work with two types of data: original images

and point clouds. In the first case, having classified

the image into ground and non-ground areas, a binary

mask can be applied to the images before the

photogrammetric process. In this way, only the

ground areas will be used to build the terrain model.

On the other hand, after the point cloud

classification process we get two files, one with the

vegetation and other with the bare ground. The last

one is that can be used for the terrain change analysis.

The approach that this paper proposes is based on

using only the RGB channels, and the indices and

HSV colour space derived from this information.

The vegetation indices are based on the visible

channels used in the VVI (Visible Vegetation Index)

and the ExG (Excessive Greenness), equation (1) and

(2) respectively (Ponti, 2013).

The VVI values are between 0 and 1, where low

values correspond to bare ground, and values near 1

to vegetation. The vector RGB

is the reference green

colour and w the weight (Ponti 2013).

𝑉𝑉𝐼



1



𝑅𝑅



𝑅𝑅





1



𝐺𝐺



𝐺𝐺





1



𝐵𝐵



𝐵𝐵







(1)

Before calculating the ExG it is necessary to

normalize the values RGB. This is a continuous

index, as in the VVI the low values indicate the

presence of bare ground and the highest imply

vegetation.

𝐸𝑥𝐺2𝐺 𝑅 𝐵

(2)

RGB colour models are commonly used since

they refer to the biological processing of colours in

the human visual system (Loesdau et al., 2014).

However, they are not always is the best. According

to the literature, the Hue-Saturation-Value (HSV)

colour space has more robustness under variable

illumination conditions. Therefore, this study further

assesses the HSV colour space for automatic

vegetation-ground classification. The conversion

from RGB to HSV is described in the equations

(3)(4)(5)

Vegetation Filtering using Colour for Monitoring Applications from Photogrammetric Data

𝑉max 𝑅𝐺𝐵 (3)

𝑆

0 𝑖𝑓max



𝑅𝐺𝐵



0

1

min



𝑅𝐺𝐵



max



𝑅𝐺𝐵



𝑜𝑡ℎ𝑒𝑟 𝑐𝑎𝑠𝑒𝑠

(4)

(5)

- Non defined 𝑖𝑓 𝑚𝑎𝑥𝑚𝑖𝑛

- 60





 0° 𝑖𝑓 𝑚𝑎𝑥𝑅 𝑎𝑛𝑑 𝐺𝐵

- 60





 360° 𝑖𝑓 𝑚𝑎𝑥𝑅 𝑎𝑛𝑑 𝐺B

- 60





120° 𝑖𝑓𝑚𝑎𝑥 𝐺

- 60





 240° 𝑖𝑓 𝑚𝑎𝑥 𝐺𝐵

The range of values for the H coordinate is from

0º to 360º, from 0 to 1 for S and V, since the RGB

values are normalized.

2.1 Data

The study was carried out in several zones of

Catalonia, although we show in detail only one of

them. This area belongs to the Riera Gavarresa

(Catalonia, Spain). It is a slope with an approximate

area of 600 m

. A set of stereo-photogrammetric

images taken in two different seasons, spring and

summer, are available with a GSD of 2 cm.

For the photogrammetric process, block

adjustment and georeferenced, we have used the

Agisoft Metashape software.

The 3D point cloud model obtained has a high

quality, Figure 1. It is dense, homogeneous and has

information of the rocky wall and the vegetation

around and on it. The vegetation covers the top of the

slope with trees, the face of the slope is covered in

some parts by shrubs, and where it finishes the ground

is covered by grass some bushes and small trees.

One option that could have been used would be

the tools implemented in Agisoft Metashape to

eliminate vegetation and to achieve a model free of

vegetation. That can be compared with the one

obtained with our approach. But it was not possible

because the software does not allow us to customize

these filtering tools and part of the vegetation

remains.

Figure 1: 3D point cloud model obtaining by

photogrammetry.

2.2 Image Classification

The image classification was carried out using the

commercial software ENVI.

In the case of working with the original images

we have chosen a supervised classification with both

visible bands, RGB, and six bands, RGB+HSV.

The first step was to build the RGB+HSV images.

After that, the regions of interest (ROI’s) were chosen

to train the classification model. We considered three

different set of classes:

 5 classes: sky, vegetation, shadow in

vegetation, rocky wall, shadow in rocky wall.

 4 classes: sky, shrubs and grass, trees, rocky

wall.

 3 classes: sky, vegetation, rocky wall.

The three classifications were applied on the two

images sets: RGB channels, and RGB+HSV.

The image classification was made with two

common methods (Richards, 2013), the Euclidean

minimum distance and the Maximum Likelihood

Estimation (MLE). After the analysis of the results of

the classification and the confusion matrix, the first

one was discarded and the process was continued

using only with MLE. Whilst the general accuracy

with MLE is approximately 98% for all the sets, it

was roughly 85% using the minimum distance

method.

2.2.1 Results

The results obtained for all the images set and all the

classification groups are satisfactory for both RGB

and RGB+HSV. However, we can highlight a slight

improvement when the number of channels is

increased.

GISTAM 2021 - 7th International Conference on Geographical Information Systems Theory, Applications and Management

100

Figure 2: Maximum Likelihood Estimation classification in

4 classes results: a) original RGB image, b) RGB channels,

c) RGB+HSV channels.

Figure 2 shows the classification results using

four classes. The upper image is the original image,

in the centre we locate the classification with an RGB

image, Figure 2. b). It can be seen that the boundary

between classes is less defined than in Figure 2.c),

which uses both RGB+HSV channels, mainly

between the sky and trees class.

With regards the general results, in all the cases

the use of RGB+HSV images allows to achieve a

slight improvement, of less than 1%, in the overall

accuracy compared to the RGB-only analysis.

However, this improvement is more pronounced in

the vegetation classification, overall in the shrubs and

grass class, when the number of classes considered is

decreased.

Using three, four, or five classes the behaviour is

similar. The accuracy in the omission decreased when

the additional information of the hue is considered in

the classification.

2.3 Point Cloud Segmentation

Working with point clouds, the segmentation has

been carried out considering three parameters: the

colour indexes VVI, the ExG, and finally with the

coordinates of the colour space HSV.

We have implemented in Python language a

program to read the point cloud and calculated the

two indices and the HSV coordinates.

For the VVI calculation the values for the vector

of the reference green colour RGB

, eq (1), are

=60.0, G

=70.0 and B

=30.0, and the weight

component w= 1.

The problem in this index is to set the reference

values, since they depend on the image and the

characteristics of the zone.

After the colour space change, it was checked that

the hue coordinate, such as the eq. (5) shows its value

range from 0º to 360º, is the one that provides most

useful information for the segmentation. For this

reason, the value and the saturation coordinates have

not been considered in the classification.

Analysing the distribution of frequencies for the

VVI, ExG and H values it can been observed that they

follow a bimodal curve, Figure 3, and the minimum

between then could be established as the threshold to

segment the point cloud. In this way, during filtering

using the ExG index we have considered vegetation

points with ExG > -0.02, Figure 3.

In all the areas classified the hue coordinate shows

a similar behaviour. The higher values, more than

300º, are located in areas covered by shadows, Figure

4. Therefore, we have to considerer the minimum

between the distributions and then the tile to segment

correctly the point cloud.

Vegetation Filtering using Colour for Monitoring Applications from Photogrammetric Data

101

Figure 3: Bimodal curve from the ExG values.

It has been considered that if we have more

classes than bare ground and vegetation this approach

should be reconsidered to find the thresholds.

Figure 4: Bimodal curve from the H values.

The result after executing the program are for each

index two files, one with the vegetation and other

with the bare ground and rocks, i.e. six files are

obtaining.

2.3.1 Results

Figure 5 shows the results of the classification using

the different indices. In all the images the bare ground

has been represented in grey colour. In a qualitative

analysis, as the images shown the best results are

obtained using the ExG index and the hue value,

Figures 5.b) and c) respectively. The worst results are

gotten with the VVI, which shows as vegetation the

shadow areas.

The Figure 5.d) shows in red the shadow areas

where the H value reaches more than 300º. These

extreme values are represented in the tile of the

histogram as we have mentioned previously.

Figure 5: Results after the classification, a) VVI, b) ExG, c)

Hue values eliminating the tile of the distribution and d)

Hue values with the values higher 300º.

A common problem is due to the lack or existence

of misleading information in the treetops. Since the

branches and leaves are not in the same position in all

the images, by the wind effect, they appear in the 3D

model without colour. Therefore, they cannot be

classified using this characteristic.

Moreover, we have classified the point cloud with

Canupo (Brodu & Lague, n.d.), Figure 6, to test the

GISTAM 2021 - 7th International Conference on Geographical Information Systems Theory, Applications and Management

102

influence of the use of geometry components in the

result improvement. In a qualitative comparison with

the results, Figure 5, we can see that there are zones

of bare ground classified as vegetation, such as with

the use of VVI.

Figure 6: Results of classification with Canupo, in red

vegetation and blue bare ground.

Table 1 shows a quantitative comparison of the

filter methods: the number of points classify as bare

ground, vegetation, and percentages of filtered

vegetation points by the three indices. The values of

shadows, the tile in the histogram, are not considered

for the hue.

Table 1: Number of points in the filtered point clouds and

the percentage of the correct vegetation filtered.

Groun

etation % filtere

VVI 3.39M 1.09M 75.8

ExG 3.40M 1.08M 94.8

Hue 3.40M 1.08M 95.3

Canupo 3.06M 1.58M 62.5

The ExG and H classify almost the whole of the

vegetation points. The percentages of commission are

around 2% and 1.5%, respectively, mainly in the

zones of shadow boundary between the ground and

the bushes. This percentage increases dramatically

when we use Canupo.

In other study areas, where the point cloud was

obtained by TLS with only an internal camera, the

results were worst. Mainly, due to the illumination is

not the best, since the sensor takes photographs in an

automatic way with the same parameter for all of

them.

3 DISCUSION AND FINAL

REMARKS

The filters that were used in this study have the

advantage that are easy to implement. Moreover, the

results let us be optimistic in the vegetation filtering

using only colour information though indices or the

values of Hue derived directly from the visible bands.

We have used images taken with a common

camera, therefore the cost of this sensor is cheaper

than other methods. The use of the NIR band could

improve the results but these cameras are more

expensive if the resolution want to be kept.

The ExG index and the H value in the case of point

clouds allows us a good classification in the areas

where the model have been built correctly. In

elements that appear moved between a photographs

and other such as the leaves moved by the wind, the

colour is not assigned to the point in the point cloud

and therefore cannot be classify by this characteristic.

Whether the classification is made directly on the

images the use of the H value added to the RGB bands

allow to improve the results without increasing

significantly the time of computation.

4 CONCLUSIONS

In the monitoring of rocky massif to prevent rockfalls

is not possible to use LiDAR due to the verticality of

the walls. Therefore, we cannot take advantage of the

use of several signal echoes to classify vegetation.

Multi-spectral cameras are an alternative. However,

their high cost makes us to discard them.

In this paper, we have proposed the use of the

RGB channels for two purposes, on one hand to

obtain vegetation indices based on the visible

channels and, on the other hand to add additional

HSV channels to the images and point cloud

previously to classify them.

The results show that the combination of the RGB

bands allows us to filter vegetation with methods of

simple application using a common camera, therefore

it is cheaper than other options.

The image classification using as additional band

the Hue value derived of the RGB improves the

result. This improvement is higher when the number

the classes decrease.

The point cloud classification has good results

using the ExG and H values. The success of VVI

index is conditioned by the reference vector, whose

adjustment depends on the scene to classify and it is

more influenced by colour variations and shadows.

However, shadows and boundary zones are where

more misleading classification exist, in all cases.

Vegetation Filtering using Colour for Monitoring Applications from Photogrammetric Data

103

ACKNOWLEDGEMENTS

This study was supported by the National Research

Project "Advances in rockfall quantitative risk

analysis (QRA) incorporating developments in

geomatics (GeoRisk)” funded by the Spanish

Ministry of Economy and Competitiveness, and co-

funded by the Agencia Estatal de Investigación

(AEI) on the framework of the State Plan of

Scientific-Technical Research and Innovation with

reference code PID2019-103974RB-I00/AEI/

10.13039/501100011033.

REFERENCES

Barbarella, M., Fiani, M., & Lugli, A. (2015). Landslide

Monitoring Using Multitemporal Terrestrial Laser

Scanning for Ground Displacement Analysis.

Geomatics, Natural Hazards and Risk 6(5–7):398–418.

Becker, C., Häni, N., Rosinskaya, E., D’Angelo, E., &

Strecha., C. (2017). Classification of aerial

photogrammetric 3D point clouds. ISPRS Annals of the

Photogrammetry, Remote Sensing and Spatial

Information Sciences 4(1W1):3–10.

Brodu, N., & Lague, D. (n.d.). 3D Terrestrial lidar data

classification of complex natural scenes using a multi-

scale dimensionality criterion : applications in

geomorphology .

Dolan, A. M., & Thompson, R. M. (2014). Integration of

drones into domestic airspace: Selected legal issues.

Domestic Drones: Elements and Considerations for the

U.S., 1–41.

Eisenbeiβ, H., Lambers, K., & Sauerbier, M. (2005).

Photogrammetric recording of the archaeological site of

Pinchango Alto (Palpa, Peru) using a mini helicopter

(UAV). In A. Figueiredo (Ed.), Annual Conference on

Computer Applications and Quantitative Methods in

Archaeology CAA (Issue March 2005, pp. 21–24).

http://www.photogrammetry.ethz.ch/general/persons/k

arsten/paper/eisenbeiss_et_al_2007.pdf

Jaboyedoff, M., Oppikofer, T., Abellán, A., Derron, M. H.,

Loye, A., Metzger, R., & Pedrazzini, A. (2012). Use of

LIDAR in landslide investigations: A review. Natural

Hazards, 61(1), 5–28. https://doi.org/10.1007/s11069-

010-9634-2

Kamps, M. T., Bouten, W., & Seijmonsbergen, A. C.

(2017). LiDAR and orthophoto synergy to optimize

object-based landscape change: Analysis of an active

landslide. Remote Sensing, 9(8).

https://doi.org/10.3390/rs9080805

Loesdau, M., Chabrier, S., & Gabillon, A. (2014). Hue and

Saturation in the RGB Color Space BT - Image and

Signal Processing. Springer International Publishing

Switzerland 2014, 203–212.

Niethammer, U., James, M. R., Rothmund, S., Travelletti,

J., & Joswig, M. (2012). UAV-based remote sensing of

the Super-Sauze landslide: Evaluation and results.

Engineering Geology, 128, 2–11.

https://doi.org/10.1016/j.enggeo.2011.03.012

Núñez-Andrés, M. A., Buill, F., Hürlimann, M., & Abancó,

C. (2019). Multi-temporal analysis of morphologic

changes applying geomatic techniques. 70 years of

torrential activity in the Rebaixader catchment (Central

pyrenees). Geomatics, Natural Hazards and Risk,

10(1), 314–335.

https://doi.org/10.1080/19475705.2018.1523235

Ponti, M.P. (2013). Segmentation of Low-Cost Remote

Sensing Images Combining Vegetation Indices and

Mean Shift. IEEE Geoscience and Remote Sensing

Letters 10(1):67–70.Varnes, DJ. 1978. “Transportation

Research Board Special Report: Slope Movement

Types and Processes.” Pp. 11–33 in Landslides,

analysis and control., edited by K. R. ( Schuster RL.

Washington D.C: National Academy of Sciences.

Roncella, R., Forlani, G., Fornari, M., & Diotri, F. (2014).

Landslide monitoring by fixed-base terrestrial stereo-

photogrammetry, ISPRS Ann. Photogramm. Remote

Sens. Spatial Inf. Sci., II-5, 297–304,

https://doi.org/10.5194/isprsannals-II-5-297-2014,

2014.

Richards, J. A. (2013). Remote sensing digital image

analysis: An introduction. (Vol. 9783642300622, pp.

1–494). Springer-Verlag Berlin Heidelberg.

https://doi.org/10.1007/978-3-642-30062-2

Stöcker, C., Bennett, R., Nex, F., Gerke, M., &

Zevenbergen, J. (2017). Review of the current state of

UAV regulations. Remote Sensing, 9(5), 33–35.

https://doi.org/10.3390/rs9050459

Travelletti, J., Delacourt, C., Allemand, P., Malet, J. P.,

Schmittbuhl, J., Toussaint, R., & Bastard, M. (2012).

Correlation of multi-temporal ground-based optical

images for landslide monitoring: Application, potential

and limitations. ISPRS Journal of Photogrammetry and

Remote Sensing, 70, 39–55.

https://doi.org/10.1016/j.isprsjprs.2012.03.007

Tziavou, O., Pytharouli, S., & Souter, J. (2018). Unmanned

Aerial Vehicle (UAV) based mapping in engineering

geological surveys: Considerations for optimum

results. Engineering Geology, 232(June 2017), 12–21.

https://doi.org/10.1016/j.enggeo.2017.11.004

Varnes, D. (1978). Transportation Research Board Special

Report: slope movement types and processes. In K. R.

( Schuster RL (Ed.), Landslides, analysis and control.

(pp. 11–33). National Academy of Sciences.

Xue, J., & Su, B. (2017). Significant remote sensing

vegetation indices: A review of developments and

applications. Journal of Sensors, 2017.

https://doi.org/10.1155/2017/1353691

Yadav, M., Lohani, B., Singh, A.K., & Husain., A. (2016).

“Identification of Pole-like Structures from Mobile

Lidar Data of Complex Road Environment.”

International Journal of Remote Sensing 37(20):4748–

77.

GISTAM 2021 - 7th International Conference on Geographical Information Systems Theory, Applications and Management

104