CREATING WALLONIA'S NEW VERY HIGH RESOLUTION LAND COVER MAPS: COMBINING GRASS GIS OBIA AND OTB PIXEL-BASED RESULTS

Abstract

The Walloon region of Belgium has launched a research project that aims at elaborating a methodology for automated, high-quality land cover mapping, based primarily on its yearly 0.25m orthophoto coverage. Whereas in urban areas an object-based (OBIA) approach has been the privileged path in the last years as it allows taking into account shape information relevant for the characterization of man-made constructions, such an approach has its limits in the rural and more natural areas due to increased difficulties for segmentation and less sharp boundaries, thus calling for a pixel-based approach. The project thus consists in developing a combination of methods, and to integrate their results through an ensemble fusion approach. As many of the more natural land cover classes have temporal profiles which cannot be detected in a one-date orthoimage, Sentinel 1 and 2 data are also included in order to take advantage of their higher spectral and temporal resolution. All methods are trained using existing regional databases. In a second step, we combine the different LC classification results by fusioning them into one high-accuracy (over 90% OA) product, using a series of different approaches ranging from rule-based to machine learning to the Dempster-Shafer method. The entire toolchain is based on free and open source software, mainly GRASS GIS and Orfeo ToolBox. Results indicate the importance of the quality of the individual classifications for the fusion results and justify the choice of combining OBIA and pixel-based approaches in order to avoid the pitfalls of each.

Full text

CREATING WALLONIA'S NEW VERY HIGH RESOLUTION LAND COVER MAPS:
COMBINING GRASS GIS OBIA AND OTB PIXEL-BASED RESULTS
M. Lennert* 1, T. Grippa 1, J. Radoux 2, C. Bassine 2, B. Beaumont 3, P. Defourny 2, E. Wolff 1
1 ANAGEO-DGES, Université Libre de Bruxelles, Belgium, moritz.lennert@ulb.ac.be
2 Earth and Life Institute, Université catholique de Louvain, Louvain-la-Neuve, Belgium
3 Remote Sensing and Geodata Unit, Institut Scientifique de Service Public, Liège, Belgium
Commission IV, WG IV/4
KEY WORDS: Remote Sensing, Land Cover, Ensemble fusion, Wallonia, GRASS GIS, Orfeo ToolBox
ABSTRACT:
The Walloon region of Belgium has launched a research project that aims at elaborating a methodology for automated, high-quality
land cover mapping, based primarily on its yearly 0.25m orthophoto coverage. Whereas in urban areas an object-based (OBIA)
approach has been the privileged path in the last years as it allows taking into account shape information relevant for the
characterization of man-made constructions, such an approach has its limits in the rural and more natural areas due to increased
difficulties for segmentation and less sharp boundaries, thus calling for a pixel-based approach. The project thus consists in
developing a combination of methods, and to integrate their results through an ensemble fusion approach. As many of the more
natural land cover classes have temporal profiles which cannot be detected in a one-date orthoimage, Sentinel 1 and 2 data are also
included in order to take advantage of their higher spectral and temporal resolution. All methods are trained using existing regional
databases. In a second step, we combine the different LC classification results by fusioning them into one high-accuracy (over 90%
OA) product, using a series of different approaches ranging from rule-based to machine learning to the Dempster-Shafer method.
The entire toolchain is based on free and open source software, mainly GRASS GIS and Orfeo ToolBox. Results indicate the
importance of the quality of the individual classifications for the fusion results and justify the choice of combining OBIA and pixel-
based approaches in order to avoid the pitfalls of each.
1. INTRODUCTION
Land cover (LC) maps, showing the characteristics of surface
elements, e.g. vegetation, artificial constructions, water, etc, are
essential components for regional decision-making, for uses as
diverse as spatial planning, environmental monitoring and
modelling, flood risk assessment, etc.
Even though the Walloon region in Belgium has compiled a
rich catalogue of vector geodata, the actual LC map currently
available dates back over a decade and an update was thus
needed. The regional administration decided to launch a
research project to develop a robust, automatized, scalable and
reproducible method for creating these data, mainly based on
the available VIS-NIR orthoimagery at 0.25 m resolution, as
well as height information derived through photogrammetry
from the raw version of that same imagery. The ultimate aim of
the project is not only to provide recent (2018), INSPIRE-
compliant maps, but also to elaborate a method that would
make it easier for the region to reproduce such data at higher
temporal frequency than in the past, ideally based on FOSS4G
software in order to avoid vendor lock-in and licence costs. As
a response to the need for high accuracy in very diversified
landscapes, from densely urbanized areas to large forests, a
combination of approaches was chosen.
This paper details the preliminary outputs of the work, which is
still in progress. We begin with a very short overview of the
current state-of-the-art in LC mapping, to then go on to
describing the data, methods and intermediate results, before
discussing the lessons already learned.
2. STATE OF THE ART
2.1 Different approaches to LC mapping
Because of its importance for many different fields, land cover
mapping has attracted a lot of attention from the research
community. Starting with simple pixel-based approaches based
on low resolution satellite images available at the time, the
advent of very high resolution (VHR) imagery has led to the
development of object-based approaches, notably for man-made
landscapes such as urban areas (Blaschke 2010; Chen et al.
2018). However, with more and more open data available at
global scale, particularly the Landsat and Sentinel satellites,
pixel-based methods continue to attract attention and
development (Grekousis, Mountrakis, and Kavouras 2015).
Such approaches as also particularly used in the environmental
sciences community, for example in the work concerning local
climate zones (Stewart and Oke 2012; Bechtel et al. 2015), the
ESA climate change initiative (Bontemps et al. 2012; Hollmann
et al. 2013) or crop mapping (Inglada et al. 2015).
Another, new actor is deep learning which has shown very
promising results for land cover mapping (Zhu et al. 2017;
Zhang, Zhang, and Du 2016). Deep learning is, however, seen
by many as a “black box”, making it difficult to use in public
decision-making which requires accountability (Dosilovic,
Brcic, and Hlupic 2018; Samek, Wiegand, and Müller 2017) .
This black box character, leading to more difficulties for the
administration to assess and accept the proposed processes than
classic, well-known methods, was one of the main reasons not
to use deep learning in this project.
* Corresponding author, moritz.lennert@ulb.ac.be
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
151

3. DEFINITION OF USER NEEDS AND OBJECTIVES
An important part of the project, although not treated in detail
here is the work with users in order to identify and prioritize
their needs (Beaumont et al. 2019). This includes the definition
of a legend for the final land cover product. As Figure 1 shows,
a legend with 6 main classes has been designed, with two of the
main classes subdivided into two subclasses each. One of the
classes, arable land, was difficult to define because it combines
two land cover types in time (bare soil and herbaceous
vegetation within the same year). This can obviously not be
detected with single-date imagery, calling for the use of images
of higher temporal resolution, but lower spatial resolution such
as Sentinel 1 and 2.
Beyond the legend, other issues need to be balanced when
working on such land cover maps: the necessary level of
accuracy, the temporal resolution and the minimum size of
objects in the map (minimum mapping unit or MMU). In order
to identify acceptable trade-offs, users were asked to allocate a
limited budget of points to these three objectives. Figure 2
shows the result of this consultation, including the contradictory
needs. As a general conclusion, it was decided to aim for an
MMU of 15m2 (possibly to be adapted according to classes), a
frequency of update between 3 and 5 years, and a minimum
overall accuracy of the product over 85%. As this first round of
mapping should provide a very solid baseline for future
updates, the team has proposed to aim for a much higher overall
accuracy (around 95%).
Figure 2: Compromises between objectives as preferred by
users
4. DATA
Different sources of data are available on the study area. Each
has its own spatial reference and specific feature for the land
cover classification. The main input is a mosaic of orthophoto
images with 25 cm pixels and 4 spectral bands (NIR-Red-Green
and Blue). The study area was covered thanks to several flights
in spring and summer 2018. The data provider also used
photogrammetry to build a digital surface model from the
original photographs. A digital height model was then derived
by subtracting a 1m resolution LIDAR-based digital elevation
model from 2013. Beside the very high resolution airborne
datasets, images from Sentinel-1 (10 m C-Band SAR data) and
Sentinel-2 (10 and 20 m multispectral data) are used to analyze
the temporal dynamic of the land cover. Finally, a 2 m open
data land cover layer from the Lifewatch project (Radoux et al.
2019) is used for training, as well as a vector database of roads,
railways and rivers provided by the Walloon Region. This
vector database was enhanced into a planimetric reference : a
linear roads dataset and three planimetric datasets for railways,
buildings and rivers. From the linear ancillary data, the road
network was completed and made continuous using toolboxes
from the open source platform QGIS and GRASS GIS. In
particular, the linear referencing toolbox from QGIS was used
to solve completeness issues in a road dataset.
5. METHODS
The methods used were chosen based on an ensemble of
criteria:
•Reflecting the state-of-the-art
•Ease of application for a regional administration
•Scalability for very large datasets (the input data is
several TB)
•Potential for automation of the entire procedure
•Existing experience in the research teams
5.1 Per pixel approach using Orfeo ToolBox
Orfeo ToolBox (OTB) is a C++ library for state of the art
remote sensing (Grizonnet et al. 2017). All of the algorithms are
accessible from Monteverdi, QGIS, Python, the command line
or C++. OTB is also the core of the SEN2AGRI toolbox (http://
www.esa-sen2agri.org/), an image processing platform used in
this study for the classification of crop types based on Sentinel-
2 images. Due to the large size of the orthophotos, command
lines are managed by the SLURM job scheduling system. The
parallel processing is driven by zones of similar flight
conditions according to the metadata of the orthophotos.
The first step of the processing consists in smoothing the
orthophotos based on the meanshift algorithm (Comaniciu and
Meer 2002) in order to reduce potential salt and pepper effects.
A reference dataset is then built based on the 2 m land cover
map, the MNH and the shadows predicted from the MNS with a
custom OTB application. This reference dataset is then eroded
using a multiclass mathematical morphology operator (Radoux
et al. 2014). The orthophotos are then classified with an a priori
probability defined by the height class. In addition, a random
forest classifier was applied on a stack of two dates (leaves on
and leaves off) of Sentinel-2 images. This classification focuses
on the discrimination of forest types to better discriminate
between broadleaved deciduous and needle-leaved evergreen
forests, as well as larch stands which are deciduous needle-
leaved trees.
5.2 OBIA approach using GRASS GIS
GRASS GIS is a full-fledged geographical information system
(Neteler et al. 2012), grown over 35 years of continuous
development. The OBIA approach used here is based on a
complete toolchain developed during the last years which
provides everything from segmentation (including unsupervised
parameter optimization) to machine learning classifiers (Grippa
et al. 2017). All tools have been specifically designed for very
large data sets, allowing parallel computing. A previous project
already allowed a first test of this toolchain on a small part of
the Walloon territory (Beaumont et al. 2017).
Figure 1: Landcover legend defined after user consultation
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
152

Recent studies have shown that using a constant segmentation
parameter across space does not lead to ideal segments, as
spatial structures differ significantly between different parts of
the territory (Grippa et al. 2017; Georganos et al. 2018; Drăguţ
et al. 2019). The OBIA approach in this study takes this into
account by optimizing the segmentation within fairly small
tiles, delineated using a module for the creation of semantically
useful cutlines (Lennert and GRASS GIS Development Team
2018), inspired by the work of Soares et al (2016). In order to
speed up the region-growing segmentation, superpixels are first
created in a rapid run of the SLIC algorithm (Kanavath, Metz,
and GRASS GIS Development Team 2018; Achanta et al.
2012).
After segmentation, diverse statistics (spectral, shape, texture,
height, x and y coordinates) are gathered for each segment,
including information about the segments neighbours. Using
the existing vector databases, segments falling into polygons of
specific classes are identified automatically as training
segments. Outliers (e.g. segments identified as artificial
construction which have a tree growing over them in the image)
are eliminated through simple tests mainly based on NDVI and
height.
The data set is divided into strata according to the date at which
the photos were taken. A subsample of the very large set of
training segments is then used to train a different random forest
model on each of the strata and the resulting models applied on
each strata’s tiles, resulting in a choice of class for each
segment, as well as the probability of each class. The OBIA
approach focuses specifically on the quality of the classification
of more man-made landscapes, for which it is particularly well
suited.
5.3 Data fusion
We test three different approaches for the fusion of the raw LC
maps into one final automated product. A set of 1500 points
labelled by visual interpretation and stratified in 3 strata to
cover different types of mismatches between the input
classifications, is used as a reference for training and validation.
At the time of writing, this fusion is still in progress, and only
preliminary results are presented here. Beyond ongoing work
on the different input classifications, we are currently
elaborating an enhanced point dataset in order to improve the
model training.
5.3.1 Dempster-Shafer
The different classification results and the clean vector database
are fused at the pixel level based on the Dempster-Shafer fusion
(Ran et al. 2012). This method combines masses of belief based
on the confusion matrix of each classification. The mass of
belief indicates the belief of each input classification present for
each label value based on a reference dataset. The confusion
matrix is computed independently for three strata selected based
on the likelihood of mislabelling. For each strata in the
reference point data set, a specific fusion is computed and the
three results are merged by an ultimate Dempster-Shafer fusion
based on the entire set of points.
5.3.2 Object-based machine learning
The fusion in this approach is based on the objects created
through the OBIA classification. For each of these objects a
series of features are extracted from the different classification
attempts such as entropy-based analysis of class probabilities
from both the per-pixel and the OBIA results, class proportions
within each object from the pixel-based results, modal classes
of the Sentinel-based results as well as modal classes from
binary version of the ancillary vector datasets.
Using the reference points a random forest classifier is then
trained to predict classes for the objects.
5.3.3 Rule-based
The rule-based approach is built interactively based on expert
assessment via visual interpretation. It currently comprises
about 15 conditional, pixel-based rules. For each pixel
(25x25cm), LC is attributed taking into account the agreement
between the two ortho-image classifications, the presence-
absence within ancillary reference datasets and contextual rules.
The Sentinel-based classifications (time-series) and digital
height model serve as support for the LC allocation of complex
remaining pixels.
Knowing that such as rule-based approach requires significant
investment of human resources to arrive at satisfying results
over a very large territory, we mainly consider it as a reference
application which allows us to benchmark the other approaches.
6. RESULTS
6.1 Results of different classifications
In this section we present a small selection of results of the
VHR pixel-based and OBIA approaches in order to illustrate
some difficulties in each. Figure 3 shows some typical issues of
pixel-based approaches: on the left, one can clearly see the salt-
and-pepper effect linked to the high variability of individual
spectral signatures, while the extract on the right illustrates the
high sensitivity to the quality (and spatial precision) of the input
data, in this case the height information which does not have the
same precision as the spectral data, and is slightly spatially
shifted.
While objects resulting from an OBIA approach often allow
mitigating the issue of high variability of individual pixel-
values by using means or other aggregated statistics, figure 4
illustrates quite well a fundamental issue with object-based
approaches in very heterogeneous environments such as forests.
Objects in these regions are often too small, increasing the risk
of misclassification.
Figure 5 reinforces this idea by showing that objects which are
not as easily segmented as buildings have a higher uncertainty
in their classification results. At the same time, the figure
demonstrates the usefulness of the OBIA technique for
delineating man-made structures fairly precisely.
Figure 3: Illustration of issues in the pixel-based classification
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
153

Figure 4: Illustration of over-segmentation in the OBIA
approach
Figure 5: Illustration of uncertainty linked to object type in
OBIA (darker color = higher uncertainty)
6.2 Results of fusion
As mentioned above, the work on fusion is still ongoing, but
some first results already are available and allow for an initial
analysis of the advantages or disadvantages of the different
approaches, as illustrated in figures 6-9 (see figure 1 for the
legend).
A major difference can be seen between the object-based,
machine learning approach and the two others is the fact that
just as in the initial classification, the objects provide a certain
smoothing effect, while ensuring sharper edges for man-made
constructions. The Dempster-Shafer approach has difficulties
dealing with the different resolutions and thus leads to a less
spatially precise result, although thematically it sometimes does
a better job identifying the right classes.
One class which is difficult to deal with in the entire processing
chain is arable land. As the VHR classifications are limited to
one point in time, this information can only come from the
Sentinel-derived classifications which have much lower
resolution.
6.3 Discussion
The above results show that fusion clearly provides qualitative
improvements over the individual classifications. However, a
major determinant of the quality of results is the accuracy of the
inputs into the fusion. This is why the team adopted an iterative
approach, going back and forth between fusion and original
classification in order to identify the best parameters at each
step.
Figure 7: Object-based fusion using machine learning (for
legend see figure 1)
Figure 8: Rule-based fusion ((for legend see figure 1)
Figure 6: Original orthophoto
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
154

Giving priority to the ancillary reference datasets provides
smoother maps for these features, but strongly depends on the
timeliness of their content (e.g. lots of omission/commission
errors if not up to date, such as for the building layer). This is
somewhat of a chicken-and-egg problem, as a procedure for
elaborating a high quality LC map depends on up-to-date input,
while potentially being one of the prime sources of such up-to-
date information.
As expected, the rule-based approach is time-consuming,
notably for the parametrization of the rules, as well as
subjective, so it is difficult to implement in an automated setting
across a large territory.
7. CONCLUSION AND PERSPECTIVES
While the precise details of classification and fusion are still
work in progress, we can already clearly see that using multiple
data, classifying them through different methods and combining
them in a final fusion step leads to better results than the
individual methods on their own. The results seem on par with
the precision level expected by the potential users, and the
entire processing chain can potentially be in-housed within the
administration as it relies entirely on free and open source
software and simple scripts that ensure complete automation.
The final LC map will be made available as open data through
the Walloon region’s geoportal.
An important aspect of the project for the FOSS4G community
is the fact that it has allowed continuous improvement of
existing tools. Some GRASS GIS modules were enhanced
during the project (e.g. i.segment.uspo, i.segment.stats,
i.cutlines, v.class.mlR), with all enhancements integrated back
into the software’s source code. Some new modules were also
developed (e.g. r.texture.tiled) and are now available for the
entire community. This clearly shows the advantage for public
administrations to fund projects based on FOSS4G as all
developments will continue to be available, including to other
parts of the administration, thus providing a potential effect of
pooling of public resources.
Within the project itself work will continue in order to finalize
the automated LC map. A period of manual corrections is
foreseen in order to create one extremely high quality base map
which can provide the foundation of a high frequency update
cycle of LC maps in Wallonia. Furthermore, the project will
integrate LC information as attribute data into existing regional
vector databases, thus supporting easier uptake of the data by
different user communities, and a combination of the LC map
with existing alpha-numeric, geocoded, databases will provide
input into automated land use (LU) mapping.
Future research will have to confront the methods developed
here to new developments, notably in deep learning. As
mentioned in the literature review, some elements of deep
learning make it less attractive to public administrations.
However, the very high quality LC map coming out of this
project has the potential to provide useful input for the training
of deep learning networks, possibly maintained within the
administration for future needs. Although this will require an
effort to make the entire process less black box, it has the
potential of providing the administration with a potent tool for
frequent updating.
ACKNOWLEDGEMENTS
The authors would like to acknowledge the Walloon
Government for the funding of the project WALOUS.
Computational resources for the OBIA and machine learning
fusion approaches are being provided by the Shared ICT
Services Centre, Université Libre de Bruxelles.
Computational resources for the pixel-based processing have
been provided by the Consortium des Équipements de Calcul
Intensif (CÉCI), funded by the Fonds de la Recherche
Scientifique de Belgique (F.R.S.-FNRS) under Grant No.
2.5020.11 and by the Walloon Region.
The authors would also like to thank the GRASS GIS
Development team for their responsiveness to any requests for
changes.
REFERENCES
Achanta, R., A. Shaji, K. Smith, A. Lucchi, P. Fua, and Sabine
Süsstrunk. 2012. “SLIC Superpixels Compared to State-of-the-
Art Superpixel Methods.” IEEE Transactions on Pattern
Analysis and Machine Intelligence 34 (11): 2274–82.
https://doi.org/10.1109/TPAMI.2012.120.
Beaumont, Benjamin, Tais Grippa, Moritz Lennert, Sabine
Vanhuysse, Nathalie Stephenne, and Eléonore Wolff. 2017.
“Toward an Operational Framework for Fine-Scale Urban
Land-Cover Mapping in Wallonia Using Submeter Remote
Sensing and Ancillary Vector Data.” Journal of Applied
Remote Sensing 11 (03): 1.
https://doi.org/10.1117/1.JRS.11.036011.
Beaumont, Benjamin, Laura Van de Vyvere, Nathalie
Stephenne, Coraline Wyard, and Eric Hallot. 2019. “Users’
Consultation Process in Building a Land Cover and Land Use
Database for the Official Walloon Georeferential.” presented at
the Joint Urban Remote Sensing Event (JURSE), Vannes.
Bechtel, Benjamin, Paul J. Alexander, Jürgen Böhner, Jason
Ching, Olaf Conrad, Johannes Feddema, Gerald Mills, Linda
See, and Iain Stewart. 2015. “Mapping Local Climate Zones for
a Worldwide Database of the Form and Function of Cities.”
ISPRS International Journal of Geo-Information 4 (1): 199–
219. https://doi.org/10.3390/ijgi4010199.
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.
Figure 9: Dempster-Shafer fusion (for legend see figure 1)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
155

Bontemps, S., M. Herold, L. Kooistra, A. van Groenestijn, A.
Hartley, O. Arino, I. Moreau, and P. Defourny. 2012.
“Revisiting Land Cover Observation to Address the Needs of
the Climate Modeling Community.” Biogeosciences 9 (6):
2145–57. https://doi.org/10.5194/bg-9-2145-2012.
Chen, Gang, Qihao Weng, Geoffrey J. Hay, and Yinan He.
2018. “Geographic Object-Based Image Analysis (GEOBIA):
Emerging Trends and Future Opportunities.” GIScience &
Remote Sensing 55 (2): 159–82.
https://doi.org/10.1080/15481603.2018.1426092.
Comaniciu, D., and P. Meer. 2002. “Mean Shift: A Robust
Approach toward Feature Space Analysis.” IEEE Transactions
on Pattern Analysis and Machine Intelligence 24 (5): 603–19.
https://doi.org/10.1109/34.1000236.
Dosilovic, Filip Karlo, Mario Brcic, and Nikica Hlupic. 2018.
“Explainable Artificial Intelligence: A Survey.” In 2018 41st
International Convention on Information and Communication
Technology, Electronics and Microelectronics (MIPRO), 0210–
15. Opatija: IEEE.
https://doi.org/10.23919/MIPRO.2018.8400040.
Drăguţ, Lucian, Mariana Belgiu, George Popescu, and Peter
Bandura. 2019. “Sensitivity of Multiresolution Segmentation to
Spatial Extent.” International Journal of Applied Earth
Observation and Geoinformation 81 (September): 146–53.
https://doi.org/10.1016/j.jag.2019.05.002.
Georganos, Stefanos, Tais Grippa, Moritz Lennert, Sabine
Vanhuysse, Brian Johnson, and Eléonore Wolff. 2018. “Scale
Matters: Spatially Partitioned Unsupervised Segmentation
Parameter Optimization for Large and Heterogeneous Satellite
Images.” Remote Sensing 10 (9): 1440.
https://doi.org/10.3390/rs10091440.
Grekousis, George, Giorgos Mountrakis, and Marinos
Kavouras. 2015. “An Overview of 21 Global and 43 Regional
Land-Cover Mapping Products.” International Journal of
Remote Sensing 36 (21): 5309–35.
https://doi.org/10.1080/01431161.2015.1093195.
Grippa, Tais, Stefanos Georganos, Moritz Lennert, Sabine
Vanhuysse, and Eléonore Wolff. 2017. “A Local Segmentation
Parameter Optimization Approach for Mapping Heterogeneous
Urban Environments Using VHR Imagery.” In . Vol. 10431.
Warsaw: International Society for Optics and Photonics. https://
doi.org/10.1117/12.2278422.
Grippa, Taïs, Moritz Lennert, Benjamin Beaumont, Sabine
Vanhuysse, Nathalie Stephenne, and Eléonore Wolff. 2017.
“An Open-Source Semi-Automated Processing Chain for Urban
Object-Based Classification.” Remote Sensing 9 (4): 358.
https://doi.org/10.3390/rs9040358.
Grizonnet, Manuel, Julien Michel, Victor Poughon, Jordi
Inglada, Mickaël Savinaud, and Rémi Cresson. 2017. “Orfeo
ToolBox: Open Source Processing of Remote Sensing Images.”
Open Geospatial Data, Software and Standards 2 (1): 15.
https://doi.org/10.1186/s40965-017-0031-6.
Hollmann, R., C. J. Merchant, R. Saunders, C. Downy, M.
Buchwitz, A. Cazenave, E. Chuvieco, et al. 2013. “The ESA
Climate Change Initiative: Satellite Data Records for Essential
Climate Variables.” Bulletin of the American Meteorological
Society 94 (10): 1541–52. https://doi.org/10.1175/BAMS-D-11-
00254.1.
Inglada, Jordi, Marcela Arias, Benjamin Tardy, Olivier Hagolle,
Silvia Valero, David Morin, Gérard Dedieu, et al. 2015.
“Assessment of an Operational System for Crop Type Map
Production Using High Temporal and Spatial Resolution
Satellite Optical Imagery.” Remote Sensing 7 (9): 12356–79.
https://doi.org/10.3390/rs70912356.
Kanavath, R., M. Metz, and GRASS GIS Development Team.
2018. I.Superpixels.Slic. Geographic Resources Analysis
Support System (GRASS) Software. USA: Open Source
Geospatial Foundation.
https://grass.osgeo.org/grass72/manuals/addons/i.superpixels.sli
c.html.
Lennert, Moritz, and GRASS GIS Development Team. 2018.
I.Cutlines. Geographic Resources Analysis Support System
(GRASS) Software. USA: Open Source Geospatial Foundation.
https://grass.osgeo.org/grass72/manuals/addons/i.cutlines.html.
Neteler, Markus, M. Hamish Bowman, Martin Landa, and
Markus Metz. 2012. “GRASS GIS: A Multi-Purpose Open
Source GIS.” Environmental Modelling & Software 31 (May):
124–30. https://doi.org/10.1016/j.envsoft.2011.11.014.
Radoux, Julien, Axel Bourdouxhe, William Coos, Marc
Dufrêne, and Pierre Defourny. 2019. “Improving Ecotope
Segmentation by Combining Topographic and Spectral Data.”
Remote Sensing 11 (3): 354.
https://doi.org/10.3390/rs11030354.
Radoux, Julien, Céline Lamarche, Eric Van Bogaert, Sophie
Bontemps, Carsten Brockmann, and Pierre Defourny. 2014.
“Automated Training Sample Extraction for Global Land Cover
Mapping.” Remote Sensing 6 (5): 3965–87.
https://doi.org/10.3390/rs6053965.
Ran, Y.H., X. Li, L. Lu, and Z.Y. Li. 2012. “Large-Scale Land
Cover Mapping with the Integration of Multi-Source
Information Based on the Dempster–Shafer Theory.”
International Journal of Geographical Information Science 26
(1): 169–91. https://doi.org/10.1080/13658816.2011.577745.
Samek, Wojciech, Thomas Wiegand, and Klaus-Robert Müller.
2017. “Explainable Artificial Intelligence: Understanding,
Visualizing and Interpreting Deep Learning Models.”
ArXiv:1708.08296 [Cs, Stat], August.
http://arxiv.org/abs/1708.08296.
Soares, Anderson Reis, Thales Sehn Körting, and Leila Maria
Garcia Fonseca. 2016. “IMPROVEMENTS OF THE DIVIDE
AND SEGMENT METHOD FOR PARALLEL IMAGE
SEGMENTATION.” Revista Brasileira de Cartografia 68 (6).
http://www.lsie.unb.br/rbc/index.php?
journal=rbc&page=article&op=view&path%5B%5D=1602.
Stewart, I. D., and T. R. Oke. 2012. “Local Climate Zones for
Urban Temperature Studies.” Bulletin of the American
Meteorological Society 93 (12): 1879–1900.
https://doi.org/10.1175/BAMS-D-11-00019.1.
Zhang, L., L. Zhang, and B. Du. 2016. “Deep Learning for
Remote Sensing Data: A Technical Tutorial on the State of the
Art.” IEEE Geoscience and Remote Sensing Magazine 4 (2):
22–40. https://doi.org/10.1109/MGRS.2016.2540798.
Zhu, Xiao Xiang, Devis Tuia, Lichao Mou, Gui-Song Xia,
Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. 2017.
“Deep Learning in Remote Sensing: A Comprehensive Review
and List of Resources.” IEEE Geoscience and Remote Sensing
Magazine 5 (4): 8–36.
https://doi.org/10.1109/MGRS.2017.2762307.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
156

APPENDIX
All scripts used for the OBIA part of the project are made
available on https://github.com/mlennert/WALOUS.
Revised June 2019
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W14, 2019
FOSS4G 2019 – Academic Track, 26–30 August 2019, Bucharest, Romania
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W14-151-2019 | © Authors 2019. CC BY 4.0 License.
157