rgb and multispectral uav image classification of
TRANSCRIPT
Student thesis series INES nr 466
Sigurbjörn Bogi Jónsson
RGB and Multispectral UAV image classification
of agricultural fields using a machine learning
algorithm
2018
Department of
Physical Geography and Ecosystem Science
Lund University
Sölvegatan 12
S-223 62 Lund
Sweden
i
Sigurbjörn Bogi Jónsson (2018).
RGB and Multispectral UAV image classification of agricultural fields using a
machine learning algorithm
Master degree thesis, 30 credits in Geomatics
Department of Physical Geography and Ecosystem Science, Lund University
Level: Master of Science (MSc)
Course duration: January 2018 until June 2018
Disclaimer
This document describes work undertaken as part of a program of study at the
University of Lund. All views and opinions expressed herein remain the sole
responsibility of the author, and do not necessarily represent those of the institute.
ii
RGB and Multispectral UAV image classification of
agricultural fields using a machine learning
algorithm
Sigurbjörn Bogi Jónsson
Master thesis, 30 credits, in Geomatics
Lars Eklundh
Lund University, Department of Physical Geography and Ecosystem
Science
Perola Olsson
Lund University, Department of Physical Geography and Ecosystem
Science
Exam committee:
Torbern Tagesson and Jonas Ardö, Lund University,
Department of Physical Geography and Ecosystem Science
iii
iv
Abstract
A common technique within image analysis is image classification which describes the
process of reducing the information content of an image into few user-defined classes.
With the emergence of unmanned aerial vehicles (UAVs), high spatial resolution (cm-
level) images can be collected. This thesis aims at testing different methods for
classifying UAV images from an agricultural crop site in the south of Skåne, Sweden.
To classify the UAV images, the Random Forest algorithm was used. Random Forest
(RF) is an ensemble classifier which consists of multiple classification-trees, where the
results of each individual tree contributes a single vote for to which class each pixel or
segment belongs to. To evaluate the results a few objectives are presented. First of all,
two cameras (RGB and multispectral) were used to examine the effect of different
wavelengths bands on classification accuracy. Furthermore, the effects of spatial
resolution, segmentation and integration of additional data were tested. To evaluate
these different strategies a few classification examples were tested; two general
classification cases, a 5-class classification and 11-class classification, and one
specialized case where the high resolution of the UAV was used to classify a crop field
consisting of two crop types.
Both RGB and multispectral cameras performed well, reaching overall accuracies
greater than 75% for all classification cases. Results from the general cases show little
difference between RGB and multispectral cameras. However, in performing the
specialized case classification, i.e., analyzing a field containing two spectrally similar
classes, the multispectral camera outperformed the RGB. The pixel size has a big
impact on resulting classification accuracy for both RGB and multispectral cameras
(30% difference in accuracy ranging from 5 cm to 1 m pixel size), where higher
accuracies are achieved at higher spatial resolutions. By integrating addition data
sources in the pixel-by-pixel classification method, accuracies increase by a factor of
>10%. The Mean – texture feature turns out to be the most important texture feature for
both cameras. The highest accuracy, for both RGB and multispectral classification, was
achieved by classifying groups of pixels into segments, reaching overall accuracy of
>90%.
Overall, UAV image classification works well for agricultural farm mapping and is a
good monitoring tool due to its quick deployment, ease of data collection and accurate
results.
v
vi
Table of Contents
1. Introduction ............................................................................................................ 1
1.1. Objectives ........................................................................................................ 3
2. Background ............................................................................................................ 5
2.1. Remote sensing and the electromagnetic spectrum ......................................... 5
2.2. Image classification ......................................................................................... 6
2.3. Random Forest Classifier ................................................................................ 7
2.4. Texture analysis and Vegetation indices ......................................................... 8
2.5. Accuracy validation ......................................................................................... 9
3. Methodology ........................................................................................................ 11
3.1. Study area ...................................................................................................... 11
3.2. Data ............................................................................................................... 13
3.2.1. Field work – sampling method .............................................................. 14
3.3. Methods ......................................................................................................... 15
3.3.1. Information classes – Classification strategy ......................................... 15
3.3.1.1. General Case ................................................................................... 15
3.3.1.2. Specialized Case ............................................................................. 17
3.3.2. Wavelengths and spatial resolution ....................................................... 17
3.3.3. Segmentation .......................................................................................... 18
3.3.4. Integration of additional data ................................................................. 19
3.3.5. Accuracy Assessment ............................................................................ 19
3.3.6. Random Forest ....................................................................................... 20
4. Results .................................................................................................................. 21
4.1. General Case ................................................................................................. 21
4.1.1. 5-class classification .............................................................................. 21
4.1.2. 11-class classification ............................................................................ 24
4.2. Specialized Case ............................................................................................ 30
5. Discussion ............................................................................................................ 33
5.1. General Thoughts .......................................................................................... 33
5.2. Effects of spatial resolution, segmentation and additional features .............. 33
5.3. UAV pre- and postprocessing ....................................................................... 34
5.4. Accuracy evaluation ...................................................................................... 34
vii
6. Conclusion ........................................................................................................... 37
7. References ............................................................................................................ 39
8. Appendices ........................................................................................................... 41
8.1. Appendix A: Classification results, images. ................................................. 41
8.2. Appendix B: Classification results, confusion matrices. ............................... 67
1
1. Introduction
In recent years, unmanned aerial vehicles (UAVs) have become a popular tool for
collecting remotely sensed data. Consequently, a wide range of UAV applications have
emerged within various scientific fields, from land cover mapping and environmental
monitoring to precise civil engineering surveys. A common application within
environmental science is the acquisition of aerial images, collected using either red-
green-blue (RGB) or multispectral cameras onboard the UAV. Aerial images contain
vast amount of information, thus allowing various interpretations to be made depending
on desired output. A common technique within image analysis is image classification
which describes the process of reducing the information content of an image into few
user-defined classes. In other words, every pixel in the original image is assigned a new
value, nominal or numeric, based on some defined classification criteria. The output
image is often named land-cover map and typically contains broad information classes
such as forest, water, crop fields, etc. (Özdogan, 2015).
Before the usage of UAVs for classifying remotely sensed images, satellites were (and
still are) the most commonly used sensor platform. There are several advantages of
using UAV images over satellite imagery, e.g. quick deployment, data can be collected
on demand and UAVs are more flexible in terms of flight height and angle (Rango et
al., 2009). Another key advantage over satellite imagery is that UAV images can be
collected in very high spatial resolution (cm-level). Spatial resolution corresponds to
the smallest identifiable surface object on the image and is often denoted as the pixel
size, even though pixels can be scaled to larger or lower sizes (Chuvieco, 2016). With
higher spatial resolution the broad information classes mentioned earlier (forest, crop
fields, etc.) can be classified even further. Therefore, more fine-scaled classification
can be performed allowing different crop types or different tree species to be detected,
as examples.
An environmental field that greatly benefits in having accurate land cover maps are
agricultural farmlands. The farming industry is highly sensitive of factors such as soil
condition, temperature, rainfall etc. and being able to monitor crop fields at regular
basis increases the chances of production optimization. UAV image classification can
play important role in this monitoring process. Having an accurate overview of the
croplands area and being able to map healthy soil and target weeds, as an example, are
factors which aid in decision making and can play a crucial role in future farming
strategies.
2
How an image gets classified ultimately depends on its pixel’s values but also on which
classification method is chosen. Various classification algorithms are available, both
parametric and non-parametric classifiers. The non-parametric algorithms such as
Random Forest classifier (RF) and Support Vector Machine (SVM) generally produce
better results, with classification accuracy of 10%-20% higher than for parametric
classifiers (Özdogan, 2015). Another factor which influences the classification result is
whether each pixel in the original image is classified (pixel-by-pixel technique) or if
classifying group of pixels forming segments, based on pixel´s spectral similarity
(object-based image analysis), is chosen.
Another important factor in the classification process is the construction of the original
image, i.e. the number of bands which the image is made up from. A standard aerial
photo, captured by an RGB (red, green, blue) camera is made up from three bands;
namely the red, green and blue band which correspond to the visible part of the
electromagnetic spectrum (EM), the part which the human eye can detect. By utilizing
other parts of the electromagnetic spectrum, e.g. the infrared band or thermal band,
more information can be gathered for the surface object. Camera sensors that gather
spectral information in other parts of the EM than the visible part are named
multispectral cameras and cameras only capable of retrieving information from the
visible part are named RGB-cameras.
Although many articles and books have been published in the field of image
classification, there are many aspects not yet thoroughly investigated, especially for
UAV images. Comparing pixel-by-pixel and object-based image analysis (OBIA)
techniques has not been substantially reported for UAV images and comparing RGB to
multispectral UAV image classification is greatly lacking scientific review.
3
1.1. Objectives
This thesis aims at testing various methods for classifying agricultural fields using UAV
images. The primary aim is to compare classification accuracies by using two different
cameras onboard the UAV, i.e. an RGB and a multispectral camera to assess the
importance of including wavelength bands from the non-visible part of the
electromagnetic spectrum. To evaluate the results, three additional objectives are
suggested below which aim to contribute to the ever-growing knowledge base of UAV
image classification by analyzing the effect of:
1) Spatial resolution. The effect of pixel size on the classification accuracy.
2) Segmentation. The pixel-by-pixel and segmentation classification processes
will be evaluated.
3) Integrating additional data to improve accuracy. In this study, digital elevation
model (DEM), texture features and vegetation indices are added as input
features in the classification algorithm and its effect on classification accuracy
documented.
For all cases, the Random Forest algorithm will be used as classifier, having been used
successfully for classifying UAV images (Baron, Hill, & Elmiligi, 2018; Franklin &
Ahmed, 2017).
4
5
2. Background
2.1. Remote sensing and the electromagnetic spectrum
Remote sensing describes the process of collecting data without physically touching the
object of interest, i.e. data is collected remotely from a distance. In its most simple case,
sensors aboard various platforms (UAV, airborne, satellite, etc.) record electromagnetic
energy signals, which are either reflected or emitted, from ground objects and that
signal later gets transformed to useful information. Reflected energy describes how
utilizing sunlight as an energy source causes a portion of that energy to reflect or bounce
back into space to be received by the sensor.
Electromagnetic radiation can vary from very short to vey long wavelengths.
Consequently, the range of wavelengths has been classified into regions or spectral
bands, which are organized in the electromagnetic spectrum displayed in Figure 1
(Chuvieco, 2016). Shorter wavelengths (Gamma or X-rays) have higher frequencies
than longer wavelengths (Microwaves or Radio waves) and therefore have a higher
energy content (Chuvieco, 2016).
Figure 1: The electromagnetic spectrum and its spectral bands division. The visible spectrum defines the range
which the human eye can sense. (source: Figure modified from Chuvieco, 2016)
The visible and infrared bands are commonly used bands on UAV platforms. The
visible spectrum ranges from 0.4 to 0.7 μm and can be further divided into blue (0.4 -
0.5 μm), green (0.5 - 0.6 μm) and red (0.6 – 0.7 μm) bands. As the name suggests, the
visible band defines the range of the electromagnetic spectrum which the human eyes
can sense. The non-visible Infrared band ranges from 0.7 to 14 μm and can be divided
into near infrared (NIR), short wave infrared (SWIR) and thermal infrared (TIR) bands
(Chuvieco, 2016).
How electromagnetic energy is reflected, absorbed or emitted across various spectral
bands is governed by the ground surface and how it reacts to different energy levels.
Furthermore, the land cover type as well as its physical and chemical properties controls
6
how the reflectance acts across the electromagnetic spectrum. This reflectance behavior
can be graphed, with reflectance on the y-axis and wavelengths on the x-axis, to
visualize the spectral “fingerprint” of the ground object. Figure 2 displays this kind of
graph, often called spectral signature, for 3 different land cover types; namely water,
vegetation and concrete (Chuvieco, 2016).
Figure 2: Spectral signatures for various ground objects. As seen from the graph, objects reflectance varies greatly
over various wavelengths (source: Figure modified from Chuvieco, 2016)
2.2. Image classification
Classically, the image classification process is divided into supervised and
unsupervised method. For the unsupervised method, an automatic search is performed
to find clusters of homogenous values within the image. These clusters then need to be
assigned to an information class manually by the user (Chuvieco, 2016). On the other
hand, the supervised method requires previous knowledge by the user, since the user
provides beforehand which information classes should be classified. To do so, a set of
training areas are created where each training area defines an area of pixels (or a single
pixel) representing a desired information class. The information within these training
areas are then used to classify the rest of the image (Chuvieco, 2016).
Multispectral cameras, like the name suggests, collect data for a given object in many
spectral bands. So, in contrast to normal RGB cameras which solely collect data for the
red, green and blue band of the electromagnetic spectrum, multispectral cameras
include addition bands, e.g. near infrared, short wave infrared or thermal infrared bands.
So, depending on how many bands the camera includes, the final image will consist of
7
as many features. Thus, for each of the layers, each pixel value represents the
reflectance for that specific band. To classify the image, the bands that best distinguish
between suggested information classes can then be used to classify the image. For
example, to classify artificial surfaces from vegetation the near infrared (NIR) band can
be used since the reflectance for vegetation is high in the NIR band but low for artificial
surfaces.
In general, if the pixel size of an image is equal in size or coarser than the object of
interest a pixel-by-pixel classification is appropriate (Blaschke, 2010). The pixel-by-
pixel technique simply refers to that each pixel in the original image will be evaluated
and classified into a new value. However, with higher resolutions pixels can be much
smaller than the object of interest, increasing the overall complexity of the image since
single objects get split up into numerous pixels. To solve the problem where multiple
pixels make up a target object, a method named object based image analyses (OBIA)
has been utilized which merges pixels into group of pixels called segments based on
their spectral similarities (Blaschke, 2010). Subsequently, every segment will be
assigned to an information class. The OBIA segmentation method has proved to be very
useful in high resolution imagery (Franklin & Ahmed, 2017; Pande-Chhetri, Abd-
Elrahman, Liu, Morton, & Wilhelm, 2017; Puliti, Talbot, & Astrup, 2018).
Reflectance values are not the only way to separate between information classes. The
usage of texture measures, a statistical method of structure useful to separate smooth
surfaces from more rough surfaces, has proved to increase overall accuracy in very high
resolution imagery (Feng et al., 2015; Laliberte et al., 2009). Other information layers,
such as elevation and shape of objects, have been used to complement pixel or segments
separation (van der Werff & van der Meer, 2008).
2.3. Random Forest Classifier
Random Forest (RF) is an ensemble classifier which consists of multiple classification-
trees, where the results of each individual tree contributes a single vote for to which
class each pixel or segment belongs to. The most frequent class, i.e. the one with the
most votes, will be the final information class the pixel or segment will be assigned to
(Rodriguez-Galiano et al., 2012). The input variables used (often named features and
typically include image spectral bands, elevation, texture information etc.) are
randomly selected for these building trees. Furthermore, the RF algorithm has the
option to produce accuracy assessment called out-of-bag error (OOB error), which
evaluates performance by leaving out part of the training data for evaluation (Millard
& Richardson, 2015). The RF classifier is highly sensitive to selection and size of
8
training areas and maximum accuracy is acquired when training areas are randomly
distributed, should be as large as possible and should have minimal spatial
autocorrelation (Millard & Richardson, 2015).
Another important factor which influences the Random Forest performance is if the
training dataset is balanced or imbalanced. If the training set is imbalanced, i.e. the
number of training pixels within each feature class is uneven, the classification may
favor the most frequent training class (Millard & Richardson, 2015). Lastly, the
Random Forest also provides features importance, i.e. what feature contributes the most
in the decision making (Pal, 2005). This can be useful for feature selection in cases
where numerous features are used.
2.4. Texture analysis and Vegetation indices
Haralick et al. (1973) wrote a paper describing fourteen texture features for image
classification. Unlike image bands, which describe spectral properties of an object,
texture describes spatial variation over a fixed area (Feng et al., 2015). Texture statistics
are derived from a gray level co-occurrence matrix (GLCM) where different angular
relationships and distance between neighboring cell pairs are calculated (Haralick et al.,
1973). Typical texture effect is differentiating between rough and smooth surfaces.
Vegetation indices (VI) are created from image bands to increase or enhance the
vegetation effects in an area (Huete et al., 2002). Therefore, for simple generalization,
it is often linked with the amount of green vegetation present on the surface. Various
VIs exist both extracted from multispectral sensors as well as RGB sensors. For
multispectral images the NDVI (Normalized Vegetation Index) is very popular, having
been used successfully in detecting healthy vegetation for numerous years (Zhou et al.,
2017, Matese et al., 2017). The VARI (Visible Atmospherically Resistance Index) is an
color index which is only utilizes the blue, green and red band from the EM ((Milas et
al., 2017). Table 1 below displays the equations for these two indices.
Table 1: Formulas for calculating the NDVI and VARI vegetation indices.
Vegetation Index Equation
NDVI (NIR–Red) / (NIR+Red)
VARI (Green-Red) / (Green+Red-Blue)
9
2.5. Accuracy validation
It is necessary to validate any classified land cover map in order to have some statistical
sense of how accurate it really is. To validate the maps, first a confusion matrix is
created which compares class classified on a map to the true class on ground (usually
true samples on ground are sampled in field). Various statistical properties can be
derived from the confusion matrix such as Overall Accuracy (correctly mapped points
divided by total number of points), user´s and producer´s accuracy (chances for each
class to be correctly mapped) and Kappa, which describes how much better than chance
alone the map is (Congalton, 1991).
Another accuracy value possible to obtain is the Random Forest´s Out of bag error
(OOB error) described in section 2.3 where the classes which are not used in building
trees are used for evaluation.
10
11
3. Methodology
To evaluate the UAV classification performance, a few subsets of tests will be
performed. Firstly, a broad image classification consisting of 5 core classes (“5-class”
classification) is performed and compared to a much more detailed classification
consisting of 11 classes (“11-class” classification). For both these instances, the same
step-by-step approach will be used and tested on both RGB and multispectral cameras.
Secondly, a more UAV focused classification will be performed by classifying a field
consisting of more than one class. In other words, the first case focuses on classical
broad land cover mapping whereas the second case really takes advantage of the UAV
high spatial resolution for examining the variation within these broadly defined classes.
3.1. Study area
For this project the agroecological field experiment plot sites (SITES) in Lönnstorp,
located in Skåne in Sweden, was chosen as the area of interest. Lönnstorp´s
environment consists of various manmade fields (Figure 3) which contain different type
of crops. Furthermore, two of the classes are mixed with old rye and legumes (clover
and lucerne). These wide range of different surface types provides a basis for
classification hierarchy, i.e. it is possible to rank easily classifiable classes from more
difficult ones. Due to the high spatial resolution of the UAV images, classifying mixed
fields (fields containing two classes) and within field variation becomes possible. At
coarser resolution, the pixel size would cause these variations to average out and thus
made harder to identify.
Figure 3: Location of Lönnstorp (yellow star) in the south of Sweden and the agricultural fields on site
containing various types of crops.
12
The agricultural fields shown on Figure 3 will, among other, be used as target
information classes for the classification. Figure 4 shows a photograph taken in the field
of all these classes.
Figure 4: The agricultural fields at Lönnstorp: a) Winter Wheat (ref), b) Winter Rapeseed, c) Kernza/Lucerne, d)
Kernza, e) Winter Wheat (orig), f) Grass Legume Lay, g) Spring Barley, h) Bare soil. Photos: Author.
a) b)
c) d)
e) f)
g) h)
13
3.2. Data
UAV images from 5th of May 2018 were collected in Lönnstorp using two different
camera sensors, i.e. the Parrot Sequoia multispectral sensor and GoPro 4 RGB sensor.
The Parrot Sequoia collects data in four spectral bands, as displayed in Table 2, whereas
the GoPro collects data only in the visible part of the EM spectrum.
Table 2: Available bands and their wavelengths for the Parrot Sequoia
Bands Parrot Sequoia [nm]
Blue X
Green 550
Red 660
Red Edge 735
Near Infrared 790
The UAV platform, a modified 3DR Solo quadcopter, was flown following a
predetermined flight plan at 70m altitude and collected 240 and 78 images using the
Parrot Sequoia and GoPro cameras respectively. The Agisoft PhotoScan software was
used to align and georeference the raw UAV photos and, by creating a dense point cloud
from the raw images, an orthophoto and digital elevation model (DEM) were exported
(Figure 5).
Figure 5: Orthophoto created from UAV images captured by the GoPro 4 camera (left) and multispectral image
created using the near infrared, red and green band from the Parrot Sequoia camera (right) from Lönnstorp.
14
The GoPro and Parrot Sequoia camera sensors will for the rest of this thesis be called
RGB and Multispectral cameras respectively. Similarly, RGB and Multispectral image
corresponds to the images created by the GoPro and Parrot Sequoia sensors.
3.2.1. Field work – sampling method
Field work was conducted in Lönnstorp on the 11th of May 2018 where a total of 595
GPS points were collected using an RTK GPS Topcon positioning system with sub-
decimeter accuracy. These points were used as training and evaluation data for the
Random Forest classifier. Additionally, 6 ground control points (GCPs) were collected
and used to georeference the raw UAV images in Agisoft (Figure 6).
Figure 6: Location of training points (blue circles), evaluation points (red circles) and ground control points (yellow
squares) collected in the field using an RTK GPS Topcon positioning system.
In collecting the GPS points, a stratified random sampling method was applied,
meaning that fixed number of random points within each information class are created.
These points were then visited in the field and their correct information value was
confirmed. At least 2 meters separated the points and if points were located close to an
edge between two classes, those were moved to new location where they were markedly
covered by a single class.
All data (GPS points and UAV images) is harmonized to the WGS84/UTM zone 33N
coordinate reference system.
15
3.3. Methods
3.3.1. Information classes – Classification strategy
To evaluate different classification strategies and compare RGB to multispectral
images, a set of tests must be performed. A total of three subtests were examined in this
project, two general cases and one specialized case (Figure 7).
Figure 7: Three different examples are used to compare RGB to multispectral image classification, i.e. two General
cases (5-class and 11-class classification) and one Specialized case.
3.3.1.1. General Case
The general classification case describes a typical classification scenario where the
surface is classified into both broad information classes (5-class classification) and
more detailed information classes (11-class classification). As the name suggests, the
broad 5-class and detailed 11-class cases describe classification scenarios consisting of
five and eleven information classes respectively (Figure 8). As seen from Figure 8, all
green vegetation surfaces listed in the 11-class classification have been merged into one
class in the 5-class case.
Figure 8: Information classes for the 11-class and 5-class classifications. All classes representing green vegetation
in the 11-class classification have been merged into a single class in the 5-class classification.
16
The reason for this summarization is that the green vegetation classes all have similar
spectral signatures which differ from the other four classes (bare soil, road, rye and
rapeseed), as seen in Figure 9. Even though “Rapeseed” has similar spectral curve to
the other vegetation classes, it is classified separately due to its visual distinction. For
the rest of this thesis, the term “5-class” and “11-class” will be used to describe these
two general case classification examples.
Figure 9: Spectral signatures for the 11-class classification information classes from the multispectral camera.
After the information classes have been decided a classification strategy needs to be
formed. To compare the RGB and multispectral images, the classification process was
the same for both sensors. In this study, the first step of the analysis was to evaluate the
effect of spatial resolution by classifying each UAV image at seven different pixel sizes.
Subsequently, the UAV image generating the highest classification accuracy was
further analyzed and two ways of improving the classification tested, i.e. by applying
image segmentation and integrating additional input features to the original UAV image
bands. Furthermore, the feature importance was documented and the least important
features removed from the Random Forest input stack. The reason for removing the
least important features was to evaluate how the classification accuracy varies by
reducing input bands of less importance. Thus, the Random Forest algorithm was run a
few times until one input band, the most important one, was remaining. Since 5-class
and 11-class cases are demonstrating the same surface (only different amount of
information classes), the last two steps (segmentation and integrating additional data)
will only be performed for the 11-class case. Figure 10 displays a simple flow chart of
the steps performed.
0
10000
20000
30000
40000
50000
60000
DN
nu
mb
er
Bare Soil Grass Legume Lay Rye
Winter Wheat (orig) Winter Wheat (ref) Spring Barley
Rapeseed Kernza Kernza/Lucerne
Natural Grass Road
Green Red Red edge NIR
17
Figure 10: Flow chart of the classification strategy for the 5-class and 11-class classification. Two ways of
improving the classification accuracy is tested on the 11-class classification, i.e. segmentation and integration of
additional input features.
3.3.1.2. Specialized Case
The specialized case describes a more detailed classification than the general case. The
high resolution provided by the UAV imagery allows for smaller objects to be detected
and more detailed within-field analysis. In this study, the intercropping field consisting
of Kernza (a type of wheatgrass) and Lucerne (legume) will be examined. Furthermore,
large amounts of dandelions (weed) grow within the field so total of three classes will
be classified within the field; i.e. Kernza, Lucerne and Dandelion. Since this a very
detailed classification, only the pixel-by-pixel method will be used.
3.3.2. Wavelengths and spatial resolution
As stated earlier, the Multispectral camera consists of four bands (green, red, red edge
and near infrared) and the RGB camera consists of the three visible bands of the EM.
To evaluate the effect of spatial resolution on classification accuracy, orthophotos at
seven different pixel sizes are created for both RGB and Multispectral images, i.e. at 5
cm, 10 cm, 20 cm, 30 cm, 40 cm, 50 cm and 1m. The resampling is done within the
Agisoft Photoscan software, by specifying appropriate pixel size for export based on
the dense point cloud created from the raw UAV images. Agisoft uses bilinear
interpolation in the creation process of the orthophotos (Agisoft PhotoScan User
Manual, 2018).
18
3.3.3. Segmentation
For testing the impact of segmentation, an object-based image analysis (OBIA) was
used to classify groups of spectrally similar pixels instead of classifying every
individual pixel (pixel-by-pixel method) in the image. To segment the UAV images,
the TerrSet Geospatial Monitoring and Modeling Software was used. The
SEGMENTATION-module within TerrSet accepts list of layers to be used for
segmentation and by tuning a similarity tolerance the user controls the generalization
level of the segmentation. The segment algorithm groups together pixels with similar
spectral properties, where lower similarity tolerance produces smaller, more detailed
segments. Three different similarity tolerance values were used in this analysis; values
of 50, 100 and 200. These values were selected by experimentation and produce three
levels of detail segment polygons (Figure 11).
The segment polygons are exported as shapefiles from TerrSet and are used to extract
zonal statistics from the UAV images, which acts as input in the Random Forest
algorithm. Since the Random Forest algorithm needs information in tabular format to
train and classify, the UAV image is saved as a python list where each list index
contains statistical information for the image for each segment polygon. Now, the
training segments (segments which include training data points) are used to train the
classifier based on the statistical information extracted from the image within these
polygons.
Figure 11: Zoom in on an area of the multispectral image. The top left displays a false color composite of the
multispectral image and the top right, bottom left and bottom right display three levels of segmentation with
similarity tolerance set to 200, 100 and 50 respectively.
19
3.3.4. Integration of additional data
Another way to improve the classification is adding other features as input to the
Random Forest algorithm. In this case, six texture features, one vegetation index (VI)
and a digital elevation model (DEM) are used as additional features.
To compute the texture features the Orfeo Toolbox, an image analysis toolbox in QGIS,
was used to calculate fourteen Haralick texture features (Haralick et al., 1973). As
suggested by Feng et al., 2015, the six least correlated texture features were chosen,
calculated from the grey level co-occurrence matrix for a single image band in four
angular directions (0°, 45°, 90°, 135°) and then taking the mean of these directions. The
texture features chosen are: correlation, entropy, inverse difference moment, mean, sum
average and variance.
The texture features are computed for a single image band. For the RGB image the red
band was chosen for texture extraction and for the multispectral image the NIR band
was chosen. These bands were chosen since they are less correlated than the other image
bands.
Another feature added as input to the Random Forest algorithm is a vegetation index.
For the multispectral image the NDVI (Normalized Vegetation Index) was chosen,
having been used successfully in detecting healthy vegetation for numerous years (Zhou
et al., 2017, Matese et al., 2017). For the RGB image, the VARI (Visible
Atmospherically Resistance Index) index was chosen. The equations for these VIs are
provided in Table 1.
Finally, a digital elevation model is exported from the Agisoft Photoscan software
which is created from the dense point cloud, generated from the raw UAV images.
3.3.5. Accuracy Assessment
To validate the classified maps, five accuracy parameters are logged; namely Overall
Accuracy (OA), Users and Producers accuracy, Kappa and the Out Of Bag score (OOB
score). The first four parameters are directly calculated from the error matrix but the
OOB score is derived from the Random Forest module.
20
3.3.6. Random Forest
All images were classified using the Random Forest classifier within the Scikit-learn
0.19.1 machine learning module in python 3.6. All parameters of the Random Forest
are the same for all classifications. A total number of 1000 trees are used, as suggested
by Millard & Richardson (2015), since using more trees does not harm the model but
rather stabilizes variable importance (Millard & Richardson, 2015). The OOB score
was logged for every classification and will act as a comparison accuracy variable to
the well-defined Overall Accuracy and Kappa value.
All the points sampled in the field are divided into training data (70%) and validation
data (30%). Since some classes have different number of points, a balanced training
sample was used where the amount of points in each class was equal to the number of
training points within the least frequent class. The feature importance was documented
during the 11-class classification to examine what features prove most valuable in the
classification process.
The size of the training data samples varies for the two cases. In the 5-class and 11-
class classification, a 25 cm buffer is created around each training point and these buffer
areas are used as training input for the classifier. The reason a 25 cm buffer is chosen
is simply to get as much variation within each class without risking overlapping into
other classes. For the specialized case (Kernza/Lucern), the training points were used
directly in the classifier, since the targeted objects are only few cm in diameter and
creating buffers in that scenario would risk averaging out the variation desired to detect.
21
4. Results
Below are the results from two test cases, i.e. the general case and specialized case
mentioned in section 3.3.1. This chapter is divided into two major sections. The first
section (4.1 General Case) contains figures and graphs from the 5-class and 11-class
classification cases whereas the second section (4.2 Specialized Case) contains results
from the specialized classification case. Only the classification yielding the highest
overall accuracy is displayed in these sections, but all output maps and confusion
matrices can be viewed in Appendix A and Appendix B respectively.
4.1. General Case
The steps followed the classification strategies mentioned in section 3.3 and the
objectives listed in section 1.1, i.e. the effect of spatial resolution, segmentation and
adding input feature were evaluated using both RGB and multispectral cameras.
4.1.1. 5-class classification
The Random Forest algorithm was tested using seven different pixel sizes and the
classification accuracy monitored. Figure 12 displays the overall accuracy curves for
the RGB and multispectral 5-class classification cases. Generally, the overall accuracy
is higher as the pixel size is smaller. The RGB camera produces the highest accuracy
of 88.3% at 10 cm pixel size and the multispectral camera produces the highest accuracy
of 86.1% at 5 cm pixel size. As seen from the graph, the RGB image accuracy generally
decreases rapidly at coarser spatial resolution. However, the multispectral image
accuracy, while also decreasing with coarser spatial resolution, decreases much less
steeply than the RGB one.
22
Figure 12: Relationship between overall accuracy and pixel size for the 5-class classification. The blue curve
represents this relationship for the RGB classification and the orange curve represents the multispectral
classification. In both cases, the overall accuracy increases as the pixel size decreases.
Figure 13 and Figure 14 display the classification maps which produced the highest
overall accuracy for the 5-class classification for the RGB and multispectral camera
respectively. Both these cases are example of pixel-by-pixel classification. As seen
from the section above, the highest accuracy was achieved having a pixel size of 10 cm
for the RGB image and pixel size of 5 cm for the multispectral one.
50
60
70
80
90
100
0 20 40 60 80 100
Ove
rall
accu
racy
[%
]
Pixel size [cm]
Overall Accuracy and Pixel Size5-class classification
RGB Multispectral
Figure 13: The 5-class classification map produced from the RGB image with 10 cm pixel size
23
Figure 14: The 5-class classification map produced from the multispectral image with 5 cm pixel size
A confusion matrix for the RGB 5-class classification is displayed in Table 3 and for
the multispectral 5-class classification in Table 4. Both tables display the user accuracy,
producer accuracy, overall accuracy, kappa value and OOB score for the classification
result. All confusion matrices for 5-class classification (both RGB and multispectral)
are displayed in Appendix B.
Table 3: Confusion matrix for the 5-class RGB classification with 10 cm pixel size. User and producer accuracies are
displayed for each information class and the final row presents the overall, kappa and OOB accuracies for the entire
classification.
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 44 1 5 0 0 0.88
Vegetation 3 82 2 7 0 0.87
Rye 0 0 10 0 1 0.91
Rapeseed 1 0 0 13 0 0.93
Road 1 0 0 0 10 0.91
Producer acc.: 0.9 0.99 0.59 0.65 0.91
Overall acc: 0.88 Kappa: 0.82 OOB: 0.88
24
Table 4: Confusion matrix for the 5-class multispectral classification with 5 cm pixel size. User and producer
accuracies are displayed for each information class and the final row presents the overall, kappa and OOB
accuracies for the entire classification.
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 46 0 0 4 0 0.92
Vegetation 7 75 3 9 0 0.8
Rye 0 0 11 0 0 1
Rapeseed 0 0 1 13 0 0.93
Road 1 0 0 0 10 0.91
Producer acc.: 0.85 1 0.73 0.5 1
Overall acc: 0.86 Kappa: 0.8 OOB: 0.92
As seen from the confusion matrices, both Rye and Rapeseed have much lower
producer accuracies then the other classes. The Rapeseed clearly gets mixed up with
the Vegetation class, which is not surprising because its spectral signature is very
similar to the other Vegetation classes (see Figure 9).
4.1.2. 11-class classification
Just like in the 5-class classification, the Random Forest algorithm was tested using
seven different pixel sizes and the classification accuracy monitored. Figure 15 displays
the overall accuracy curves for the RGB and multispectral 11-class classification cases.
Unlike in the 5-class case, here the trend is very similar for both RGB and multispectral
cameras. Again, the lowest pixel sizes generate the highest overall accuracy for both
RGB and multispectral cameras and accuracy decreases as pixel size gets larger. Both
RGB and multispectral images have the highest accuracy of 76.8% at 5 cm pixel size.
25
Figure 15: Relationship between overall accuracy and pixel size for 11-class classification. The blue curve
represents this relationship for the RGB classification and the orange curve represents the multispectral
classification. In both cases, the overall accuracy increases as the pixel size decreases.
Figure 16 and Figure 17 display the resulting classification maps for the 11-class
classification for the RGB and multispectral cameras respectively. In this section, three
classification maps for each camera sensor are created; namely pixel-by-pixel
classification (with pixel size equal to 5 cm), pixel-by-pixel with integrated additional
features and segmented classification. The segmentation classification provided the
highest accuracy when the similarity tolerance was equal to 200 for both RGB and
multispectral UAV images.
The multispectral classification map, classified using segmentation with ST equal to
200, has an overall accuracy of 94%, which is very high considering the complexity of
the surface and spectral similarities of the vegetated objects. Table 5 demonstrates the
overall accuracy, Kappa and the OOB score for the 5 cm multispectral UAV image with
and without integration of addition data, and segmentation. It is interesting that the
overall accuracy and Kappa are in similar range for all cases but the OOB score does
not follow that trend.
Table 5: Overall accuracy, Kappa and OOB score for the 11-case multispectral classification
Overall accuracy [%] Kappa [%] OOB score [%]
Image bands only 77 71 71
Additional features 88 85 99
Segmentation 94 92 64
30
40
50
60
70
80
0 10 20 30 40 50 60 70 80 90 100
Ove
rall
Acc
ura
cy [
%]
Pixel Size [cm]
Overall Accuracy and Pixel Size11-class classification
RGB multispectral
26
Figure 16: Classification maps from the 11-class RGB classification: a) Result from the pixel-by-pixel method using
blue, green and red bands as input features. b) Pixel-by-pixel classification with added features (texture, VI and
DEM). c) Result from the segmented classification with similarity tolerance equal to 200.
a)
b)
c)
27
Figure 17: Classification maps from the 11-class multispectral classification: a) Result from the pixel-by-pixel
method using green, red, red edge and NIR bands as input features. b) Pixel-by-pixel classification with added
features (texture, VI and DEM). c) Result from the segmented classification with similarity tolerance equal to 200.
a)
b)
c)
28
The feature importance was also examined when adding the texture, VI and DEM
features. By analyzing the feature importance, for both RGB and Multispectral cases,
the DEM and Mean texture features score high (Figure 18). The top three important
features for the multispectral classification are the DEM, red band and Mean-texture
input features. For the RGB the Mean-texture, DEM and Sum average-texture were in
the top three list. It is also interesting to notice that all visible bands (B, G and R) are
left out after the first Random Forest iteration in the RGB classification but the red band
is the second most important one in the multispectral classification.
A confusion matrix for the RGB 11-class classification is displayed in Table 6. Due to
the table size, only the RGB table is displayed in this section. The table displays the
user accuracy, producer accuracy, overall accuracy, kappa value and OOB score for the
classification result. All confusion matrices for the 11-class classification (both RGB
and multispectral) are displayed in Appendix B. For both RGB and multispectral, the
highest accuracy was recorded in the segmentation classification with similarity
tolerance equal to 200.
Figure 18: Feature importance for 11-class classification for the RGB (left) and Multispectral (right) cases. The digital elevation model
(DEM) and the Mean texture features both prove to be important features for the classification in both cases.
29
Table 6: Confusion matrix for the RGB 11-class classification produced by segmented classification, with similarity
tolerance equal to 200. User and producer accuracies are displayed for each information class and the final row
presents the overall, kappa and OOB accuracies for the entire classification.
It is interesting to note that all classes are classified rather accurately, with both user
and producer accuracy greater than 80% for all information clasees, except for the Rye
which has the lowest producer and user accuracy of all the classes. The low producer
accuracy suggests that the chances of a randomly chosen point in the field to have the
same class on the map is low. By comparing the RGB 11-class to the RGB 5-class it is
evident that in both cases Rye gets misclassified for either Bare Soil or Grass legume
lay, which is understandable since it is only present within the Grass legume lay class.
Bare Soil
Grass Leg. Lay Rye
Winter Wheat (orig)
Winter Wheat
(ref) Spring Barlay Rape-seed Kernza
Kernza/ Lucerne
Natural Grass Road
User acc:
Bare Soil 127 0 4 2 1 5 1 0 0 0 0 0.91
Grass Leg. Lay 0 22 3 0 0 0 0 0 1 0 0 0.85
Rye 0 3 9 0 0 0 0 0 0 0 0 0.75
Winter Wheat (orig) 0 1 0 10 0 0 0 0 0 0 0 0.91
Winter Wheat (ref) 0 0 0 0 11 0 0 0 0 0 0 1.00
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.00
Rapeseed 0 0 2 0 0 0 21 0 0 0 0 0.91
Kernza 0 0 0 0 0 0 0 22 1 0 0 0.96
Kernza/Lucerne 0 0 0 0 0 0 0 0 30 0 0 1.00
Natural Grass 0 0 0 0 0 0 0 0 0 11 0 1.00
Road 0 0 0 0 0 0 0 0 0 0 11 1.00
Producer acc.: 1.00 0.85 0.50 0.83 0.92 0.83 0.95 1.00 0.94 1.00 1.00
Overall acc: 0.93 Kappa: 0.91 OOB: 0.68
30
4.2. Specialized Case
The final classification case is classifying the intercropping system Kernza/Lucerne as
well as classifying dandelion (a yellow weed) which grows wildly within that field. The
Kernza and Lucerne were sown in systematic row-wise manner as seen from Figure 19.
It is hard to visually see the row-wise pattern on the UAV images, since these two
classes are spectrally similar, but that pattern is more clearin the photograph.
The classified maps are displayed in Figure 20 and are the results of the Random Forest
pixel-by-pixel classification produced using image bands, Mean texture feature (the
most important texture feature according to section 4.1.2.2) and VI as input. As seen on
Figure 20 the multispectral camera did a much better job in classifying the Lucerne
from the Kernza. The RGB image has captured a vague row-wise pattern of the
Kernza/Lucerne system whereas the multispectral image clearly captures the row by
row division. The confusion matricies for the classification are displayed in Appendix
B.
Figure 19: Location of the Kernza/Lucern field (top left) and zoom in on the RGB and multispectral image of the
area (bottom left). To the right is a photograph displaying the row-wise pattern of the Kernza and Lucerne. The
dandelions are also visible (Photo: Author).
31
It is hard to verify the classification accuracy of the dandelions except for visual
confirmation. Figure 21 displays a zoomed-in area of the Kernza/Lucerne field and only
pixels classified as dandelions are overlaid on the map. Visually it looks good and this
distinctive yellow weed really stands out. Both vegetation indices, NVDI and VARI,
are ranked as the most important feature. Lastly, Figure 22 displays two pie charts
showing the area coverage for each class. As seen from the chart, the multispectral
classification classifies more Kernza than Lucerne which is understandable as the
Kernza stood a little higher off the ground and is therefore more likely to cover the
Lucerne at certain places.
Figure 20: Classification maps for the Kernza/Lucerne intercropping system, produced using the pixel-by-pixel
classification method with pixel size equal to 5 cm. On the left is the classification map from the RGB camera and
on the righ is the classification map from the multispectral camera. Blue is Kernza, red is Lucerne and yellow is
Dandelion.
32
Figure 21: Zoom-in area of the Kernza/Lucerne field. At top are the original RGB and multispectral images where
dandelions are clearly visible (yellow in the RGB, bright white in the multispectral). At the bottom are the same
RGB and multispectral images which are overlaid with the classified dandelions pixels (orange areas).
Figure 22: Area division for the Kernza, Lucerne and Dandelion classes for the multispectral and RGB classification
results.
33
5. Discussion
The aim of this thesis was to compare RGB and multispectral UAV imagery using the
machine learning algorithm Random Forest. Even though the subtests designed for this
project, i.e. to evaluate the effect of pixel size, texture, segmentation etc., show very
promising results, there are a great many other factors which can influence the final
output and improvements can always be made.
5.1. General Thoughts
In general, the Random Forest algorithm produced very good results and some very
interesting trends are noticeable. Firstly, the main emphasize of this project was to
compare an RGB camera to a multispectral one. By running multiple classification
examples, it is clear that an RGB camera works very well in producing general
classification maps and usually provided similar or higher overall accuracies then the
multispectral one. However, RGB cameras run into problems when spectrally similar
objects coexist at proximity, as seen from the specialized Kernza/Lucerne intercropping
map. With that being said, the multispectral camera really demonstrated the importance
of having more wavelengths bands in these kinds of scenarios. However, for general
mapping purposes an RGB camera can do the job fine.
5.2. Effects of spatial resolution, segmentation and additional features
The pixel size proves to be very important in UAV image classification. Classifying at
higher spatial resolution produces the highest accuracy maps for all cases and both
sensors. Furthermore, by integrating addition data much higher classification accuracies
can be achieved (>10% overall accuracy difference). The effects of segmenting an
image before classification and integrating additional input features also improved the
classification accuracy. Table 7 displays the highest overall accuracy, Kappa and OOB
score for every subtest of the RGB 11-class classification. By taking the sum of all the
accuracies, the additional input features to the pixel-by-pixel technique ranks the
highest. However, the OOB score is in no agreement with the other accuracy values and
thus greatly affects the outcome.
As stated before, the segmented classification produces the highest map accuracy, but
it also works well on detecting edges between classes. It also produces a speckle free
map which is not the case for pixel-by-pixel classification maps. So, for a simple
overview map of the area, the segmentation method generates the most clean and
uniform maps.
34
Table 7: Accuracy comparison for every subtest of the 11-class classification. The highest overall accuracy (OA),
kappa and OOB score are displayed for each subtest, i.e. the pixel-by-pixel method with original image bands
(RGB), the pixel-by-pixel method with additional input features (RGB/VARI/Texture/DEM) and the three
segmentation with similarity tolerance equal to 50, 100 and 200 (RGB seg50, RGB seg100, RGB seg200).
RGB RGB/VARI/Texture/DEM RGB seg50 RGB seg100 RGB seg200
OA 0.77 0.89 0.85 0.89 0.93
Kappa 0.71 0.87 0.82 0.87 0.91
OOB 0.73 0.99 0.86 0.84 0.68
Total 2.21 2.75 2.53 2.6 2.52
5.3. UAV pre- and postprocessing
The first problem regarding this UAV classification experiment is that only one image
from the phenology cycle is used for analysis. It is hard to estimate how better or worse
it will be to classify classes at different crop stages. Therefore, it would have been
interesting to compare these results with UAV images acquired later in the summer to
see if any classes dramatically increase or decrease in accuracy.
In this thesis the classification was solely based on the raw digital number (DN) values,
i.e. no correction is made and the typical conversion from DN values to reflectance was
not performed in this study. This also means that these results cannot be directly
compared to images from other dates since rainfall, atmospheric- and sun condition etc.
all influence the DN value creation process. However, for this thesis purpose,
classifying using the DN values is fine since no comparison between flights is made.
Another factor which could affect the accuracy is the fact that Agisoft Photoscan uses
bilinear interpolation in creating the orthophotos, which are then used for classifying
the surface. However, the bilinear process interpolates the true spectral value, thus
producing a DN value which is not necessarily correct. Since the scope of this effect is
not known, it would be very interesting to classify the raw UAV images before they go
through the alignment and georeferencing phase in Agisoft and compare the results.
5.4. Accuracy evaluation
The highest ranked classified maps from all subtest have an overall accuracy greater
than 85%, which is often the threshold for what counts as a good map (Foody, 2002),
except for the RGB and multispectral pixel-by-pixel 11-class maps with overall
accuracy of just under 80%.
35
The OOB score extracted from the Random Forest algorithm does not always correlate
well with the overall accuracy (OA) or kappa values. Figure 23 displays all OA values
graphed against the OOB score for every classification. As seen from the graph, a rather
weak linear relationship exists for the OOB score and OA where it seems that for a
given Overall accuracy value the OOB score will be lower.
Figure 23: Relationship between Overall accuracy and OOB score for all classification examples
R² = 0.4917
0
0.2
0.4
0.6
0.8
1
1.2
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OO
B s
core
Overall accuracy
Overall accuracy and OOB score
36
By examining the same relationship, grouped by each classification type, the texture
classification generates correlated high OOB score and OA. The two highest OA
classified maps, the multispectral and RGB segmented classification, have considerably
lower OOB score.
Figure 24: Relationship between Overall accuracy and OOB score, grouped by each classification type, i.e. 5-class
classification (class5), 11-class classification (class11), segmentation classification (seg), classification with
additional features (texture) and the specialized Kernza/Lucerne classification (KL).
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
OO
B s
core
Overall accuracy
Overall accuracy and OOB score
class5
class11
seg
texture
KL
37
6. Conclusion
An RGB and multispectral camera sensors were tested and compared during an UAV
image classification process, where the target information classes consisted mostly of
different agricultural crop fields. Both RGB and multispectral cameras performed well,
reaching overall accuracies greater than 90%, and the RGB in many cases the better
option. However, when a more detailed classification was performed, classifying a field
containing two spectrally similar classes, the multispectral camera outperformed the
RGB.
The pixel size has a huge impact on resulting classification accuracy, where overall
accuracy varies from e.g. 58% (1 m pixel size) to 88% (5 cm pixel size) for the RGB 5-
class classification. For both RGB and multispectral cameras, higher accuracies are
achieved at higher spatial resolutions. By integrating addition data sources in the pixel-
by-pixel classification method, overall accuracies increase by a factor of >10%, e.g.
from 77% to 88% for the 11-class multispectral classification. The Mean – texture
feature turns out to be the most important texture feature for both cameras.
The highest accuracy for both RGB and multispectral classification was achieved by
classifying groups of pixels into segments where larger segments generated higher
classification accuracies, differing by a factor of 10% for the largest segments compared
to the smallest segments classification.
So, overall the objectives were all met and successfully answered. UAV works really
well for agricultural farm mapping and is a very good monitoring tool due to its quick
deployment, ease of data collection and accurate results.
38
39
7. References
Agisoft PhotoScan User Manual: Professional Edition, Version 1.4 (2018). Agisoft
LLC.
Baron , J., Hill, D. J., & Elmiligi, H. (2018). Combining image processing and machine
learning to identify invasive plants in high-resolution images. International
Journal of Remote Sensing, 0(0), 1–20.
https://doi.org/10.1080/01431161.2017.1420940
Blaschke, T. (2010). Object based image analysis for remote sensing. ISPRS Journal of
Photogrammetry and Remote Sensing, 65(1), 2–16.
https://doi.org/10.1016/j.isprsjprs.2009.06.004
Chuvieco, E. (2016). Fundamentals of Satellite Remote Sensing. An Environmental
Approach. (2nd ed.). Boca Raton, FL: CRC Press.
Congalton, R. G. (1991). A review of assessing the accuracy of classifications of
remotely sensed data. Remote Sensing of Environment, 37(1), 35–46.
https://doi.org/10.1016/0034-4257(91)90048-B
Feng, Q., Liu, J., & Gong, J. (2015). UAV Remote Sensing for Urban Vegetation
Mapping Using Random Forest and Texture Analysis. Remote Sensing, 7(1),
1074–1094. https://doi.org/10.3390/rs70101074
Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote
Sensing of Environment, 80(1), 185–201. https://doi.org/10.1016/S0034-
4257(01)00295-4
Franklin, S. E., & Ahmed, O. S. (2017). Deciduous tree species classification using
object-based analysis and machine learning with unmanned aerial vehicle
multispectral data. International Journal of Remote Sensing, 0(0), 1–10.
https://doi.org/10.1080/01431161.2017.1363442
Haralick, R. M., Shanmugam, K., & Dinstin, I. (1973). Textural features for image
classification. IEEE Transactions on Systemst. Man and Cybernetics, 3, 610–
621.
Huete, A., Didan, K., Miura, T., Rodriguez, E. P., Gao, X., & Ferreira, L. G. (2002).
Overview of the radiometric and biophysical performance of the MODIS
vegetation indices. Remote Sensing of Environment, 83(1), 195–213.
https://doi.org/10.1016/S0034-4257(02)00096-2
Laliberte, A. S., & Rango, A. (2009). Texture and Scale in Object-Based Analysis of
Subdecimeter Resolution Unmanned Aerial Vehicle (UAV) Imagery. IEEE
Transactions on Geoscience and Remote Sensing, 47(3), 761–770.
https://doi.org/10.1109/TGRS.2008.2009355
Matese, A., Gennaro, S. F. D., Miranda, C., Berton, A., & Santesteban, L. G. (2017).
Evaluation of spectral-based and canopy-based vegetation indices from UAV
and Sentinel 2 images to assess spatial variability and ground vine parameters.
40
Advances in Animal Biosciences, 8(2), 817–822.
https://doi.org/10.1017/S2040470017000929
Milas, A. S., Arend, K., Mayer, C., Simonson, M. A., & Mackey, S. (2017). Different
colours of shadows: classification of UAV images. International Journal of
Remote Sensing, 38(8–10), 3084–3100.
https://doi.org/10.1080/01431161.2016.1274449
Millard, K., & Richardson, M. (2015). On the Importance of Training Data Sample
Selection in Random Forest Image Classification: A Case Study in Peatland
Ecosystem Mapping. Remote Sensing, 7, 8489–8515.
https://doi.org/10.3390/rs70708489
Özdogan, M. (2015). Image Classification Methods in Land Cover and Land Use. In
Remotly sensed data characterization, classification and accuracies (pp. 231–
246). Boca Raton: FL: CRC Press.
Pal, M. (2005). Random forest classifier for remote sensing classification. International
Journal of Remote Sensing, 26(1), 217–222.
https://doi.org/10.1080/01431160412331269698
Pande-Chhetri, R., Abd-Elrahman, A., Liu, T., Morton, J., & Wilhelm, V. L. (2017).
Object-based classification of wetland vegetation using very high-resolution
unmanned air system imagery. European Journal of Remote Sensing, 50(1),
564–576. https://doi.org/10.1080/22797254.2017.1373602
Puliti, S., Talbot, B., & Astrup, R. (2018). Tree-Stump Detection, Segmentation,
Classification, and Measurement Using Unmanned Aerial Vehicle (UAV)
Imagery. Forests, 9(3), 102. https://doi.org/10.3390/f9030102
Rango, A., Laliberte, A., Herrick, J. E., Winters, C., Havstad, K., Steele, C., &
Browning, D. (2009). Unmanned aerial vehicle-based remote sensing for
rangeland assessment, monitoring, and management. Journal of Applied Remote
Sensing, 3(1), 033542. https://doi.org/10.1117/1.3216822
Rodriguez-Galiano, V. F., Ghimire, B., Rogan, J., Chica-Olmo, M., & Rigol-Sanchez,
J. P. (2012). An assessment of the effectiveness of a random forest classifier for
land-cover classification. ISPRS Journal of Photogrammetry and Remote
Sensing, 67, 93–104. https://doi.org/10.1016/j.isprsjprs.2011.11.002
van der Werff, H. M. A., & van der Meer, F. D. (2008). Shape-based classification of
spectrally identical objects. ISPRS Journal of Photogrammetry and Remote
Sensing, 63(2), 251–258. https://doi.org/10.1016/j.isprsjprs.2007.09.007
Zhou, X., Zheng, H. B., Xu, X. Q., He, J. Y., Ge, X. K., Yao, X., … Tian, Y. C. (2017).
Predicting grain yield in rice using multi-temporal vegetation indices from
UAV-based multispectral and digital imagery. ISPRS Journal of
Photogrammetry and Remote Sensing, 130, 246–255.
https://doi.org/10.1016/j.isprsjprs.2017.05.003
41
8. Appendices
8.1. Appendix A: Classification results, images.
RGB 5-class classification. Pixel-by-pixel classification.
Figure A 1: RGB 5-class classification, pixel size: 5 cm
Figure A 2: RGB 5-class classification, pixel size: 10 cm
42
Figure A 3: RGB 5-class classification, pixel size: 20 cm
Figure A 4: RGB 5-class classification, pixel size: 30 cm
43
Figure A 5: RGB 5-class classification, pixel size: 40 cm
Figure A 6: RGB 5-class classification, pixel size: 50 cm
44
Figure A 7: RGB 5-class classification, pixel size: 100 cm
45
Multispectral 5-class classification. Pixel-by-pixel classification.
Figure A 8: Multispectral 5-class classification, pixel size: 5 cm
Figure A 9: Multispectral 5-class classification, pixel size: 10 cm
46
Figure A 10: Multispectral 5-class classification, pixel size: 20 cm
Figure A 11: Multispectral 5-class classification, pixel size: 30 cm
47
Figure A 12: Multispectral 5-class classification, pixel size: 40 cm
Figure A 13: Multispectral 5-class classification, pixel size: 50 cm
48
Figure A 14: Multispectral 5-class classification, pixel size: 100 cm
49
RGB 11-class classification. Pixel-by-pixel classification.
Figure A 15: RGB 11-class classification, pixel size: 5 cm
Figure A 16: RGB 11-class classification, pixel size: 10 cm
50
Figure A 17: RGB 11-class classification, pixel size: 20 cm
Figure A 18: RGB 11-class classification, pixel size: 30 cm
51
Figure A 19: RGB 11-class classification, pixel size: 40 cm
Figure A 20: RGB 11-class classification, pixel size: 50 cm
52
Figure A 21: RGB 11-class classification, pixel size: 100 cm
53
Multispectral 11-class classification. Pixel-by-pixel classification.
Figure A 22: Multispectral 11-class classification, pixel size: 5 cm
Figure A 23: Multispectral 11-class classification, pixel size: 10 cm
54
Figure A 24: Multispectral 11-class classification, pixel size: 20 cm
Figure A 25: Multispectral 11-class classification, pixel size: 30 cm
55
Figure A 26: Multispectral 11-class classification, pixel size: 40 cm
Figure A 27: Multispectral 11-class classification, pixel size: 50 cm
56
Figure A 28: Multispectral 11-class classification, pixel size: 100 cm
57
RGB 11-class classification, pixel size equal to 5 cm. Integration of addition data;
Texture[Corr, Entr, IDM, Mean, Sum_avg, Var], VARI, DEM
Figure A 29: RGB 11-class classification. Features: B, G, R, Texture[corr, entr, IDM, mean, sum_avg, var], VARI,
DEM
Figure A 30: RGB 11-class classification. Features: Texture [ entr, mean, sum_avg, var], VARI, DEM
58
Figure A 31: RGB 11-class classification. Features: Texture [ mean, sum_avg], DEM
Figure A 32: RGB 11-class classification. Features: Texture [ mean], DEM
59
Multispectral 11-class classification, pixel size equal to 5 cm. Integration of addition
data; Texture[Corr, Entr, IDM, Mean, Sum_avg, Var], NDVI, DEM
Figure A 33: Multispectral 11-class classification. Features: G, R, R edge, NIR, Texture [corr, entr, IDM, mean,
sum_avg, var], VARI, DEM
Figure A 34: Multispectral 11-class classification. Features: G, R, Texture [ mean, sum_avg], VARI, DEM
60
Figure A 35: Multispectral 11-class classification. Features: R, Texture [ mean], DEM
Figure A 36: Multispectral 11-class classification. Features: R, DEM
61
Figure A 37: Multispectral 11-class classification. Features: DEM
62
RGB 11-class classification, pixel size equal to 5 cm. Segmentation.
Figure A 38: RGB 11-classification, segmentation (similarity tolerance: 50)
Figure A 39: RGB 11-classification, segmentation (similarity tolerance: 100)
63
Figure A 40: RGB 11-classification, segmentation (similarity tolerance: 200)
64
Multispectral 11-class classification, pixel size equal to 5 cm. Segmentation.
Figure A 41: Multispectral 11-class classification, segmentation (similarity tolerance: 50).
Figure A 42: Multispectral 11-class classification, segmentation (similarity tolerance: 100).
65
Figure A 43: Multispectral 11-class classification, segmentation (similarity tolerance: 200).
66
RGB and Multispectral classification – Kernza, Lucerne and Dandelions (pizel size
equal to 5 cm)
Figure A 44: RGB classification; Kernza, Lucerne and Dandelion
Figure A 45: Multispectral classification; Kernza, Lucerne and Dandelion
67
8.2. Appendix B: Classification results, confusion matrices.
RGB 5-class classification. Pixel-by-pixel classification.
Table B 1:RGB 5-class classification, pixel size: 5cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 48 0 2 0 0 0.96 Vegetation 7 75 4 8 0 0.80 Rye 0 0 10 1 0 0.91 Rapeseed 0 1 1 12 0 0.86
Road 0 0 0 0 11 1
Producer acc.: 0.87 0.99 0.59 0.57 1 Overall acc: 0.87 Kappa: 0.80 OOB: 0.92
Table B 2: RGB 5-class classification, pixel size: 10cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 44 1 5 0 0 0.88 Vegetation 3 82 2 7 0 0.87 Rye 0 0 10 0 1 0.91 Rapeseed 1 0 0 13 0 0.93
Road 1 0 0 0 10 0.91
Producer acc.: 0.9 0.99 0.59 0.65 0.91
Overall acc: 0.88 Kappa: 0.82 OOB: 0.88
Table B 3: RGB 5-class classification, pixel size: 20 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 44 0 6 0 0 0.88 Vegetation 2 80 0 12 0 0.85 Rye 1 1 9 0 0 0.81 Rapeseed 0 1 0 13 0 0.93
Road 0 0 1 0 10 0.91
Producer acc.: 0.94 0.98 0.56 0.52 1
Overall acc: 0.87 Kappa: 0.80 OOB: 0.83
68
Table B 4: RGB 5-class classification, pixel size: 30 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 43 1 4 2 0 0.86 Vegetation 3 70 6 15 0 0.74 Rye 1 0 7 2 1 0.64 Rapeseed 0 3 1 10 0 0.71 Road 1 0 1 0 9 0.82
Producer acc.: 0.9 0.95 0.37 0.34 0.9
Overall acc: 0.77 Kappa: 0.67 OOB: 0.71
Table B 5: RGB 5-class classification, pixel size: 40 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 46 3 1 0 0 0.92 Vegetation 4 70 5 15 0 0.74
Rye 3 0 7 1 0 0.64 Rapeseed 0 2 2 10 0 0.71 Road 0 0 1 0 10 0.91
Producer acc.: 0.86 0.93 0.44 0.38 1
Overall acc: 0.79 Kappa: 0.70 OOB: 0.76
Table B 6: RGB 5-class classification, pixel size: 50 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 42 0 4 4 0 0.84 Vegetation 3 72 2 17 0 0.77 Rye 2 1 7 0 1 0.64 Rapeseed 0 1 1 12 0 0.86 Road 0 0 0 0 11 1
Producer acc.: 0.89 0.97 0.5 0.36 0.92
Overall acc: 0.8 Kappa: 0.71 OOB: 0.83
Table B 7: RGB 5-class classification, pixel size: 100 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 15 6 2 27 0 0.3 Vegetation 2 75 0 17 0 0.8 Rye 2 5 1 2 1 0.1 Rapeseed 1 7 0 3 3 0.21 Road 0 0 0 0 11 1
Producer acc.: 0.75 0.80 0.33 0.1 0.73
Overall acc: 0.58 Kappa: 0.38 OOB: 0.73
69
Multispectral 5-class classification. Pixel-by-pixel classification.
Table B 8: Multispectral 5-class classification. Pixel size: 5 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 46 0 0 4 0 0.92
Vegetation 7 75 3 9 0 0.78
Rye 0 0 11 0 0 1.0
Rapeseed 0 0 1 13 0 0.93
Road 1 0 0 0 10 0.91
Producer acc.: 0.85 1.0 0.73 0.5 1.0
Overall acc: 0.86 Kappa: 0.8 OOB: 0.92
Table B 9: Multispectral 5-class classification. Pixel size: 10 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 43 0 3 4 0 0.86
Vegetation 4 73 2 15 0 0.78
Rye 0 0 11 0 0 1.0
Rapeseed 1 1 0 12 0 0.86
Road 2 0 0 0 9 0.82
Producer acc.: 0.86 0.99 0.69 0.39 1.0
Overall acc: 0.82 Kappa: 0.74 OOB: 0.9
Table B 10: Multispectral 5-class classification. Pixel size: 20 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 42 0 5 3 0 0.84
Vegetation 6 75 1 12 0 0.8
Rye 1 0 9 1 0 0.81 Rapeseed 3 2 1 8 0 0.57
Road 0 0 1 0 10 0.91
Producer acc.: 0.81 0.97 0.53 0.33 1.0
Overall acc: 0.8 Kappa: 0.7 OOB: 0.88
Table B 11: Multispectral 5-class classification. Pixel size: 30 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 40 4 3 3 0 0.8
Vegetation 1 73 3 17 0 0.78
Rye 0 0 11 0 0 1 Rapeseed 1 1 0 12 0 0.86
Road 1 0 1 0 9 0.82
Producer acc.: 0.93 0.94 0.61 0.38 1.0
Overall acc: 0.81 Kappa: 0.72 OOB: 0.81
70
Table B 12: Multispectral 5-class classification. Pixel size: 40 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 40 6 3 1 0 0.8
Vegetation 4 76 0 14 0 0.81
Rye 1 2 8 0 0 0.73 Rapeseed 1 1 2 10 0 0.71
Road 0 0 1 0 10 0.91
Producer acc.: 0.87 0.89 0.57 0.4 1.0
Overall acc: 0.8 Kappa: 0.7 OOB: 0.81
Table B 13: Multispectral 5-class classification. Pixel Size: 50 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 43 2 2 3 0 0.86
Vegetation 5 70 6 13 0 0.74
Rye 3 1 7 0 0 0.64 Rapeseed 1 2 2 9 0 0.64
Road 1 0 0 0 10 0.91
Producer acc.: 0.81 0.93 0.41 0.36 1.0
Overall acc: 0.77 Kappa: 0.67 OOB: 0.75
Table B 14: Multispectral 5-class classification. Pixel size: 100 cm
Bare Soil Vegetation Rye Rapeseed Road User acc:
Bare Soil 34 1 1 14 0 0.68
Vegetation 4 75 0 14 1 0.8
Rye 2 4 3 2 0 0.27 Rapeseed 1 2 0 10 1 0.71
Road 1 1 1 0 8 0.73
Producer acc.: 0.81 0.9 0.6 0.25 0.8
Overall acc: 0.72 Kappa: 0.59 OOB: 0.73
71
RGB 11-class classification. Pixel-by-pixel classification.
Table B 15: RGB 11-class classification. Pixel size: 5 cm
Table B 16: RGB 11-class classification, pixel size: 10 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 121 0 4 0 2 11 0 0 1 0 1 0.86
Grass Leg. Lay 0 14 0 1 0 0 0 2 7 2 0 0.54
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 4 1 0 0 5 1 0 0 0.36
Winter Wheat (ref) 0 0 0 0 7 0 0 0 4 0 0 0.64
Spring Barlay 1 0 0 0 0 23 1 0 0 0 0 0.92
Rapeseed 0 0 2 1 0 0 18 0 0 2 0 0.78
Kernza 0 0 0 7 0 0 1 13 1 1 0 0.57
Kernza/Lucerne 0 3 0 1 7 0 0 0 18 1 0 0.6
Natural Grass 0 2 0 0 0 0 0 1 0 8 0 0.73
Road 0 0 1 0 0 0 0 0 0 0 10 0.91
Producer acc.: 1 0.74 0.63 0.29 0.412 0.68 0.9 0.62 0.56 0.57 0.91
Overall acc: 0.77 Kappa: 0.71 OOB: 0.73
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 111 0 8 1 0 15 1 0 2 0 2 0.79
Grass Leg. Lay 0 12 0 2 1 0 0 6 4 1 0 0.46
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 5 0 0 0 4 2 0 0 0.45
Winter Wheat (ref) 0 0 0 0 9 1 0 0 1 0 0 0.82
Spring Barlay 0 0 0 0 1 24 0 0 0 0 0 0.96
Rapeseed 0 0 0 0 1 2 15 3 0 2 0 0.65
Kernza 0 1 1 9 0 0 1 9 1 1 0 0.39
Kernza/Lucerne 0 2 0 2 3 0 0 1 21 1 0 0.7
Natural Grass 0 1 0 1 0 0 0 0 2 7 0 0.64
Road 0 0 1 0 0 0 0 0 0 0 10 0.91
Producer acc.: 1.0 0.75 0.55 0.25 0.6 0.57 0.88 0.39 0.64 0.58 0.83
Overall acc: 0.73 Kappa: 0.66 OOB: 0.68
72
Table B 17: RGB 11-class classification, pixel size: 20 cm
Table B 18: RGB 11-class classification, pixel size: 30 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 104 0 7 0 1 24 1 1 1 0 1 0.71
Grass Leg. Lay 0 15 0 1 2 0 0 3 3 2 0 0.5
Rye 0 0 12 0 0 0 0 0 0 0 0 0.75
Winter Wheat (orig) 0 1 1 4 0 0 1 3 1 0 0 0.36
Winter Wheat (ref) 0 0 0 1 7 0 0 0 3 0 0 0.55
Spring Barlay 0 0 0 0 1 24 0 0 0 0 0 0.88
Rapeseed 0 0 0 0 1 1 18 3 0 0 0 0.65
Kernza 0 1 2 4 1 0 2 12 1 0 0 0.48
Kernza/Lucerne 0 5 0 4 2 0 0 0 18 1 0 0.5
Natural Grass 0 2 0 0 0 0 0 0 0 9 0 0.64
Road 0 0 2 0 0 0 0 0 0 0 9 0.82
Producer acc.: 0.97 0.5 0.47 0.21 0.3 0.42 0.8 0.46 0.75 0.7 0.75
Overall acc: 0.65 Kappa: 0.57 OOB: 0.55
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 99 0 7 1 1 30 0 1 0 0 1 0.71
Grass Leg. Lay 0 13 0 2 2 0 2 4 2 1 0 0.5
Rye 0 0 9 0 0 0 0 2 0 0 1 0.75
Winter Wheat (orig) 0 1 0 4 2 0 1 2 0 1 0 0.36
Winter Wheat (ref) 0 1 0 3 6 0 0 0 1 0 0 0.55
Spring Barlay 2 0 0 0 0 22 0 1 0 0 0 0.88
Rapeseed 1 1 1 1 1 1 15 0 1 0 1 0.65
Kernza 0 2 0 7 1 0 0 11 1 1 0 0.48
Kernza/Lucerne 0 5 0 1 6 0 0 3 15 0 0 0.5
Natural Grass 0 3 0 0 1 0 0 0 0 7 0 0.64
Road 0 0 2 0 0 0 0 0 0 0 9 0.82
Producer acc.: 0.97 0.5 0.47 0.21 0.3 0.42 0.83 0.46 0.75 0.7 0.75
Overall acc: 0.65 Kappa: 0.57 OOB: 0.55
73
Table B 19: RGB 11-class classification, pixel size: 40 cm
Table B 20: RGB 11-class classification, pixel size: 50 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 97 1 10 0 2 29 0 0 0 0 1 0.69
Grass Leg. Lay 0 17 0 2 0 0 0 0 5 2 0 0.65
Rye 0 0 7 1 0 0 0 0 0 0 4 0.58
Winter Wheat (orig) 0 1 0 5 0 0 1 3 1 0 0 0.45
Winter Wheat (ref) 0 1 0 0 10 0 0 0 0 0 0 0.91
Spring Barlay 2 0 0 0 0 23 0 0 0 0 0 0.92
Rapeseed 0 2 6 0 0 1 10 2 0 2 0 0.43
Kernza 0 5 0 6 1 1 1 9 0 0 0 0.39
Kernza/Lucerne 0 2 0 1 7 2 0 1 15 2 0 0.5
Natural Grass 0 1 0 0 1 0 1 1 1 6 0 0.55
Road 0 0 1 0 0 0 0 0 0 0 10 0.91
Producer acc.: 0.98 0.57 0.29 0.33 0.48 0.41 0.77 0.56 0.68 0.5 0.67
Overall acc: 0.65 Kappa: 0.57 OOB: 0.54
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 102 0 7 0 2 29 0 0 0 0 0 0.73
Grass Leg. Lay 0 13 0 1 1 0 0 2 6 3 0 0.5
Rye 1 0 8 1 0 0 0 0 0 0 2 0.67
Winter Wheat (orig) 0 1 0 5 1 0 0 4 0 0 0 0.45
Winter Wheat (ref) 0 0 0 0 10 0 0 0 1 0 0 0.91
Spring Barlay 1 0 0 0 1 23 0 0 0 0 0 0.92
Rapeseed 1 1 1 0 0 1 17 1 1 0 0 0.74
Kernza 0 2 0 7 0 0 0 13 0 1 0 0.57
Kernza/Lucerne 0 4 0 0 3 1 0 3 19 0 0 0.63
Natural Grass 0 3 0 0 0 0 0 1 0 7 0 0.64
Road 0 0 1 0 0 0 0 0 0 0 10 0.91
Producer acc.: 0.97 0.54 0.47 0.36 0.56 0.43 1.0 0.54 0.7 0.64 0.83
Overall acc: 0.7 Kappa: 0.64 OOB: 0.63
74
Table B 21: RGB 11-class classification, pixel size: 100 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 33 0 24 1 0 78 1 0 1 1 1 0.24
Grass Leg. Lay 0 6 1 2 0 5 0 5 1 6 0 0.23
Rye 0 0 3 2 0 3 1 1 1 0 1 0.25
Winter Wheat (orig) 0 3 0 1 0 1 0 3 1 2 0 0.1
Winter Wheat (ref) 0 0 0 0 0 0 0 1 7 3 0 0.0
Spring Barlay 1 0 0 2 0 22 0 0 0 0 0 0.88
Rapeseed 0 0 2 1 0 5 6 0 0 6 3 0.26
Kernza 0 3 0 8 1 0 0 9 0 2 0 0.39
Kernza/Lucerne 0 5 0 0 0 3 1 5 10 6 0 0.33
Natural Grass 0 0 0 0 0 1 0 1 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.97 0.35 0.1 0.1 0.0 0.19 0.67 0.36 0.48 0.26 0.69
Overall acc: 0.34 Kappa: 0.27 OOB: 0.45
75
Multispectral 11-class classification. Pixel-by-pixel classification.
Table B 22: Multispectral 11-class classification, pixel size: 5 cm
Table B 23: Multispectral 11-class classification, pixel size: 10 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 110 0 3 0 0 20 3 2 0 2 0 0.79
Grass Leg. Lay 0 16 0 4 0 0 0 0 4 2 0 0.62
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 1 0 4 2 2 0 1 1 0 0 0.36
Winter Wheat (ref) 0 2 0 3 5 0 0 1 0 0 0 0.45
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 1 0 0 0 0 0 21 1 0 0 0 0.91
Kernza 0 0 0 6 0 1 0 16 0 0 0 0.7
Kernza/Lucerne 0 1 0 3 3 0 0 4 19 0 0 0.63
Natural Grass 0 0 0 0 1 0 0 1 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.99 0.8 0.8 0.2 0.45 0.52 0.88 0.62 0.79 0.69 1.0
Overall acc: 0.77 Kappa: 0.71 OOB: 0.71
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 112 0 2 0 0 17 4 3 0 2 0 0.8
Grass Leg. Lay 0 14 0 3 0 0 0 0 9 0 0 0.54
Rye 0 0 11 0 0 0 0 0 0 1 0 0.92
Winter Wheat (orig) 0 1 0 3 4 2 0 0 1 0 0 0.27
Winter Wheat (ref) 0 4 0 2 4 0 0 1 0 0 0 0.36
Spring Barlay 2 0 0 2 1 17 0 3 0 0 0 0.68
Rapeseed 0 0 0 1 0 2 16 3 1 0 0 0.7
Kernza 1 0 0 5 3 1 1 12 0 0 0 0.52
Kernza/Lucerne 0 1 0 3 1 0 0 3 21 1 0 0.7
Natural Grass 0 1 0 0 1 0 0 0 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.97 0.67 0.85 0.16 0.29 0.44 0.76 0.48 0.66 0.69 1.0
Overall acc: 0.71 Kappa: 0.64 OOB: 0.69
76
Table B 24: Multispectral 11-class classification, pixel size: 20 cm
Table B 25: Multispectral 11-class classification, pixel size: 30 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 111 0 1 0 0 20 5 0 0 2 1 0.79
Grass Leg. Lay 0 13 0 0 4 0 0 2 6 1 0 0.5
Rye 0 0 11 0 0 0 0 0 1 0 0 0.92
Winter Wheat (orig) 0 2 0 3 1 1 0 2 1 1 0 0.27
Winter Wheat (ref) 0 0 0 2 9 0 0 0 0 0 0 0.82
Spring Barlay 0 0 0 2 0 18 1 4 0 0 0 0.72
Rapeseed 0 0 1 0 0 3 17 2 0 0 0 0.74
Kernza 0 0 2 2 1 3 1 13 1 0 0 0.57
Kernza/Lucerne 0 1 0 2 3 1 2 2 19 0 0 0.63
Natural Grass 0 0 0 0 0 0 0 1 0 10 0 0.91
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 1.0 0.81 0.73 0.27 0.5 0.39 0.65 0.5 0.68 0.71 0.92
Overall acc: 0.73 Kappa: 0.66 OOB: 0.63
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 100 0 3 0 0 17 16 2 0 2 0 0.71
Grass Leg. Lay 0 11 0 5 2 0 0 1 6 1 0 0.42
Rye 0 0 8 0 0 0 3 1 0 0 0 0.67
Winter Wheat (orig) 0 3 0 3 1 2 0 1 1 0 0 0.27
Winter Wheat (ref) 0 1 0 2 6 0 0 0 1 1 0 0.55
Spring Barlay 1 0 0 0 0 21 0 3 0 0 0 0.84
Rapeseed 2 0 2 0 0 2 15 2 0 0 0 0.65
Kernza 0 0 0 2 0 3 0 15 2 1 0 0.65
Kernza/Lucerne 0 2 0 2 1 0 1 7 17 0 0 0.57
Natural Grass 0 0 0 0 1 0 0 0 0 10 0 0.91
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.97 0.65 0.62 0.21 0.55 0.47 0.43 0.47 0.63 0.67 1.0
Overall acc: 0.67 Kappa: 0.6 OOB: 0.62
77
Table B 26: Multispectral 11-class classification, pixel size: 40 cm
Table B 27: Multispectral 11-class classification, pixel size: 50 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 99 1 3 0 0 22 8 5 1 1 0 0.71
Grass Leg. Lay 0 12 1 3 3 0 0 0 7 0 0 0.46
Rye 0 1 7 0 0 1 1 0 1 0 1 0.58
Winter Wheat (orig) 0 0 0 2 5 1 1 0 2 0 0 0.18
Winter Wheat (ref) 0 0 0 0 7 0 0 1 2 1 0 0.64
Spring Barlay 2 0 0 1 1 21 0 0 0 0 0 0.84
Rapeseed 1 0 1 2 0 1 16 1 0 1 0 0.7
Kernza 0 0 0 3 3 4 0 11 2 0 0 0.48
Kernza/Lucerne 0 1 0 1 3 0 0 2 22 1 0 0.73
Natural Grass 0 0 0 1 1 0 0 0 1 8 0 0.73
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.97 0.8 0.58 0.15 0.30 0.42 0.62 0.55 0.58 0.67 0.92
Overall acc: 0.67 Kappa: 0.6 OOB: 0.58
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 98 1 4 1 0 35 1 0 0 0 0 0.7
Grass Leg. Lay 0 15 0 3 7 0 0 0 0 1 0 0.58
Rye 0 1 9 0 0 0 1 0 0 1 0 0.75
Winter Wheat (orig) 0 1 0 4 2 1 0 2 1 0 0 0.36
Winter Wheat (ref) 0 0 0 3 8 0 0 0 0 0 0 0.73
Spring Barlay 3 0 0 1 0 19 0 2 0 0 0 0.76
Rapeseed 1 0 3 1 0 0 16 2 0 0 0 0.7
Kernza 2 0 1 3 3 3 1 9 1 0 0 0.39
Kernza/Lucerne 0 2 1 4 2 1 0 3 17 0 0 0.57
Natural Grass 0 0 0 0 1 0 0 1 1 8 0 0.73
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.94 0.75 0.5 0.2 0.35 0.32 0.84 0.47 0.85 0.8 1.0
Overall acc: 0.66 Kappa: 0.59 OOB: 0.49
78
Table B 28: Multispectral 11-class classification, pixel size: 100 cm
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 59 0 13 0 0 41 25 1 0 1 0 0.42
Grass Leg. Lay 0 9 0 1 0 0 1 1 7 7 0 0.35
Rye 0 1 3 0 0 1 2 1 1 3 0 0.25
Winter Wheat (orig) 0 3 0 0 0 0 0 6 1 1 0 0.0
Winter Wheat (ref) 0 2 0 2 0 0 0 1 0 6 0 0.0
Spring Barlay 1 0 0 0 0 17 1 4 2 0 0 0.68
Rapeseed 2 0 2 0 0 2 7 9 0 0 1 0.3
Kernza 0 1 0 1 0 9 0 8 4 0 0 0.35
Kernza/Lucerne 0 4 0 3 0 1 2 4 12 3 1 0.4
Natural Grass 0 0 1 1 0 0 0 2 0 7 0 0.64
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.95 0.45 0.16 0.0 0 0.24 0.18 0.22 0.44 0.25 0.85
Overall acc: 0.41 Kappa: 0.32 OOB: 0.26
79
RGB 11-class classification, pixel size equal to 5 cm. Pixel-by-pixel classification
including Texture (Correlation, Entropy, IDM, Mean, Sum_average, Variance), NDVI
and DEM.
Table B 29: RGB 11-class classification, pixel size: 5 cm. Features: B, G, R, Texture [Corr, Ent, IDM, Mean, Sum_Avg,
Var], VARI, DEM
Table B 30: RGB 11-class classification, pixel size: 5 cm. Features: Texture [ Ent, Mean, Sum_Avg, Var], VARI, DEM
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 119 3 6 0 0 12 0 0 0 0 0 0.85
Grass Leg. Lay 0 26 0 0 0 0 0 0 0 0 0 1.0
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 8 0 0 0 0 0 3 0 0.73
Winter Wheat (ref) 0 1 0 0 10 0 0 0 0 0 0 0.91
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 1 0 0 0 0 0 22 0 0 0 0 0.96
Kernza 0 0 0 4 0 0 0 19 0 0 0 0.83
Kernza/Lucerne 0 0 0 0 0 0 0 2 28 0 0 0.93
Natural Grass 0 2 0 0 0 0 0 0 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.99 0.81 0.67 0.67 1.0 0.68 1.0 0.9 1.0 0.75 1.0
Overall acc: 0.89 Kappa: 0.87 OOB: 0.99
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 119 3 7 0 0 11 0 0 0 0 0 0.85
Grass Leg. Lay 0 25 0 1 0 0 0 0 0 0 0 0.96
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 8 0 0 0 0 0 3 0 0.73
Winter Wheat (ref) 0 1 0 0 10 0 0 0 0 0 0 0.91
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 0 0 0 0 0 0 23 0 0 0 0 1.0
Kernza 0 0 0 4 0 0 0 19 0 0 0 0.83
Kernza/Lucerne 0 0 0 0 0 0 0 2 28 0 0 0.93
Natural Grass 0 2 0 0 0 0 0 0 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 1.0 0.81 0.63 0.62 1.0 0.69 1.0 0.9 1.0 0.75 1.0
Overall acc: 0.89 Kappa: 0.87 OOB: 0.99
80
Table B 31: RGB 11-class classification, pixel size: 5 cm. Features: Texture [ Mean, Sum_Avg,], DEM
Table B 32: RGB 11-class classification, pixel size: 5 cm. Features: Texture [ Mean,], DEM
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 114 2 12 1 0 11 0 0 0 0 0 0.81
Grass Leg. Lay 0 23 0 3 0 0 0 0 0 0 0 0.88
Rye 1 0 11 0 0 0 0 0 0 0 0 0.92
Winter Wheat (orig) 0 0 1 8 0 0 0 0 0 2 0 0.73
Winter Wheat (ref) 0 2 0 0 8 0 0 0 0 1 0 0.73
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 2 0 0 0 0 0 21 0 0 0 0 0.91
Kernza 0 0 0 2 0 0 0 21 0 0 0 0.91
Kernza/Lucerne 0 0 0 0 0 0 0 3 27 0 0 0.9
Natural Grass 0 2 0 0 1 0 0 0 0 8 0 0.73
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.97 0.79 0.46 0.57 0.89 0.69 1.0 0.88 1.0 0.73 1.0
Overall acc: 0.86 Kappa: 0.82 OOB: 0.98
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 105 2 25 1 0 5 2 0 0 0 0 0.75
Grass Leg. Lay 0 20 0 3 3 0 0 0 0 0 0 0.77
Rye 1 0 11 0 0 0 0 0 0 0 0 0.92
Winter Wheat (orig) 0 0 0 7 0 1 0 0 0 3 0 0.64
Winter Wheat (ref) 0 2 0 0 8 0 0 0 0 1 0 0.73
Spring Barlay 0 2 0 0 0 23 0 0 0 0 0 0.92
Rapeseed 1 0 0 0 0 0 22 0 0 0 0 0.96
Kernza 0 0 0 2 0 0 3 17 1 0 0 0.74
Kernza/Lucerne 0 0 0 0 0 0 0 4 26 0 0 0.87
Natural Grass 0 1 0 0 3 0 0 0 0 7 0 0.64
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.98 0.74 0.31 0.54 0.57 0.79 0.81 0.81 0.96 0.64 1.0
Overall acc: 0.8 Kappa: 0.75 OOB: 0.96
81
Multispectral 11-class classification, pixel size equal to 5 cm. Pixel-by-pixel
classification including Texture (Correlation, Entropy, IDM, Mean, Sum_average,
Variance), NDVI and DEM.
Table B 33: Multispectral 11-class classification, pixel size: 5 cm. Features: G, R, R edge, NIR, Texture [Corr, Ent, IDM,
Mean, Sum_Avg, Var], NDVI, DEM
Table B 34: Multispectral 11-class classification, pixel size: 5 cm. Features: G, R, Texture [ Mean, Sum_Avgr], NDVI,
DEM
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 119 0 3 0 0 16 0 0 0 2 0 0.85
Grass Leg. Lay 0 22 0 1 2 0 0 0 0 1 0 0.85
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 8 1 2 0 0 0 0 0 0.73
Winter Wheat (ref) 0 3 0 1 7 0 0 0 0 0 0 0.64
Spring Barlay 1 0 0 0 0 24 0 0 0 0 0 0.96
Rapeseed 0 0 0 0 0 0 23 0 0 0 0 1.0
Kernza 0 0 0 0 0 0 0 23 0 0 0 1.0
Kernza/Lucerne 0 0 0 0 0 0 1 2 27 0 0 0.9
Natural Grass 0 1 0 0 1 0 0 0 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.99 0.85 0.8 0.8 0.64 0.57 0.96 0.92 1.0 0.75 1.0
Overall acc: 0.88 Kappa: 0.85 OOB: 0.99
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 118 0 2 0 0 17 1 0 0 2 0 0.84
Grass Leg. Lay 0 20 0 2 2 0 0 0 0 2 0 0.77
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 8 1 2 0 0 0 0 0 0.73
Winter Wheat (ref) 0 0 0 1 9 0 0 0 0 1 0 0.82
Spring Barlay 1 0 0 0 0 24 0 0 0 0 0 0.96
Rapeseed 0 0 0 0 0 0 22 1 0 0 0 0.96
Kernza 0 0 0 0 0 0 0 23 0 0 0 1.0
Kernza/Lucerne 0 0 0 0 0 0 3 2 25 0 0 0.83
Natural Grass 0 1 0 0 1 0 0 0 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.99 0.95 0.86 0.73 0.69 0.56 0.85 0.88 1.0 0.64 1.0
Overall acc: 0.87 Kappa: 0.84 OOB: 0.99
82
Table B 35: Multispectral 11-class classification, pixel size: 5 cm. Features: R, Texture [ Mean], DEM
Table B 36: Multispectral 11-class classification, pixel size: 5 cm. Features: R, DEM
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 117 0 2 0 0 18 1 0 0 2 0 0.84
Grass Leg. Lay 0 17 0 4 2 0 0 0 0 3 0 0.65
Rye 0 1 11 0 0 0 0 0 0 0 0 0.92
Winter Wheat (orig) 0 0 0 7 3 1 0 0 0 0 0 0.64
Winter Wheat (ref) 0 0 0 2 8 0 0 0 0 1 0 0.73
Spring Barlay 1 0 0 0 0 24 0 0 0 0 0 0.96
Rapeseed 0 0 0 0 0 0 20 2 1 0 0 0.87
Kernza 0 0 0 0 0 0 1 21 0 1 0 0.91
Kernza/Lucerne 0 0 0 0 0 0 2 3 25 0 0 0.83
Natural Grass 0 2 0 1 1 0 0 0 0 7 0 0.64
Road 0 0 0 0 0 0 0 0 1 0 10 0.91
Producer acc.: 0.99 0.85 0.85 0.5 0.57 0.56 0.83 0.81 0.93 0.5 1.0
Overall acc: 0.83 Kappa: 0.73 OOB: 0.97
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 61 1 26 1 0 40 2 6 0 3 0 0.44
Grass Leg. Lay 0 17 0 3 3 0 0 0 0 3 0 0.65
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 7 3 0 0 0 0 1 0 0.64
Winter Wheat (ref) 0 4 0 0 6 0 0 0 0 1 0 0.55
Spring Barlay 1 3 1 1 0 19 0 0 0 0 0 0.76
Rapeseed 0 0 0 0 0 0 20 3 0 0 0 0.87
Kernza 0 0 0 0 0 0 1 17 1 4 0 0.74
Kernza/Lucerne 0 0 0 0 0 0 2 3 24 1 0 0.8
Natural Grass 0 2 0 1 1 0 0 0 0 7 0 0.64
Road 0 0 0 0 0 0 2 0 0 0 9 0.82
Producer acc.: 0.98 0.63 0.31 0.54 0.46 0.32 0.74 0.59 0.96 0.35 1.0
Overall acc: 0.62 Kappa: 0.56 OOB: 0.90
83
Table B 37: Multispectral 11-class classification, pixel size: 5 cm. Features: DEM
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 31 11 6 23 20 29 2 7 0 11 0 0.22
Grass Leg. Lay 0 8 5 3 3 5 0 0 0 2 0 0.31
Rye 0 6 4 0 1 1 0 0 0 0 0 0.33
Winter Wheat (orig) 1 1 0 6 1 1 0 0 0 1 0 0.55
Winter Wheat (ref) 1 1 1 1 3 4 0 0 0 0 0 0.27
Spring Barlay 5 2 1 7 3 6 0 0 0 1 0 0.24
Rapeseed 0 0 0 0 0 0 11 3 4 0 5 0.48
Kernza 0 0 0 0 0 0 0 17 1 5 0 0.74
Kernza/Lucerne 0 0 0 0 0 0 6 2 13 2 7 0.43
Natural Grass 0 1 3 2 0 1 0 2 1 1 0 0.1
Road 0 0 0 0 0 0 7 0 1 0 3 0.27
Producer acc.: 0.82 0.27 0.2 0.14 0.1 0.13 0.42 0.55 0.65 0.04 0.2
Overall acc: 0.32 Kappa: 0.24 OOB: 0.62
84
RGB 11-class classification, segmentation.
Table B 38: RGB 11-class classification, segmentation classification [Similarity tolerance: 50]
Table B 39: RGB 11-class classification, segmentation classification [Similarity tolerance: 100]
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 112 0 5 3 0 19 0 0 0 1 0 0.8
Grass Leg. Lay 0 19 0 0 0 0 0 3 3 1 0 0.73
Rye 0 0 12 0 0 0 0 0 0 0 0 1.0
Winter Wheat (orig) 0 0 0 10 1 0 0 0 0 0 0 0.91
Winter Wheat (ref) 0 0 0 1 10 0 0 0 0 0 0 0.91
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 0 0 0 0 0 0 23 0 0 0 0 1.0
Kernza 0 4 0 1 1 0 0 17 0 0 0 0.74
Kernza/Lucerne 0 1 0 0 0 0 0 0 29 0 0 0.97
Natural Grass 0 0 0 1 1 0 0 0 0 9 0 0.82
Road 0 0 1 0 0 0 0 0 0 0 10 0.91
Producer acc.: 1.0 0.79 0.67 0.63 0.77 0.57 1.0 0.85 0.90625 0.82 1.0
Overall acc: 0.85 Kappa: 0.82 OOB: 0.86
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 124 0 6 1 0 7 0 0 0 2 0 0.89
Grass Leg. Lay 0 24 0 0 0 0 0 2 0 0 0 0.92
Rye 0 1 11 0 0 0 0 0 0 0 0 0.92
Winter Wheat (orig) 0 0 0 8 1 2 0 0 0 0 0 0.73
Winter Wheat (ref) 0 0 0 0 11 0 0 0 0 0 0 1.0
Spring Barlay 2 0 0 0 0 23 0 0 0 0 0 0.92
Rapeseed 0 0 3 0 0 0 20 0 0 0 0 0.87
Kernza 0 4 0 0 0 0 0 19 0 0 0 0.83
Kernza/Lucerne 0 2 0 0 0 0 0 0 28 0 0 0.93
Natural Grass 0 0 0 1 0 0 0 0 0 10 0 0.91
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.98 0.77 0.55 0.8 0.92 0.72 1.0 0.9 1.0 0.83 1.0
Overall acc: 0.89 Kappa: 0.87 OOB: 0.84
85
Table B 40: RGB 11-class classification, segmentation classification [Similarity tolerance: 200]
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 127 0 4 2 1 5 1 0 0 0 0 0.91
Grass Leg. Lay 0 22 3 0 0 0 0 0 1 0 0 0.85
Rye 0 3 9 0 0 0 0 0 0 0 0 0.75
Winter Wheat (orig) 0 1 0 10 0 0 0 0 0 0 0 0.91
Winter Wheat (ref) 0 0 0 0 11 0 0 0 0 0 0 1.0
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 0 0 2 0 0 0 21 0 0 0 0 0.91
Kernza 0 0 0 0 0 0 0 22 1 0 0 0.96
Kernza/Lucerne 0 0 0 0 0 0 0 0 30 0 0 1.0
Natural Grass 0 0 0 0 0 0 0 0 0 11 0 1.0
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 1.0 0.85 0.5 0.83 0.92 0.83 0.95 1.0 0.94 1.0 1.0
Overall acc: 0.93 Kappa: 0.91 OOB: 0.68
86
Multispectral 11-class classification, segmentation.
Table B 41: Multispectral 11-class classification, segmentation classification [Similarity tolerance: 50]
Table B 42: Multispectral 11-class classification, segmentation classification [Similarity tolerance: 100]
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 121 0 3 0 1 13 0 0 0 2 0 0.86
Grass Leg. Lay 0 22 0 0 0 0 0 0 3 1 0 0.85
Rye 0 0 11 0 0 0 1 0 0 0 0 0.92
Winter Wheat (orig) 0 0 0 6 3 2 0 0 0 0 0 0.55
Winter Wheat (ref) 0 0 0 3 8 0 0 0 0 0 0 0.73
Spring Barlay 1 0 0 0 0 24 0 0 0 0 0 0.96
Rapeseed 0 0 0 0 0 0 22 1 0 0 0 0.96
Kernza 0 1 0 3 0 2 0 16 1 0 0 0.7
Kernza/Lucerne 0 1 0 1 0 0 0 6 22 0 0 0.73
Natural Grass 0 1 0 1 0 0 0 0 1 8 0 0.73
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 0.99 0.88 0.79 0.43 0.67 0.59 0.96 0.7 0.81 0.73 1.0
Overall acc: 0.84 Kappa: 0.8 OOB: 0.81
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 129 0 3 2 0 4 1 0 0 0 1 0.92
Grass Leg. Lay 0 24 0 1 0 0 0 0 1 0 0 0.92
Rye 0 1 10 0 0 0 0 0 1 0 0 0.83
Winter Wheat (orig) 0 0 0 9 1 0 0 0 1 0 0 0.82
Winter Wheat (ref) 0 0 0 1 10 0 0 0 0 0 0 0.91
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 0 0 0 0 0 0 23 0 0 0 0 1.0
Kernza 0 0 0 0 0 0 0 21 0 2 0 0.91
Kernza/Lucerne 0 3 0 0 0 0 0 3 24 0 0 0.8
Natural Grass 0 2 0 0 1 0 0 0 0 8 0 0.73
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 1.0 0.8 0.77 0.69 0.83 0.86 0.96 0.88 0.89 0.8 0.92
Overall acc: 0.91 Kappa: 0.89 OOB: 0.84
87
Table B 43: Multispectral 11-class classification, segmentation classification [Similarity tolerance: 200]
Bare Soil
Grass Leg. Lay
Rye Winter Wheat (orig)
Winther Wheat
(ref)
Spring Barlay
Rape-seed Kernza Kernza/ Lucerne
Natural Grass
Road User acc:
Bare Soil 133 0 1 0 2 0 2 0 0 2 0 0.95
Grass Leg. Lay 0 21 4 0 0 0 0 0 0 1 0 0.81
Rye 0 4 8 0 0 0 0 0 0 0 0 0.67
Winter Wheat (orig) 0 0 0 9 0 0 0 0 0 2 0 0.82
Winter Wheat (ref) 0 0 0 0 11 0 0 0 0 0 0 1.0
Spring Barlay 0 0 0 0 0 25 0 0 0 0 0 1.0
Rapeseed 0 0 0 0 0 0 23 0 0 0 0 1.0
Kernza 0 0 0 0 0 0 0 23 0 0 0 1.0
Kernza/Lucerne 0 0 0 0 0 0 0 0 30 0 0 1.0
Natural Grass 0 2 0 0 0 0 0 0 0 9 0 0.82
Road 0 0 0 0 0 0 0 0 0 0 11 1.0
Producer acc.: 1.0 0.78 0.62 1.0 0.85 1.0 0.92 1.0 1.0 0.64 1.0
Overall acc: 0.94 Kappa: 0.92 OOB: 0.69
88
Kernza – Lucerne – Dandelion classification. Pixel-by-pixel classification.
Table B 44: Kernza/Lucerne multispectral classification (pixel-by-pixel classification)
Kernza Lucerne Dandelion User acc:
Kernza 7 2 0 0.78
Lucerne 4 3 0 0.43
Dandelion 1 0 7 0.88
Producer acc.: 0.58 0.6 1.0 Overall acc: 0.71 Kappa: 0.55
OOB: 0.55
Table B 45: Kernza/Lucerne RGB classification (pixel-by-pixel classification)
Kernza Lucerne Dandelion User acc:
Kernza 7 2 0 0.78
Lucerne 4 3 0 0.43
Dandelion 0 0 8 1.0
Producer acc.: 0.64 0.6 1.0 Overall acc: 0.75 Kappa: 0.62
OOB: 0.92