International Journal of

Geo-Information

Technical Note
Development of a QGIS Plugin to Obtain Parameters
and Elements of Plantation Trees and Vineyards with
Aerial Photographs
Lia Duarte 1,2, * ID
, Pedro Silva 1 and Ana Cláudia Teodoro 1,2 ID

1 Department of Geosciences, Environment and Land Planning, Faculty of Sciences, University of Porto,
Porto 4169-007, Portugal; [email protected] (P.S.); [email protected] (A.C.T.)
2 Earth Sciences Institute (ICT), Faculty of Sciences, University of Porto, Porto 4169-007, Portugal
* Correspondence: [email protected]; Tel.: +351-220-402-477

Received: 4 January 2018; Accepted: 12 March 2018; Published: 14 March 2018

Abstract: Unmanned Aerial Vehicle (UAV) imagery allows for a new way of obtaining geographic
information. In this work, a Geographical Information System (GIS) open source application was
developed in QGIS software that estimates several parameters and metrics on tree crown through image
analysis techniques (image segmentation and image classification) and fractal analysis. The metrics that
have been estimated were: area, perimeter, number of trees, distance between trees, and a missing tree
check. This methodology was tested on three different plantations: olive, eucalyptus, and vineyard.
The application developed is free, open source and takes advantage of QGIS integration with external
software. Several tools available from Orfeo Toolbox and Geographic Resources Analysis Support
System (GRASS) GIS were employed to generate a classified raster image which allows calculating
the metrics referred before. The application was developed in the Python 2.7 language. Also, some
functions, modules, and classes from the QGIS Application Programming Interface (API) and PyQt4
API were used. This new plugin is a valuable tool, which allowed for automatizing several parameters
and metrics on tree crown using GIS analysis tools, while considering data acquired by UAV.

Keywords: unmanned aerial vehicle; multispectral imagery; NDVI imagery; QGIS; python; plugin;
imagery segmentation; imagery classification; fractal analysis; plantations; orthophotos

1. Introduction
Remote sensing techniques provide useful information for agricultural management enhancing
the reliable automation of mapping procedures [1]. In recent years, Unmanned Aerial Vehicle (UAV)
imagery has allowed a new way of getting geographic information from remote sensing data [2]. UAV
technology provides several advantages compared to a manned aircraft such as: (i) significantly lower
qualification requirements for UAV pilots; (ii) allows for mapping small areas, resulting in very high
spatial resolution (VHR) images, which are ideal for detecting small objects; (iii) low cost acquisition;
and, (iv) it can couple different sensors [2]. UAVs are increasingly used in different contexts and in
several research areas, such in cartography, environmental studies, cultural heritage, civil engineering,
forestry, and agriculture/precision agriculture [3–8]. Therefore, when considering UAV imagery, it is
possible to acquire information more efficiently and inexpensively [9]. UAV imagery have been used
to estimate metrics from trees in several crops or farms, which can be of interest of farmers in order to
obtain a better planning of the irrigation process, helping in the forecast of the production date and
climate change effects as a consequence of deforestation and degradation of forests, which leads to
global warming, precision forestry, forest management planning, and biomass estimation [2,10–13].
Most of the work conducted with UAV data has been related with forest areas [11,12,14–19]. However,
in the field of agriculture crops, it is relevant to study the geometric characterization of trees and

ISPRS Int. J. Geo-Inf. 2018, 7, 109; doi:10.3390/ijgi7030109 www.mdpi.com/journal/ijgi

ISPRS Int. J. Geo-Inf. 2018, 7, 109 2 of 20

plantations. This is a complex task, being particularly important to understand the tree growth and
productivity [1,9,13,20–22]. In the field of precision agriculture, several image processing methods for
automatic detection and delineation of individual trees were already implemented [1]. The information
regarding tree crown areas of a specific crop is the key in the study of ecological value and ratio of
the occupied area by that crop [9,15,23]. Traditionally, tree counting has been performed manually
based on field information and statistical data, being an expensive process, time consuming, and a
tedious task subject to several errors [13,14,22]. Satellite, manned aircraft, or UAV imagery can be
used to obtain these data. Therefore, the implementation of automatic processes is crucial to extract
information and provide forest or crops inventories from remote sensing data [2,13,22].
Several studies have been developed with different techniques to estimate the tree crown
delineation, tree counting, and tree characterization from remote sensing imagery. Deng et al. [23]
proposed an artificial intelligence algorithm based on seeded based region growth method to extract
tree crown from Quickbird satellite images. In this work, after the segmentation process, a condition
regarding the Normalized Difference Vegetation Index (NDVI > 0.38) was considered. Santoro et al. [1]
proposed an algorithm based on an asymmetric weighted kernel optimized for tree counting using
GeoEye-1 satellite imagery. Park et al. [10] developed a tree isolation algorithm based on aerial imagery
for precise tree detection and crown delineation. Through UAV imagery acquired from a Hexacoptor,
Bazi et al. [2] extracted a set of candidate key points using Scale Invariant Feature Transform (SIFT)
analyzed with the Extreme Learning Machine (ELM), which is a kernel-based classification method.
Processing operations using mathematical morphology were used to distinguish palm trees from
other type of vegetation. Kang et al. [9] proposed an identification method for tree crown areas from
imagery captured by UAV. The methodology was based on mathematical morphological filtering,
unsupervised segmentation based on J-value SEGmentation (JSEG), local spatial statistics and Iterative
Self-Organizing Data Analysis Technique Algorithm (ISODATA). Panagiotidis et al. [24] used the
Structure from motion (Sfm) approach to generate a point cloud that was used to calculate a Digital
Surface Model (DSM) and an orthomosaic. This information was used as input data for tree crown
delineation using local maxima filtering (for tree heights) and Inverse Watershed Segmentation
(IWS) for tree crown detection. Thiel and Schmullius [17] also used Sfm for generating remote
sensing products from UAV imagery and LASTools software package to generate the point cloud.
The tree detection was based on pit-free Canopy Height Models. Hassaan et al. [11] created a script
in Matlab to remove barrel distortions from UAV images, k-means clustering in the segmentation
stage, and texture analysis. Finally, to count the trees a circle fitting approximation technique was used.
Nevalainen et al. [12] investigated the performance of UAV and hyperspectral imagery in individual
tree detection and tree species classification in forests computing the spectral features using Matlab.
Chianucci et al. [25] used Red-Green-Blue (RGB) UAV images to segment tree canopies (tree/non tree
mask) based on visible-vegetation indices and colour space conversion and conclude that the true
colour UAV images can be effectively used to obtain meaningful estimations of forest crown attributes.
Object-Based Image Analysis (OBIA) has proven to be a valuable approach to image segmentation
and classification when considering the form, color, and textures of the pixels, minimizing the boundary
problem [22,26–29]. OBIA groups the local homogeneous pixels, i.e., it allows for mapping the VHR
imagery into meaningful objects [27].
In the referred studies proprietary software were considered. However, there are several open
source software available such as Spring, Orfeo ToolBox (OTB)/Monteverdi, GRASS and QGIS which
are complemented with processing image algorithms [28]. The open source Geographical Information
System (GIS) software, such as QGIS, presented some advantage since it integrates algorithms from
OTB and GRASS through Processing Toolbox framework [30–33]. The UAV imagery can be considered
as an input in GIS applications as it provides the tools that are required to correctly manipulate,
analyze, and incorporate geographic information. In fact, the open source concept implies the
creation/development of new applications. For instance, Grippa et al. [34] presented a semi-automated
processing chain, developed in Python language, for object-based land-cover and land-use classification
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 3 of 20

concept implies the creation/development of new applications. For instance, Grippa et al. [34]
presented
ISPRS a semi-automated
Int. J. Geo-Inf. 2018, 7, 109 processing chain, developed in Python language, for object-based 3 of 20
land-cover and land-use classification using GRASS GIS and R open source software. Also, Duarte et
al. [35] developed a GIS open source application for image processing automatizing the
using GRASS GIS and
photogrammetric R open source
procedures, such as software.
computing Also,tieDuarte
pointsetbetween
al. [35] developed a GIS open
pairs of images, source
computing
application for image processing automatizing the photogrammetric
relative orientations based on calibration models, sparsing three-dimensional (3D) models, procedures, such as computing
tie points between
generating orthomosaics,pairs of images,
point cloud,computing
DSM, andrelative
shaded orientations
reliefs, through based
the on calibration
integration models,
of MicMac
sparsing three-dimensional
open source software in QGIS. (3D) models, generating orthomosaics, point cloud, DSM, and shaded
reliefs, through the integration of MicMac open source software in QGIS.
The objective of this work was the creation of a GIS open source application/plugin—Tree
The objective
Crown—that of thisseveral
estimates work was the creation
parameters of a GISon
and metrics opentreesource
crownapplication/plugin—Tree
through image analysis
Crown—that estimates several parameters and metrics
techniques (image segmentation and classification) and fractal analysis. on tree crown through image analysis
techniques (image segmentation
The innovation of this work and
is classification)
highlighted by: and(i)fractal analysis. analysis—there is not a GIS
the literature
The innovation of this work is highlighted by: (i)
application available that provides specifically the metrics regarding treethe literature analysis—there
crown; (ii) theisdeveloping
not a GIS
application available that provides specifically the metrics regarding tree
and implementation of the methodology in a new plugin, so the users that are not familiar with thecrown; (ii) the developing
and implementation
algorithms presents in ofQGIS
the methodology
software can,ineasily,
a newestimate
plugin, the so the users thatwith
parameters are anot familiar
simple tool;with
(iii) the
the
algorithms presents in QGIS software can, easily, estimate the parameters
GIS application was developed in Python language, which is a free programming language, so that itwith a simple tool; (iii) the
GIS application
allows was developed
for the modification in Python language,
or adaptation by any user which is a free
according toprogramming
specific rules; language, so thatthe
(iv) we respect it
allows
freedoms for the
of modification
open sourceorsoftwareadaptation by any
using QGISuser software
accordingwhich to specific rules; the
provide (iv) we respect the
possibility of
freedoms of open source software using QGIS software which provide the possibility
incorporating algorithms and functions from others software; and, (v) the fractal analysis was only of incorporating
algorithms
feasible with andanfunctions
automaticfrom others software; and, (v) the fractal analysis was only feasible with an
implementation.
automatic implementation.
2. Study case and Data Acquisition
2. Study case and Data Acquisition
2.1. Study Case
2.1. Study Case
Three study cases were considered: an orthomosaic from an olive crop acquired with an UAV
Three study cases were considered: an orthomosaic from an olive crop acquired with an UAV
imagery with 2.75 ha (Mirandela, Bragança, Portugal in August 2016,◦ 41°37′18.895′′ N; 7°10′20.182′′
imagery with 2.75 ha (Mirandela, Bragança, Portugal in August 2016, 41 370 18.89500 N; 7◦ 100 20.18200 W),
W), another from a eucalypt crop with 0.77 ha (Chamusca, Santarém, Portugal in December 2016,
another from a eucalypt crop with 0.77 ha (Chamusca, Santarém, Portugal in December 2016,
39°20′39.686′′ N; 8°23′28.989′′ W) and the last one considered was from a vineyard crop with 1.57 ha
39◦ 200 39.68600 N; 8◦ 230 28.98900 W) and the last one considered was from a vineyard crop with 1.57 ha
(Corvo, Fafe, Portugal in May 2017, 41°31′03.124′′ N; 8°13′00.997′′ W). In Figure 1 are presented the
(Corvo, Fafe, Portugal in May 2017, 41◦ 310 03.12400 N; 8◦ 130 00.99700 W). In Figure 1 are presented the
three orthomosaics that are considered in this work.
three orthomosaics that are considered in this work.

(a)

Figure 1. Cont.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 4 of 20
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 4 of 20

(b)

(c)
Figure 1. Study cases: (a) olive crop; (b) eucalyptus crop; and, (c) vineyard crop.
Figure 1. Study cases: (a) olive crop; (b) eucalyptus crop; and, (c) vineyard crop.

2.2. Data Acquisition

2.2. Data Acquisition
The image corresponding to each study case was obtained with a rotary-wing UAV (3DR Solo)
The image corresponding to each study case was obtained with a rotary-wing UAV (3DR Solo)
with a MicaSense RedEdge camera which is composed by five bands: Red, Green, Blue, Red edge, and
with a MicaSense RedEdge camera which is composed by five bands: Red, Green, Blue, Red edge,
Near-Infrared (NIR). The coordinate reference system adopted was Universal Transverse Mercator
and Near-Infrared (NIR). The coordinate reference system adopted was Universal Transverse Mercator
(UTM) Zone 29 North—World Geodetic System 1984 (WGS84; EPSG:32629) for olive crop image
(UTM) Zone 29 North—World Geodetic System 1984 (WGS84; EPSG:32629) for olive crop image (with
(with a spatial resolution of 2.8 cm), and European Terrestrial Reference System 1989 (ETRS89)
a spatial resolution of 2.8 cm), and European Terrestrial Reference System 1989 (ETRS89) Portugal
Portugal Transverse Mercator 2006 (PTTM06; EPSG:3763) for eucalypt (with a spatial resolution of
Transverse Mercator 2006 (PTTM06; EPSG:3763) for eucalypt (with a spatial resolution of 3.6 cm) and
3.6 cm) and vineyard images (with a spatial resolution of 5.2 cm). It was performed one single
vineyard images (with a spatial resolution of 5.2 cm). It was performed one single cloud-free flight to
cloud-free flight to each study case. The flight height to olive crop was 50 m, for eucalypt was 120 m
and for vineyard was 80 m. After, for each image, the NDVI was computed using Red and NIR
bands [36].
ISPRS Int. J. Geo-Inf. 2018, 7, 109 5 of 20

each study
ISPRS case. 2018,
Int. J. Geo-Inf. The flight height
7, x FOR to olive
PEER REVIEW crop was 50 m, for eucalypt was 120 m and for vineyard5 of was
20
80 m. After, for each image, the NDVI was computed using Red and NIR bands [36].
3. Methodology
3. Methodology
3.1. GIS Software, Python Libraries and Tree Crown Application
3.1. GIS Software, Python Libraries and Tree Crown Application
The application was created and implemented in QGIS software (version 2.18 Las Palmas), an
openThesourceapplication
software, was created
which and implemented
consider the Stallman in four
QGISfreedoms,
software (version
namely 2.18 Las Palmas),toanstudy
the possibility open
source software, which consider the Stallman four freedoms, namely
and modify the code, to distribute the program with different modified versions [32,33]. QGIS the possibility to study and
modifyfor
allows thethe
code, to distribute
development ofthe
new program
pluginswith different
in Python 2.7modified
programming versions [32,33].[37]
language QGIS andallows
providefor
the development of new plugins in Python 2.7 programming language [37]
online support through forums, tutorials, and documentation [33]. It allows for vector and raster and provide online support
through
files forums, tutorials,
manipulation, and documentation
visualization, [33]. It allows
analysis, and acquisition. Tofor vector
create andplugin,
a new raster files manipulation,
different libraries
visualization,
and/or analysis,
Application and acquisition.
Programming To create
Interface (API) aare
new plugin, different
commonly used, such libraries and/or Application
as Geographic Resources
Programming
Analysis SupportInterface
System (API) are commonly
(GRASS), QGIS API, used,andsuch as Geographic
PyQt4 API [30,38,39]. Resources Analysis
Also, from QGISSupport
2.0, a
System (GRASS), QGIS API, and PyQt4 API [30,38,39]. Also, from
Processing Toolbox, a framework containing several algorithms of standalone GIS software, such QGIS 2.0, a Processing Toolbox,
as
a framework
GRASS containing
7.2.2 and OTB 5.0.0,several
wasalgorithms
integrated of standalone
[31]. The OTBGIS software,
software is ansuch
open assource
GRASSproject
7.2.2 and
for
OTB 5.0.0,
remote was integrated
sensing and it is [31]. The OTB
composed bysoftware
several is an open source
algorithms, whichproject
allows forfor
remote sensinghigh
processing and
it is composed
resolution, by severaland
multispectral, algorithms,
RADAR images,which allows for processing
and includes several high
imageresolution,
processingmultispectral,
procedures,
and RADAR images, and includes several image processing procedures,
such as ortho-rectification, pan sharpening, classification algorithms, image manipulation, and such as ortho-rectification,
pan sharpening,
segmentation classification algorithms, image manipulation, and segmentation [31].
[31].
This plugin
This plugin was
was concretized
concretized as as aa button
button in
in QGIS
QGIS mainmain menu.
menu. The The plugin
plugin graphic
graphic interface
interface waswas
developed through Qt Designer package where it is possible to create and
developed through Qt Designer package where it is possible to create and personalize the widgets personalize the widgets such
as combo
such as comboboxes, pushpush
boxes, buttons, labels,
buttons, among
labels, othersothers
among [40]. [40].
The main
The main graphic
graphic interface
interface is is aa window
window composed
composed by by four
four fields,
fields, two
two inputs
inputs andand two
two outputs
outputs
(Figure 2). This fields were defined as push buttons and line edits. The functions
(Figure 2). This fields were defined as push buttons and line edits. The functions related to each push related to each push
button are,
button are,respectively,
respectively, inputfile, inputfile2,
inputfile, output,output,
inputfile2, and output2.
and The main The
output2. procedure
main was implemented
procedure was
in the function run.
implemented in the function run. In this function the image details, such as extension, numberand
In this function the image details, such as extension, number of bands of
multispectral
bands bands, werebands,
and multispectral accessed wereusing functions
accessed from
using QgsRasterLayer()
functions class [38]. In order
from QgsRasterLayer() classto[38].
access
In
to Processing
order to accessToolbox
to Processing the Processing
algorithms,Toolbox module
algorithms, thewas imported
Processing and thewas
module function runAlgorithm()
imported and the
was used. The experimental procedure performed allowed to select the
function runAlgorithm() was used. The experimental procedure performed allowed to select the most most accurate methodology
to be implemented
accurate methodology in thetoplugin. The methodology
be implemented in the implementation is described implementation
plugin. The methodology in the Section 3.2.1, is
Section 3.2.2,
described andSections
in the Section 3.2.1,
3.2.3. 3.2.2, and 3.2.3.

Figure 2. Plugin graphic interface.

Figure 2. Plugin graphic interface.

3.2. Procedures Tested

3.2. Procedures Tested
To define the methodology to be implemented in the plugin, different tools were tested and
To define the methodology to be implemented in the plugin, different tools were tested and
compared. In the end, the more accurate methodology was chosen. Based on the literature consulted
compared. In the end, the more accurate methodology was chosen. Based on the literature consulted
and in the tools existing in QGIS, the OBIA classification approach was chosen. The Sections 3.2.1
and in the tools existing in QGIS, the OBIA classification approach was chosen. The Sections 3.2.1
and 3.2.2 describe the tools that are available in QGIS to perform OBIA segmentation and
and 3.2.2 describe the tools that are available in QGIS to perform OBIA segmentation and classification.
classification.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 6 of 20

3.2.1. OBIA Segmentation

For the segmentation process, two algorithms from GRASS and OTB, respectively, i.segment and
Watershed segmentation were tested. The first approach performs the segmentation based in the region
growing algorithm and the second uses the watershed algorithm [41]. The i.segment sequentially
examines all of the segments in the raster file and calculates a similarity measure between the current
segment and each of its neighbors, according to a given distance [30]. This algorithm was computed
considering two parameters: the thresholding, which represents the similarity between 0 and 1;
and, the object minimum size (minsize), which corresponds to the minimum size value that each
object can have. In thresholding, a higher value aggregates objects with similar properties. Through
a neighbor relation, the pixels that presented similar spectral and spatial characteristics to the seed
points were aggregated, dividing the image into objects.
When considering the region growing method, different values of thresholding and minsize were
tested in order to optimize the procedure. The thresholding was tested with values ranging from 0.05
to 0.95 (unitless), in which a loop cycle was created, and where the first image created (with 0.05
of thresholding) contained the seed points and the following images were obtained recursively from
the first one; was considered an increment in thresholding value of 0.05 [41]. Also, the minsize value
was randomly tested when considering the values 10, 100, and 200. In this segmentation process,
the tool accepts the multispectral image with separated bands and with non-separated bands as inputs.
The other parameters were defined by default.
The Watershed segmentation algorithm from OTB was tested with the default values that were
provided by QGIS [33].
In Section 4.1.1, the results that were obtained with the two algorithms tested will be presented.
The more robust algorithm implemented was the i.segment. In the Tree Crown application, to split
the multispectral image by bands, the split image, algorithm from OTB, was employed. The first seed
image was processed with 0.05 of thresholding. Then, a for loop was created to use i.segment algorithm
with an increment of 0.05 and the thresholding values varied between 0.1 and 0.65. The final image was
obtained considered a thresholding of 0.65.

3.2.2. OBIA Classification

In image classification, two algorithms, from OTB, were tested: supervised classification using the
Image classification algorithm, through the Support Vector Machine (SVM) classifier, and unsupervised
classification through the Unsupervised kmeans image classification algorithm [42,43].
In the supervised classification, the mean and standard deviation values of each band were
obtained from a XML file generated from Compute images second order statistics algorithm from OTB.
The next step consisted in a shapefile creation with two training areas: the tree crown class (assigning as
1) and remaining objects (assigning as 0). The SVM method was used through TrainImagesClassifier (svm),
algorithm from OTB, to perform classifier training from pairs of images [30]. Finally, the classification
was performed using the Image Classification algorithm with the multispectral image and the NDVI
image as inputs, and the XML file containing the mean and standard deviation values, the segmented
image generated before and the SVM resulting file [42]. The classification accuracy was obtained
through r.kappa algorithm from GRASS GIS 7, which compute the error matrix of the classification
result by crossing the classified map layer with respect to reference map layer (training areas). The error
matrix and the kappa value were returned.
In the unsupervised classification (k-means algorithm), the clusters (center points of each class)
are defined and the pixels are classified based on their distance to the cluster. When considering
that the NIR or RED bands from low-cost UAV sensors are usually less radiometrically robust than
RGB bands, the multispectral image, the NDVI image, and the RGB true colour combination were
used as inputs into the classification process. The training areas were defined with the Training set
size parameter, assuming the value of 100,000 pixels [44]. The validation mask corresponded to the
segmented image and the other parameters were defined by default.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 7 of 20

The classification algorithm chosen was the unsupervised classification algorithm implemented
through the Unsupervised kmeans image classification from OTB. In this algorithm the input can be the
RGB combination, the multispectral or the NDVI image, and the segmented image as validation mask.
The training set size parameter was obtained when multiplying the total number of pixels by 0.00705 to
obtain the value 100,000 and the remaining parameters were defined by default. Then, the maximum
value of the resulted image was extracted and associated to the tree crown using functions from
QgsRasterLayer(). To distinguish the tree crown from the other objects, the Rastercalculator, algorithm
from GDAL, was used, assigning the value 1 to tree crown. The image obtained was converted to
shapefile through Polygonize algorithm, from GDAL, which was interpreted using functions from
QgsVectorLayer() class. To restrict the objects only to tree crown, were not considered: (i) objects with
an area lower than 1 m2 or higher than 50 m2 ; (ii) elements with the class value different from 1;
and, (iii) objects with the ratio between height and width of the polygon bounding box, which is
considered as a regular polygon where the extension values (x minimum, x maximum, y minimum,
and y maximum) correspond to the extension of an irregular polygon. In the case of tree crown,
it is expected a quadrangular polygon, so this ratio should be closest to 1. However, in this plugin,
a tolerance of 0.55 was considered. The ratio was obtained implementing a for loop to extract the area,
height and width. In the end, the features, which satisfy at least one of the above criteria, were excluded
from the shapefile.

3.2.3. Fractal Analysis

The fractal analysis was applied in order to obtain the tree crown parameters and metrics in
shapefile format. To perform this, the resulting image from OBIA need to be converted to binary
image, where 1 corresponded to the tree crown and 0 to the remaining objects. The resulted image was
then converted to shapefile through Polygonize, algorithm from Geospatial Data Abstraction Library
(GDAL) [45]. The area was estimated, and the tree crown objects were identified when considering an
area between 1 m2 and 50 m2 .
In the GIS application, the fractal analysis was generated and included in a text file composed
by the following metrics: area, perimeter, centroid of each tree, total number of trees, total area of
tree crown, and total number of missing trees. The information was saved in table format with the
following headers: Id, Area, Perimeter, x (centroid) and y (centroid).
The code that was implemented used the write() function to insert each parameter. Through
functions from QgsVectorLayer() class, the Id attribute was added to the shapefile attribute table. Two
for loops were created to count the number of trees, sum the total area of tree crown, and to extract the
trees metrics. To obtain the distance between the trees in each row, it was necessary to estimate the
rhumb value. Because the trees positions are not regular, the rhumb between each tree was calculated
using the azimuth function from QgsPoint() class. It was considered the positive rhumb. This value
was calculated considered all of the centroids. From the list obtained, the mode value of the rhumb
was extracted. Then, two for loops were defined to calculate the rhumb of one point related with the
other points to find the distances in the existent rows, with a tolerance of 10. The minimum distance,
which corresponds to the distance between two consecutive points in the same row, was extracted and
listed. These values allowed for detecting the missing trees. For that, the mean distance was obtained
and, if a distance value was higher than the mean value, with a specific tolerance of 1.5, a missing tree
was detected. Finally, the total number of trees, the total area, the mean distance between trees and the
number of missing trees was recorded in the text file.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 8 of 20

4. Results and Discussion

4.1. Experimental Procedure

In order to define the methods to be implemented in the new plugin, some tests were performed
in the segmentation and classification stages when considering the olive crop. This (Section 4) section
presents
ISPRS Int. the results
J. Geo-Inf. obtained
2018, in these
7, x FOR PEER tests.
REVIEW 8 of 20

4.1.1.
4.1.1. OBIA
OBIA Segmentation
Segmentation
Figure
Figure 33 presents
presents the
the segmentation
segmentation result
result obtained with i.segment
obtained with i.segment algorithm
algorithm (Figure
(Figure 3a)
3a) and
and
considered the Watershed algorithm (Figure 3b), both performed when considering the default values.
considered the Watershed algorithm (Figure 3b), both performed when considering the default values.

(a)

(b)
Figure 3. (a) Segmentation with i.segment algorithm; (b) segmentation with Watershed algorithm.
Figure 3. (a) Segmentation with i.segment algorithm; (b) segmentation with Watershed algorithm.

Several tests were computed using different values of thresholding and minsize. Table 1 presents
the results obtained regarding the tests that were performed using different values of thresholding
(0.65 or 0.95) and minsize (10, 100, or 200) for segmentation process and different inputs
(multispectral image, RGB image, or NDVI) for classification process with i.segment algorithm. Also,
the algorithm can be applied with separated bands or joined bands. So, the tests presented
considered these two cases.
Table 1 presents the results of several tests performed considered the multispectral image,
ISPRS Int. J. Geo-Inf. 2018, 7, 109 9 of 20

Several tests were computed using different values of thresholding and minsize. Table 1 presents
the results obtained regarding the tests that were performed using different values of thresholding (0.65
or 0.95) and minsize (10, 100, or 200) for segmentation process and different inputs (multispectral image,
RGB image, or NDVI) for classification process with i.segment algorithm. Also, the algorithm can be
applied with separated bands or joined bands. So, the tests presented considered these two cases.
ISPRSTable 1 presents
Int. J. Geo-Inf. 2018, the results
7, x FOR PEER ofREVIEW
several tests performed considered the multispectral image, NDVI 9 of 20
image and RGB true colour image as inputs in unsupervised classification. For this study case, the results
accuracy considering
obtained when compared with NDVI image
the multispectral image (asand
input). For this
the RGB imagereason, the NDVI
presented worstimage was used
accuracy when as
input.
compared with NDVI image (as input). For this reason, the NDVI image was used as input.
Thesegmentation
The segmentation accuracy
accuracy waswas evaluated
evaluated basedbased on “Goodness
on “Goodness for fit” for fit” statistical
statistical model, asmodel,
provided as
provided by i.segment algorithm, which allows for describing how
by i.segment algorithm, which allows for describing how well it fits a set of observations, summarizingwell it fits a set of observations,
summarizing
the discrepancythe discrepancy
between observed between observed
and expected and Inexpected
values. the GRASS values.
library In (i.segment
the GRASS library
algorithm),
(i.segment
this parameter algorithm),
is basedthis on parameter
the distanceisof based on the
the pixel to distance
the objectofthatthe itpixel to the
belongs to.object
A valuethatofit1belongs
means
to. A value of 1 means identical values, a perfect fit, and a value
identical values, a perfect fit, and a value of 0 means maximum possible distance, worst possible fit. of 0 means maximum possible
distance,
Given worst
the lack of possible
field datafit. to Given
validatethethelack of field
results data this
obtained, to validate
model was the used.
results Inobtained, this model
order to estimate the
was used. In order to estimate the percentage of pixels with specific
percentage of pixels with specific values, r.report (algorithm from GRASS) was computed. The model was values, r.report (algorithm from
GRASS)towas
applied computed. The
the segmentation model
tests was performed
that were applied tointhe segmentation
Table 1. Based on tests that were
the report that wasperformed
returned in it
Table
was 1. Basedthat:
concluded on the report
in case (a) that
all ofwas returned
the image it was
pixels wereconcluded
above 0.96that: and in case of
47.33% (a)them
all ofwere
the above
image
pixels
0.99; in were
case (b) above
all of0.96 and 47.33%
the pixels of them
were above 0.94were above of
and 69.82% 0.99;
the in casewere
pixels (b) all of the
above pixels
0.99; in thewere
case above
(c) all
0.94 and 69.82% of the pixels were above 0.99; in the case (c) all
the pixels were above 0.95 and 44.85% of them were above 0.99; in case (d) all the pixels were above the pixels were above 0.95 0.93
and
44.85%
and 73.79%of themof themwere wereaboveabove 0.99; in case
0.99; in case(d)(e)
allall
thethepixels
pixelswere
wereabove
above0.930.95andand73.79%
44.85% of them
were; in were
case
(f) all the pixels were above 0.97 and 64.94% of them were above 0.99. The values obtained provedwere
above 0.99; in case (e) all the pixels were above 0.95 and 44.85% were; in case (f) all the pixels that
above
the 0.97 and 64.94%
segmentation of them algorithm
using i.segment were above was0.99. The values
satisfactory. obtained
Figure proved
4 presents that the of
an example segmentation
“Goodness
using
for fit”i.segment
map. algorithm was satisfactory. Figure 4 presents an example of “Goodness for fit” map.

Figure
Figure 4. “Goodnessforforfit”fit”
4. “Goodness map
map considering
considering a thresholding
a thresholding of 0.65
of 0.65 and and a minsize
a minsize of with
of 100, 100,
with separated
separated bands.
bands.

The i.segment
The i.segment presented
presented consistent
consistent results
results (Figure
(Figure 3a),
3a), whereas
whereas the the results
results that
that were
were obtained
obtained
withthe
with theOTB
OTBalgorithm
algorithmwere
were not
not satisfactory
satisfactory (Figure
(Figure 3b).3b).
ForFor
thisthis reason,
reason, the i.segment
the i.segment algorithm
algorithm was
was chosen to be implemented in the plugin. The tests performed with different
chosen to be implemented in the plugin. The tests performed with different thresholding and minsize thresholding and
minsizeallowed
values values to allowed to more
define the defineadequate
the more adequate methodology
methodology (Table 1). The (Table 1). The
thresholding thresholding
parameter was
parameter was defined as 0.65 as it presented a significant agglomeration of the pixels
defined as 0.65 as it presented a significant agglomeration of the pixels in the objects with smaller in the objects
with smaller
values. values.value
The minsize The minsize valueas
was defined was
100defined as the
pixels as 100results
pixels as the results
obtained wereobtained were
better with better
smaller
with smaller
values and veryvalues andwith
similar veryhigher
similarvalues
with higher
(Table 1).values (Table 1).
 ISPRS
ISPRS Int.
Int. J. Geo-Inf.
J.J. Geo-Inf. 2018,
J.2018, 7,x2018,
7,7,xx7, 109 1020of 20 10 of 20
ISPRS Int.
ISPRS
ISPRS J. ISPRS
Int. Int.Int.
J.Geo-Inf.
Int.ISPRS 2018,
J.2018,
Geo-Inf.
Geo-Inf.
Geo-Inf.
Geo-Inf.
2018, FORx FOR
2018,
7,
FOR PEER
FOR xPEER
7, 7,
PEER FOR
PEER REVIEW
xREVIEW
FOR PEER
PEER REVIEW
REVIEW
REVIEW
REVIEW 10
10of
1020
of 10 of 20
of 20

Table
Table 1.
1.1.TableTests
1.Table
Tests performed
1.Tests
performed for for
Testsperformed
performed segmentation
segmentation and and
forforsegmentation
segmentation classification
classification processes
andclassification
classification
processes considering
processes
considering different
considering
different values
different
values of of thresholding
values and
ofofthresholding
thresholding minsize
thresholding
and minsizeand
for for i.segment
minsize algorithm,
forfori.segment
i.segment i.segment
algorithm, in
in the
algorithm,
in theininthe
the
Table
Table
Table 1. Tests
1.
Tests
Tests performed
performed
performed forfor
for segmentation
segmentation
segmentation andand
and classification
and
classification
classification processes
processes
processes considering
processes
considering
considering different
considering
different
different values
different
values
values of
ofvalues
of thresholding
thresholding
thresholding andand
and minsize
and
minsize
minsize forfor
minsize
for i.segment
i.segment
i.segment in the
algorithm,
algorithm,
algorithm, in
algorithm,
the the
segmentation process
segmentation and
processconsidering
and different
considering inputs
different (multispectral
inputs image,
(multispectral Red-Green-Blue
image, (RGB)
Red-Green-Blue image,
(RGB) or Normalized
image,oror Difference
Normalized Vegetation
Difference Index
Vegetation (NDVI))
Index in
segmentation
segmentation
segmentation process
segmentation and
process
process
process and considering
and
process
and considering
and
considering
considering different inputs
different
considering
different
different (multispectral
inputs
different
inputs
inputs (multispectral
inputs image,
(multispectral
(multispectral image,
(multispectral Red-Green-Blue
image,
image, Red-Green-Blue
image, (RGB)
(RGB)
Red-Green-Blue
Red-Green-Blue
Red-Green-Blue image,
(RGB)
(RGB) image,
(RGB) or
image,
image, or
or Normalized
Normalized
image,
or
Normalized Difference
Difference
Normalized
Normalized Vegetation
Vegetation
Difference
Difference
Difference Index
Vegetation
Vegetation
Vegetation (NDVI))
Index
Index Index
(NDVI))
(NDVI)) in(NDVI))
in the
(NDVI))
in inin
the the classification
the
classification
classification
the process.
classification
process.
process.
classification process.
process.
thethe classification
classification process.
process.
Unsupervised
Unsupervised Classification
Unsupervised
Unsupervised
Unsupervised with
Classification
Classification with
Classification
Classification with with Unsupervised
withUnsupervised
Unsupervised Classification
Unsupervised
Classification
Unsupervised withwith
Classification
Classification
Classification with
with with Unsupervised
Unsupervised Classification
Unsupervised
Classification
Unsupervised
Unsupervised withwith
Classification
Classification
Classification with
with with
Unsupervised Classification
Unsupervised with
Classification with Unsupervised Classification
Unsupervised Classificationwith
with Unsupervised
Unsupervised Classification
Classification with
with
Segmentation
Segmentation
Segmentation
Segmentation
Segmentation Segmented
Segmentation Segmented Image
Segmented
Image
Segmented
Segmented and
Image
Segmented and
Image
Image
and
Image Multispectral
and
Multispectral
andMultispectral
Multispectral
Multispectral
and Segmented
Multispectral
Segmented Image
Segmented
Image
Segmented
Segmented
Segmented and
Image
Image and
Image
NDVI
Image
and
and NDVI
and
andImage
NDVI
NDVI Image
NDVI
NDVI as
Image
Image as
Image
Image Segmented
asas
asasSegmented Image
Segmented
Image
Segmented
Segmented
Segmented Image
Image and
andand
Image
RGB
Image
and
RGB RGB
RGBandImage
Image
and RGB
RGBasas
Image
Image as
Image
as asas
Image
Segmentation Segmented Image and Multispectral Segmented Image and NDVI Image as Segmented Image and RGB Image as
Image
Image as
Image as
Image
Image
as Inputs
Image
Inputs asas
as
Inputs Inputs
Inputs
Inputs Inputs
Inputs
Inputs Inputs
Inputs
Inputs Inputs
Inputs
Inputs Inputs
Inputs
Inputs
Image as Inputs Inputs Inputs

(a) (a) Segmentation:

(a) Thresholding
Segmentation: === 0.65,
Thresholding minsize
=minsize
0.65, ==100,
100,
==minsize =separated
= 100, bands
separated bands
(a)Segmentation:
(a) (a)(a)
Segmentation:
Segmentation: Thresholding
Segmentation: ==0.65,
Thresholding
Segmentation: Thresholding
Thresholding
Thresholding 0.65, minsize
0.65,
= minsize
0.65, 0.65,
minsize 100,
=100,
minsize separated
100,separated
100, bands
bands
separated
separated
separated bands
bands
bands

(b) (b) Segmentation:

(b) (b)
Segmentation: Thresholding
Segmentation:
Segmentation:
(b) Segmentation: == 0.65,
Thresholding
Thresholding ==0.65,
Thresholding
Thresholding minsize
=minsize
0.65,
minsize
0.65,
= minsize
0.65, ==100,
100,
=100,
minsize100,
==minsize =joined
joined= 100,
joined bands
joined
bands
100,bands
joined bands
bandsbands
(b)(b) Segmentation:
Segmentation: Thresholding
Thresholding 0.65,
0.65, minsize 100, joined
joined bands
 ISPRS Int. J. Geo-Inf. 2018, 7, 109 11 of 20

Table 1. Cont.

Unsupervised Classification with Unsupervised Classification with Unsupervised Classification with

ISPRS
ISPRS Int.
Int.ISPRS
J. J.
Geo-Inf.
ISPRS Segmentation
J. Geo-Inf.
Int. J.2018,
Int. J. 2018,
7, x 7,
Geo-Inf. FOR
Geo-Inf. x2018,
x FOR
2018, 7, xxPEER
PEER
7, FOR REVIEW
PEER
xREVIEW
FOR Segmented
REVIEW Image and Multispectral
REVIEW
PEER Segmented Image and NDVI Image as Segmented Image and RGB Image11as of1120of 20
11 of
11 20
of 20
ISPRS
ISPRS J.ISPRS
Int.
Int.ISPRS Int.
Geo-Inf.
Int.
Geo-Inf. J. J.
2018, Geo-Inf.
2018,
7, x 7,
Geo-Inf. 2018,
2018,
FOR 7,7,
FOR
PEER xFOR
FOR
PEER PEER
REVIEW
PEER
REVIEW REVIEW
REVIEW 1111
11 of1120of 20 ofof2020
Image as Inputs Inputs Inputs

(c) (c)
(c) Segmentation:
(c)(c)
(c)Segmentation:
Segmentation:
Segmentation:
(c)
Segmentation: Thresholding
Thresholding
Thresholding
Thresholding
Segmentation:
Segmentation: 0.95,
== 0.95, == =minsize
== 0.95,
Thresholding
Thresholding
Thresholding 0.95, 0.95,
minsize
0.95,minsize
minsize
0.95,
minsize = =100,
100,
==minsize
100,
minsize
100, == =separated
100,
100,separated
separated
separated
100,
separated bands
bands
bands
separated
separated
bands bands
bands
bands

(d)(d) Segmentation:
(d) Thresholding
Segmentation: Thresholding
(d) (d) Segmentation:
Segmentation:
(d)
(d) Thresholding
Thresholding
Segmentation:
Segmentation: == 0.65,
0.65,
== 0.65,
Thresholding
Thresholding minsize
==minsize
0.65,
minsize
=0.65, =10,10,
minsize
=10,
==minsize
0.65,
minsize ==joined
joined
joined 10, joined
bands
bands
=10,
10, bands
bands
joined
joined bands
bands
Segmentation: Thresholding 0.65, minsize 10, joined bands

(e) (e)
(e) Segmentation:
(e)
(e)Segmentation:
Segmentation:
(e) Thresholding
Segmentation:
Segmentation:
Segmentation: Thresholding
Thresholding = 0.95,
== 0.95,
Thresholding
Thresholding
Thresholding 0.95, == =0.95,
minsize
0.95,minsize
minsize
0.95,
minsize = 10,
==minsize
10,
minsize
10, ==separated
=10,
10,separated
separated
10,
separated bands
bands bands
separated
separated
bands bands
bands
 (c) Segmentation: Thresholding
(c)Segmentation:
Segmentation:
(c) Segmentation:
(c) = 0.95,
Thresholding minsize
Thresholding
= 0.95,
Thresholding 0.95, =minsize
= =minsize
0.95, 100, separated
minsize
= 100, =100, bands
100,separated
separated
=separated bands
bandsbands

ISPRS Int. J. Geo-Inf. 2018, 7, 109 12 of 20

Table 1. Cont.

Unsupervised Classification with Unsupervised Classification with Unsupervised Classification with

Segmentation Segmented Image and Multispectral Segmented Image and NDVI Image as Segmented Image and RGB Image as
(d)Image asSegmentation:
(d) Segmentation:Inputs
Thresholding
(d)Segmentation:
Segmentation:
(d) = 0.65,
Thresholding minsize
Thresholding = =minsize
= 0.65,
Thresholding 0.65,
0.65, Inputs
=minsize
10,
= joined
minsize bands
= =10,
10,joined
10, joined joined bands
bandsbands Inputs

ISPRS
ISPRS Int.
Int.ISPRS J. Geo-Inf.
J. ISPRS
Geo-Inf.
Int.
Int. J. 2018,
2018, 7, x 7,
FOR
Geo-Inf.
J. Geo-Inf. x2018,
FOR
2018,PEERxPEER
7, 7, REVIEW
xREVIEW
FOR
FOR PEER
PEER REVIEW
REVIEW 1220
12 of of 20
1212
of of
2020
(e) Segmentation:
(e) Segmentation:
(e)(e) Thresholding
(e)Segmentation:
Segmentation: = 0.95,
Thresholding
Thresholding
Segmentation: minsize
= 0.95,
Thresholding
Thresholding =minsize
=minsize
0.95,
= minsize
0.95, 0.95, 10, separated
minsize
==10, = =10,
10,separated bands
10,separated
separated
separated bands
bandsbands
bands

(f) Segmentation:
(f) Segmentation: Thresholding
Thresholding
(f)Segmentation:
(f)(f) Segmentation:
Segmentation: ==0.65,
= 0.65,
Thresholding
Thresholding
Thresholding minsize
minsize
= =minsize
0.65, ==200,
200,
= minsize
200,
0.65,minsize
0.65, =separated
separated
=200, bandsbands
bands
200,separated
separated separated
bands bands
ISPRS Int. J. Geo-Inf. 2018, 7, 109 13 of 20
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 12 of 20

4.1.2. 4.1.2.
OBIAOBIA
Classification
Classification
In this process,
In this both
process, classification,
both classification, supervised and
supervised and unsupervised,
unsupervised, obtained
obtained similarsimilar
results,results,
as
as presented inin
presented Figure
Figure5.5.

(a)

(b)
Figure 5. (a) Supervised classification result; and, (b) Unsupervised classification result, for olive
Figure 5. (a) Supervised classification result; and, (b) Unsupervised classification result, for olive crop.
crop.

The input
The input defined
defined in the
in the supervisedclassification
supervised classification (Figure
(Figure5a)5a)was
wasthethe
multispectral image,
multispectral image,
because the NDVI image did not presented valid results. Regarding the classification
because the NDVI image did not presented valid results. Regarding the classification accuracy, accuracy, in
the supervised classification was obtained an overall accuracy of 92.84% and a kappa
in the supervised classification was obtained an overall accuracy of 92.84% and a kappa value of 0.71.value of 0.71.
While in the unsupervised classification the NDVI image presented results more robust (Figure 5b).
While in the unsupervised classification the NDVI image presented results more robust (Figure 5b).
Both classifications used the segmented image with separated bands as input. As both provided
Both classifications used the segmented image with separated bands as input. As both provided very
very similar results, the algorithm that was implemented in the plugin was the unsupervised
similar results, the because
classification algorithm it that
is awas implemented
procedure more in the plugin
objective whenwas the unsupervised
compared classification
to the supervised
because it is a procedure more objective when compared to the supervised classification,
classification, given the fact that the user doesn’t need to define the training areas to classify.given the fact
that the user doesn’t need to define the training areas to classify.
The final result was obtained from a segmentation through i.segment algorithm using the
multispectral image with separated bands, followed by an unsupervised classification through
Unsupervised kmeans image classification when considering the NDVI image as input (Figure 5a).
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 13 of 20
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 13 of 20
The final result was obtained from a segmentation through i.segment algorithm using the
The finalimage
multispectral result with
was separated
obtained from
bands, a segmentation
followed by an through i.segmentclassification
unsupervised algorithm using the
through
multispectral
Unsupervised
ISPRS
image
kmeans
Int. J. Geo-Inf.
with
2018,image
separated bands, followed by an unsupervised classification
7, 109 classification when considering the NDVI image as input (Figure 5a).
through
14 of 20
Unsupervised kmeans image classification when considering the NDVI image as input (Figure 5a).
4.1.3. Fractal Analysis
4.1.3. Fractal
4.1.3. Fractal Analysis
Analysis
The classified image was then converted to binary image using Raster Calculator, resulting in an
image The
Thewith classified
classified 1image
values imageto tree was
was thenand
then
crown converted
converted to
0 to theto binaryimage
binary
remaining image using
using
objects. Raster
Raster
After, Calculator,
Calculator,
this image was resulting
resulting inan
in
converted an
to
image
image with
with values
values
shapefile (Figure 6). 11 to
to tree
tree crown
crown and
and 00 to
to the
the remaining
remaining objects.
objects. After,
After, this
this image
image was
was converted
converted to
to
shapefile (Figure
shapefile (Figure 6). 6).

Figure 6. Shapefile with tree crown well identified.

Figure6.
Figure 6. Shapefile
Shapefilewith
withtree
treecrown
crownwell
wellidentified.
identified.
In the shapefile showed in Figure 6, some objects were not correctly classified. Based on in situ
In the objects
knowledge, shapefile showed
with an areain Figure 6, some objects were notwith
correctly classified. Based on in situ
In the shapefile showed in different
Figure 6,from
some (1–50) m2 were
objects or objects an atypical
not correctly shape
classified. for the
Based on general
in situ
knowledge,
tree objectsbe
crown should with an areaThe
removed. different from objects
remaining (1–50) m
areor
2
2objects
considered with an atypical (Figure
as tree shape for the general
knowledge, objects with an area different from (1–50) m or objects withcrown 5b).
an atypical shape for the
tree crown should be removed. The remaining objects are considered as tree crown (Figure 5b).
general tree crown should be removed. The remaining objects are considered as tree crown (Figure 5b).
4.2. GIS Open Source Application
4.2. GIS
4.2. GIS Open
Open Source
SourceApplication
Application
To test the plugin created, the study case presented in the experimental procedure was
Totest
considered.
To testthethe
The plugin
result
plugin created,
obtained
created, the the
the was
study study
case case presented
shapefile
presentedwith in the
theexperimental
in the tree crownexperimental
and procedure
the text
procedure wasfile with was
the
considered.
considered.
metrics The
estimated. result
Figureobtained
7a,b was
present the
the shapefile
resulting with
shapefilethe tree
from crown
plugin and
and the
the text file
shapefile
The result obtained was the shapefile with the tree crown and the text file with the metrics estimated. with the
obtained
metrics
manually,
Figure estimated. Figure
7a,b respectively. 7a,b present the resulting shapefile from plugin and the shapefile
present the resulting shapefile from plugin and the shapefile obtained manually, respectively. obtained
manually, respectively.

(a) (b)
(a) (b)
Figure 7. (a) Shapefile obtained with the plugin; and, (b) shapefile obtained with the manual process.
Figure 7. (a) Shapefile obtained with the plugin; and, (b) shapefile obtained with the manual process.
Figure 7. (a) Shapefile obtained with the plugin; and, (b) shapefile obtained with the manual process.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 15 of 20
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 14 of 20

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 14 of 20

The
The code
code implemented
implemented allowed
allowed forfor removing
removing the the objects
objects with
with aa bounding
bounding box box higher
higher than
than the
the
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 14 of 20
ratio defined.
ratioThe
defined. The
The metrics
metrics estimated
estimated presented
presented a
a total
total count
count of
of 178
178 trees
trees and
and a
a total
total area
area of
of 1401.3
1401.3 m22..
m
code implemented allowed for removing the objects with a bounding box higher than the
The
The mean
mean distance
distance between
between the
the trees was 7.8
treespresented
was 7.8 m
maand
and itit were
were detected
detected 23 missing trees.
The code
ratio defined. implemented
The metrics allowed
estimated for removing the
total objects
count of with
178 a23
trees missing
bounding
and trees.
box
a total higher
area than m
of 1401.3 the2.
In
In the
theplugin
plugintesting phase
testing phasesome limitations
some were
limitations founded.
were For
founded. instance,
For two
instance,trees
twowere wrongly
trees were
ratiomean
The defined. The metrics
distance between estimated
the treespresented
was 7.8 maand totalit count of 178 trees
were detected and a total
23 missing area of 1401.3 m2.
trees.
removed
wrongly and some and
removed of them
some were
of not totally
them were separated.
not totally The proximity
separated. The between
proximitythem was notthem
between detected
was
The mean
In thedistance
plugin between the trees
testing phase waslimitations
some 7.8 m and itwere were founded.
detected 23 missing
For instance,trees.
two trees were
by the algorithm
not detected (Figure
by theand 8).
algorithm
In the
wrongly plugin
removed testing of(Figure
somephase themsome8). limitations
were not totallywere founded.
separated. The For instance,
proximity two trees
between themwere
was
wrongly
not removed
detected by theand some of
algorithm them 8).
(Figure were not totally separated. The proximity between them was
not detected by the algorithm (Figure 8).

Figure 8. Incorrect objects identification (closer objects).

Figure 8. Incorrect objects identification (closer objects).
Figure 8. Incorrect objects identification (closer objects).
Also, some vegetation species
Figure with similar
8. Incorrect spectral characteristics
objects identification with olives were wrongly
(closer objects).
Also, as
identified
Also, some vegetation
olivevegetation
some species
crop (Figure with similar
9). with
species similar spectral
spectral characteristics
characteristics with
with olives
olives were
were wrongly
wrongly
identified as
Also, as
identified olivevegetation
some
olive crop (Figure
crop (Figure 9).
species
9). with similar spectral characteristics with olives were wrongly
identified as olive crop (Figure 9).

Figure 9. Incorrect objects identification (other type of vegetation).

Figure 9. Incorrect objects identification (other type of vegetation).
Figure 9. Incorrect
The detected number objects identification
seems to be(other type of vegetation).
Figure 9.ofIncorrect
missingobjects
trees identification higher
(other than
type the expected. This fact can be
of vegetation).
related
Thewith the roads
detected that cross
number this area,
of missing treessoseems
it wastoidentified
be higher as than
a missing tree (Figure
the expected. 10).
This fact can be
The
The
related detected
detected
with number
number
the roads of missing
of missing
that cross trees
trees
this area, seems
so seems to
it was to be higher
be higher
identified than
as athan the expected.
the expected.
missing This
This
tree (Figure fact can
10).fact can be
be
related with the roads that cross this area, so it was identified as a missing tree (Figure
related with the roads that cross this area, so it was identified as a missing tree (Figure 10). 10).

Figure 10. Representation of the road that cross the study area, which can influence the results.
Figure 10. Representation of the road that cross the study area, which can influence the results.
Two different
Figure crops were also used to test thethe
plugin, eucalyptus, andinfluence
vineyards. In the case of
Figure 10.
10. Representation
Representation of
of the
the road
road that
that cross
cross the study area, which
study area, which can
can the results.
influence the results.
eucalyptus, the resulted shapefile was not satisfactory, because most of
Two different crops were also used to test the plugin, eucalyptus, and vineyards. In the the trees were
casenot
of
recognized
eucalyptus, and
Two differentsome shadows
crops were
the resulted were
shapefile identified
also used
wastonot as trees (Figure
test satisfactory, 11).
the plugin, eucalyptus,
because most and vineyards.
of the treesIn the
werecasenot
of
In this
eucalyptus, study
the case,
resultedsome polygons
shapefile were
was notnot completely
satisfactory,
recognized and some shadows were identified as trees (Figure 11). separated,
because because
most of the
the eucalyptus
trees were trees
not
are very
recognizedclose
In this andtogether.
studysome Also,polygons
case,shadows
some the
wereshade of the
identified
were notastrees
treesis(Figure
different,
completely 11).so it was
separated, not the
because possible to obtain
eucalyptus treesa
In this
are very study
close case, some
together. Also,polygons
the shadewere nottrees
of the completely separated,
is different, because
so it was the eucalyptus
not possible trees
to obtain a
are very close together. Also, the shade of the trees is different, so it was not possible to obtain a
ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 15 of 20

homogeneous colour maybe due to the light conditions. In this test, 558 eucalyptuses were detected
ISPRS Int. J. Geo-Inf. 2018, 7, 109 16 of 20
and an area of 2632.1 m2 was estimated. The mean distance obtained was 14.9 m and 101
eucalyptuses were not identified.
In vineyard
Two differentcrop study
crops werecase,
alsosome
used species vegetation
to test the and other trees,
plugin, eucalyptus, different from
and vineyards. vineyard,
In the case of
were identified.
eucalyptus, In this case,
the resulted 254 trees
shapefile were
was not identified because
satisfactory, and 886.3 m2 of the
most areatrees
waswere
obtained. The mean
not recognized
distance
and somewas 13.3 mwere
shadows and identified
41 trees were missing
as trees (Figure
(Figure 11). 12).

ISPRS Int. J. Geo-Inf. 2018, 7, x FOR PEER REVIEW 15 of 20

homogeneous colour maybe due to the light conditions. In this test, 558 eucalyptuses were detected
and an area of 2632.1 m2 was estimated. The mean distance obtained was 14.9 m and 101
eucalyptuses were not identified.
In vineyard crop study case, some species vegetation and other trees, different from vineyard,
were identified. In this case, 254 trees were identified and 886.3 m2 of area was obtained. The mean
distance was 13.3 m and 41 trees were missing (Figure 12).

Figure 11. Eucalyptus

Figure 11. Eucalyptus crop
crop plugin
plugin result.
result.

In this study case, some polygons were not completely separated, because the eucalyptus trees
are very close together. Also, the shade of the trees is different, so it was not possible to obtain a
homogeneous colour maybe due to the light conditions. In this test, 558 eucalyptuses were detected
and an area of 2632.1 m2 was estimated. The mean distance obtained was 14.9 m and 101 eucalyptuses
were not identified.
In vineyard crop study case, some species vegetation and other trees, different from vineyard,
were identified. In this case, 254 trees were identified and 886.3 m2 of area was obtained. The mean
distance was 13.3 m and 41 treesFigure 11. Eucalyptus
were missing crop
(Figure plugin result.
12).

Figure 12. Vineyard crop result.

From the results obtained, testing the plugin with different plantations, some limitations were
founded. These limitations could be overcome with the use of DSM, which can be generated from
UAV imagery. Several studies proved that, with the DSM, it is possible to achieve a good
performance in the extraction of individual tree metrics [46]. The plantation background is relevant
to understand the accuracy of the results obtained with this plugin. As in olive crop some shrubs
were wrongly identified as olives (Figure 9). In these cases, the DSM could help into distinguish tree
crowns from the surrounding vegetation more reliably. In the future a new method should be
implemented in this plugin integrating the12.
Figure
Figure
DSM
12. in thecrop
Vineyard
Vineyard
thatresult.
crop is methodology proposed.
result.

From the results obtained, testing the plugin with different plantations, some limitations were
founded. These limitations could be overcome with the use of DSM, which can be generated from
UAV imagery. Several studies proved that, with the DSM, it is possible to achieve a good
performance in the extraction of individual tree metrics [46]. The plantation background is relevant
to understand the accuracy of the results obtained with this plugin. As in olive crop some shrubs
ISPRS Int. J. Geo-Inf. 2018, 7, 109 17 of 20

From the results obtained, testing the plugin with different plantations, some limitations were
founded. These limitations could be overcome with the use of DSM, which can be generated from
UAV imagery. Several studies proved that, with the DSM, it is possible to achieve a good performance
in the extraction of individual tree metrics [46]. The plantation background is relevant to understand
the accuracy of the results obtained with this plugin. As in olive crop some shrubs were wrongly
identified as olives (Figure 9). In these cases, the DSM could help into distinguish tree crowns from
the surrounding vegetation more reliably. In the future a new method should be implemented in this
plugin integrating the DSM in the that is methodology proposed.

4.3. Processing Time

In order to compare the performance of the plugin developed, the processing time of the
experimental tests were recorded (Table 2), for olive crop. These procedures were performed in
the same environment: same computer, same CPU (Intel® , Santa Clara, CA, USA, Core™ i7-5500U
CPU@ 2.40 GHz), same memory (8.00 GB) and same operating system (Windows 10 Home, Microsoft,
Seattle, WA, USA). The objective of this analysis was to evaluate the performance of the plugin created
comparing the processing time of it with the manual procedure.

Table 2. Processing time of manually procedure and plugin.

Time (s)
Procedure Tool Algorithm
Experimental
Plugin
Procedure
GRASS GIS 7 i.segment (default values) 1837
OBIA segmentation OTB Watershed segmentation 319
GRASS GIS 7 i.segment (thresholding 0.05 until 0.65) 2017
OTB Compute second order statistics 3
OBIA classification—supervised
OTB TrainImageClassifier (svm) 14
classification
OTB Image Classification 8
OBIA
classification—unsupervised OTB Unsupervised kmeans image classification 3
classification
GDAL Raster Calculator 8
Fractal Analysis
GDAL Polygonize 69
Segmentation with i.segment (default values) 1917
TOTAL (with unsupervised
Segmentation with Watershed algorithm 399 2118
classification)
Segmentation with thresholding 2097

Several conclusions can be drawn when considering the results that are presented in Table 2.
The total processing time was calculated when considering three different segmentation methods.
The Watershed algorithm took 60 3900 . The segmentation with i.segment, defined with the default values,
took 310 5700 and the segmentation with the thresholding takes 340 5700 . The plugin processing time
was also recorded and took 350 1800 . From the manual procedures, the algorithm implemented in the
plugin (with thresholding method) was the slowest method. However, the evaluation of the results
that were obtained in this work proved that the i.segment with thresholding were the most satisfactory
approach. The processing time was similar with the segmentation performed manually or with
the plugin using the i.segment with thresholding. Even though the manual processing took less 2100
than the automatized processing implemented in the plugin, it must be considered that the manual
processing was performed step by step and the plugin ran all of the algorithms automatically in one
step. Furthermore, the fractal analysis including the estimation of the number of trees, the total area
of the tree crown, the centroid of a tree, the perimeter and the total number of missing trees were
implemented in the plugin automatizing this process. A method was defined where the missing
trees are automatically detected, which improved the plugin capabilities, as this process can be very
complex to perform without a specific script. When combining this with the high quality of the results
obtained, this methodology proved to be the best one and was chosen to be implemented in the plugin.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 18 of 20

The total processing time of the plugin for eucalyptus and vineyards resulted in 3 h 300 3100 and 1 h
200 1800 , respectively. These results proved that the processing time depends on the crop used and the
image size.

5. Conclusions
The automatic trees detection using a UAV image is a complex process. Different types of objects
can be identified in the UAV images, so it is required a robust algorithm to detect the correct objects.
UAV technology combined with GIS software proved to be efficient in the detection of tree crown, as it
combines tools to visualize, manipulate, analyse and process geographic information data acquired by
UAVs. The open source GIS software, as QGIS, provides the possibility of developing new plugins in
order to automatize procedures and integrates several external algorithms, for instance, processing image
algorithms. This functionality allowed for developig/creating a new plugin with the implementation
of the more accurate methodology to obtain the parameters/metrics of tree crown. In this work, a new
plugin was created automatizing the process of obtaining the parameters/metrics of tree crown. Different
methodologies were tested and the more accurate was implemented in the plugin. To evaluate the
performance of the plugin, the time processing of the set of procedures implemented was compared
with the manual procedure. From the results that were obtained, it was concluded that the methodology
implemented in the segmentation stage using i.segment with thresholding method presented the most
satisfactory result in terms of processing time and accuracy. However, some limitations were founded in
the code implemented, which were related to the precise identification of trees of a specific crop. Therefore,
in the future, the plugin can be improved with a more efficient code and new methods. For instance,
it can be improved when considering: (i) the correctly overlap of the bands; (ii) to use hyperspectral
images instead of multispectral images, as the hyperspectral images presents continuous values allowing
for a better distinction between the objects; (iii) to provide morphological operations, such as erosion to
remove surrounding pixels and dilatation of the objects adding pixels to the surrounding areas; (iv) added
supervised classification algorithms; and, (v) incorporation of DSM in the methodology. This plugin was
created in the context of plantations, however the plugin presents a huge potential for other different
areas such in small-crop delineations (cabbage fields, strawberries plantations, among others) or even
non-environmental applications such as counting cars or buildings. The greater advantages of the
developed plugin are the possibility of estimating the parameters with a simple and intuitive tool,
not requiring the knowledge of the algorithms existent in QGIS and the possibility of the incorporation of
algorithms and functions from others software. The first version of this tool is presented in this paper.
The plugin is free, open source and available in http://www.fc.up.pt/pessoas/liaduarte/treecrown.rar.

Acknowledgments: The authors would like to thank Eye2Map Lda, Porto, Portugal, for the data provided.
Author Contributions: Ana Cláudia Teodoro conceived and designed the experiments; Pedro Silva and Lia Duarte
performed the experiments; Ana Cláudia Teodoro and Lia Duarte analyzed the data; and Lia Duarte and
Ana Cláudia Teodoro wrote the paper.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Santoro, F.; Tarantino, E.; Figorito, B.; Gualano, S.; D’Onghia, A.M. A tree counting algorithm for precision
agriculture tasks. Int. J. Digit. Earth 2013, 6. [CrossRef]
2. Bazi, Y.; Malek, S.; Alajlan, N.; AlHichri, H. An Automatic Approach for Palm Tree Counting in UAV Images.
Presented at the IEEE Geoscience and Remote Sensing Symposium IGARS, Quebec City, QC, Canada,
13–18 July 2014.
3. Pierrot-Deseilligny, M. Presentation, IGN/ENSG, France, Institut Geographique National. Available online:
http://perso.telecomparistech.fr/~tupin/JTELE/PRES012009/Deseilligny.pdf (accessed on 10 March 2016).
4. Tian, J.Y.; Wang, L.; Li, X.J.; Gong, H.L.; Shi, C.; Zhong, R.F.; Liu, X.M. Comparison of UAV and WorldView-2
imagery for mapping leaf area index of mangrove forest. Int. J. Appl. Earth Obs. Geoinform. 2017, 61, 22–31.
[CrossRef]
ISPRS Int. J. Geo-Inf. 2018, 7, 109 19 of 20

5. Senthilnath, J.; Kandukuri, M.; Dokania, A.; Ramesh, K.N. Application of UAV imaging platform for
vegetation analysis based on spectral-spatial methods. Comput. Electron. Agric. 2017, 140, 8–24. [CrossRef]
6. Malehmir, A.; Dynesius, L.; Paulusson, K.; Paulusson, A.; Johansson, H.; Bastani, M.; Wedmark, M.;
Marsden, P. The potential of rotary-wing UAV-based magnetic surveys for mineral exploration: A case study
from central Sweden. Lead. Edge 2017, 36, 552–557. [CrossRef]
7. Suh, J.; Choi, Y. Mapping hazardous mining-induced sinkhole subsidence using unmanned aerial vehicle
(drone) photogrammetry. Environ. Earth Sci. 2017, 76. [CrossRef]
8. McKenna, P.; Erskine, P.D.; Lechner, A.M.; Phinn, S. Measuring fire severity using UAV imagery in semi-arid
central Queensland, Australia. Int. J. Remote Sens. 2017, 38, 4244–4264. [CrossRef]
9. Kang, J.; Wang, L.; Chen, F.; Niu, Z. Identifying tree canopy areas in undulating eucalyptus plantations using
JSEG multi-scale segmentation and unmanned aerial vehicle near-infrared imagery. Int. J. Remote Sens. 2017,
38, 2296–2312.
10. Park, T.; Cho, J.; Lee, J.; Lee, W.; Choi, S.; Kwak, D.; Kim, M. Unconstrained approach for isolating individual
trees using high-resolution aerial imagery. Int. J. Remote Sens. 2014, 35, 89–114. [CrossRef]
11. Hassaan, O.; Nasir, A.K.; Roth, H.; Khan, M.F. Precision Forestry: Trees Counting in Urban Areas Using
Visible Imagery based on an Unmanned Aerial Vehicle. IFAC-PapersOnLine 2016, 49–16, 16–21. [CrossRef]
12. Nevalainen, O.; Honkavaara, E.; Tuominen, S.; Viljanen, N.; Hakala, T.; Yu, X.; Hyyppa, J.; Saari, H.;
Polonen, I.; Imai, N.N. Individual Tree Detection and Classification with UAV-Based Photogrammetric Point
Clouds and Hyperspectral Imaging. Remote Sens. 2017, 9, 185. [CrossRef]
13. Özcan, A.H.; Hisar, D.; Sayar, Y.; Ünsalan, C. Tree canopy detection and delineation in satellite images using
probabilistic voting. Remote Sens. Lett. 2017, 8, 761–770. [CrossRef]
14. Koukoulas, S.; Blacknurn, G.A. Mapping individual tree location, height and species in broadleaved deciduous
forest using airborne LIDAR and multi-spectral remotely sensed data. Int. J. Remote Sens. 2007, 26, 431–455.
[CrossRef]
15. Katoh, M.; Gougeon, F.A. Improving the Precision of Tree Counting by Combining Tree Detection with
Canopy Delineation and Classification on Homogeneity Guided Smoothed High Resolution (50 cm)
Multispectral Airborne Digital Data. Remote Sens. 2012, 4, 1411–1424. [CrossRef]
16. Tanhuanpaa, T.; Saarinen, N.; Kankare, V.; Nurminen, K.; Vastaranta, M.; Honkavaara, E.; Karjalainen, M.;
Yu, X.; Holopainen, M.; Hyyppa, J. Evaluating the Performance of High-Altitude Aerial Image-Based Digital
Surface Models in Detecting Individual Tree Canopys in Mature Boreal Forests. Forests 2016, 7, 143. [CrossRef]
17. Thiel, C.; Schmullius, C. Comparison of UAV photograph-based and airborne lidar-based point clouds over
forest from a forestry application perspective. Int. J. Remote Sens. 2017, 38, 2411–2426. [CrossRef]
18. Wei, T.; Lin, Y.; Yan, L.; Zhang, L. Tree species classification based on stem-related feature parameters derived
from static terrestrial laser scanning data. Int. J. Remote Sens. 2016, 37, 4420–4440. [CrossRef]
19. Yin, D.; Wang, L. How to assess the accuracy of the individual tree-based forest inventory derived from
remotely sensed data: A review. Int. J. Remote Sens. 2016, 37, 4521–4553. [CrossRef]
20. Rosell, J.R.; Sanz, R. A Review of Methods and Applications of the Geometric Characterization of Tree Crops
in Agricultural Activities. Comput. Electron. Agric. 2012, 81, 124–141. [CrossRef]
21. Jiang, H.; Chen, S.; Li, D.; Wang, C.; Yang, J. Papaya Tree Detection with UAV Iamges Using a
GPU-Accelerated Scale-Space Filtering Method. Remote Sens. 2017, 9, 721. [CrossRef]
22. Guerra-Hernández, J.; González-Ferreiro, E.; Monleón, V.J.; Faias, S.P.; Tomé, M.; Díaz-Varela, R.A. Use of
Multi-Temporal UAV-Derived Imagery for Estimating Individual Tree Growth in Pinus pinea Stands. Forests
2017, 8, 300. [CrossRef]
23. Deng, G.; Li, Z.; Wu, H.; Zhang, X. Automated Extracting Tree Canopy from Quickbird Stand Image.
In Proceedings of the 4th Conference on Computer and Computing Technologies in Agriculture (CCTA),
Nanchang, China, 22–25 October 2010; IFIP Advances in Information and Communication Technology,
AICT-344 (Part I); Springer: Berlin/Heidelberg, Germany, 2010; pp. 304–311.
24. Chianucci, F.; Disperati, L.; Guzzi, D.; Bianchini, D.; Nardino, V.; Lastri, C.; Rindinella, A.; Corona, P.
Estimation of canopy attributes in beech forests using true colourdigital images from a small fixed-wing
UAV. Int. J. Appl. Earth Obs. Geoinform. 2016, 47, 60–68. [CrossRef]
25. Panagiotidis, D.; Abdollahnejad, A.; Surový, P.; Chiteculo, V. Determining tree height and canopy diameter
from high-resolution UAV imagery. Int. J. Remote Sens. 2017, 38, 2392–2410.
ISPRS Int. J. Geo-Inf. 2018, 7, 109 20 of 20

26. Blaschke, T. Object based image analysis for remote sensing. ISPRS J. Photogr. Remote Sens. 2010, 65, 2–16.
[CrossRef]
27. Cheng, G.; Han, J. A survey on object detection in optical remote sensing images. ISPRS J. Photogr. Remote Sens.
2016, 117, 11–28. [CrossRef]
28. Teodoro, A.C.; Araújo, R. A Comparison of Performance of OBIA Techniques Available in Open Source
Software (Spring and OTB/Monteverdi) considering Very High Spatial Resolution Data. J. Appl. Remote Sens.
2016, 10, 016011. [CrossRef]
29. Shahzad, N.; Ahmad, S.R.; Ashraf, S. An assessment of pan-sharpening algorithms for mapping mangrove
ecosystems: A hybrid approach. Int. J. Remote Sens. 2017, 38, 1579–1599. [CrossRef]
30. GRASS GIS. 2017. The World’s Leading Free GIS Software. Available online: http://grass.osgeo.org/
(accessed on 10 March 2017).
31. Orfeu Toolbox (OTB). 2017. Orfeo Toolbox Is Not a Black Box. Available online: https://www.orfeo-toolbox.
org/ (accessed on 10 March 2017).
32. GNU Operating System. 2017. Stallman Four Freedoms. Available online: https://www.gnu.org/
philosophy/free-sw.html (accessed on 12 March 2017).
33. QGIS. 2017. QGIS Project. Available online: http://www.qgis.org/ (accessed on 10 March 2017).
34. Grippa, T.; Lennert, M.; Beaumont, B.; Vanhuysse, S.; Stephenne, N.; Wolff, E. An Open-Source
Semi-Automated Processing Chain for Urban Object-Based Classification. Remote Sens. 2017, 9, 358. [CrossRef]
35. Duarte, L.; Teodoro, A.C.; Moutinho, O.; Gonçalves, J.A. Open-source GIS application for UAV photogrammetry
based on MicMac. Int. J. Remote Sens. 2016, 38, 8–10. [CrossRef]
36. Reed, B.C.; Brown, J.F.; VanderZee, D.; Loveland, T.R.; Merchant, J.W.; Ohlen, D.O. Measuring phenological
variability from satellite imagery. J. Veg. Sci. 1994, 5, 703–714. [CrossRef]
37. Python. 2017. Python Programming Language. Available online: http://python.org/ (accessed on 9 March 2016).
38. QGIS API. 2017. QGIS API Documentation. Available online: http://www.qgis.org/api/ (accessed on
10 March 2017).
39. PyQt4 API. 2017. PyQt Class Reference. Available online: http://pyqt.sourceforge.net/Docs/PyQt4/classes.
html (accessed on 10 March 2016).
40. Qt Designer. 2017. Qt Documentation. Available online: http://doc.qt.io/qt-4.8/designer-manual.html
(accessed on 10 March 2017).
41. Momsen, E.; Metz, M. Grass Osgeo Manual i.segment. Available online: https://grass.osgeo.org/grass73/
manuals/i.segment.html (accessed on 5 December 2016).
42. Awf-Wik. 2017. Forest Inventory Remote Sensing. Supervised Classification (Tutorial). Available online:
http://wiki.awf.forst.uni-goettingen.de/wiki/index.php/Supervised_classification (accessed on 12 March 2017).
43. Awf-Wik. 2017. Forest Inventory Remote Sensing. Unsupervised Classification (Tutorial). Available
online: http://wiki.awf.forst.uni-goettingen.de/wiki/index.php/Unsupervised_classification (accessed on
12 March 2017).
44. Klinger, Riccardo. Digital Geography, Unsupervised Classification in QGIS: Kmeans or Part Two. Available
online: http://www.digital-geography.com/unsupervised-classification-in-qgis-kmeans-or-part-two/#.
WT5ixbpFzIV (accessed on 13 August 2013).
45. GDAL/OGR. 2017. Geospatial Data Abstraction Library. Available online: http://www.gdal.org/ (accessed
on 10 March 2017).
46. Adão, T.; Peres, E.; Pádua, L.; Hruska, J.; Sousa, J.J.; Morais, R. UAS-based hyperspectral sensing
methodology for continuous monitoring and early detection of vineyard anomalies. In Proceedings of
the UAS4ENVIRO—Small Unmanned Aerial Systems for Environmental Reasearch—5th Edition, Vila Real,
Portugal, 28–30 June 2017.

© 2018 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).