SPUSPO: Spatially Partitioned Unsupervised Segmentation Parameter Optimization for Efficiently Segmenting Large Heterogeneous Areas

Abstract

Very-High-Resolution (VHR) Remote Sensing (RS) data are crucial for deriving essential geospatial information on cities, e.g. for urban planning, population estimation and socioeconomic assessments with particular merit in sub-Saharan Africa (SSA) due to the scarcity or absence of reference data. One of the cornerstones of information that can be produced from RS is classified Land Use and Land Cover (LULC) maps. For VHR imagery, Object Based Image Analysis (OBIA) is the most efficient methodology to produce such outputs. A crucial intermediate step in OBIA is the selection of a suitable segmentation scale. However, for large, heterogeneous areas (e.g., at city level), little effort has been made to optimize OBIA algorithms. Supervised methods to optimize segmentation parameters are subjective and time consuming while Unsupervised Segmentation Parameter Optimization (USPO) techniques, assume spatial stationarity for the whole image. This is problematic for geographically large heterogenous areas and does not capture intra-urban variations due to building size, materials and fractions of LULC intrinsically varying in space. In this study, we employ a novel framework named Spatially Partitioned Unsupervised Segmentation Parameter Optimization (SPUSPO) that optimizes segmentation parameters locally for two SSA cities, Dakar and Ouagadougou. The framework employs the open access GRASS GIS software that is suitable for large scale computing. Our results suggest that SPUSPO is an efficient way to optimize segmentation parameters for large and heterogeneous urban areas.

Full text

SPUSPO: SPATIALLY PARTITIONED UNSUPERVISED SEGMENTATION PARAMETER
OPTIMIZATION FOR EFFICIENTLY SEGMENTING LARGE HETEROGENEOUS AREAS
S. Georganos1, T. Grippa1, M. Lennert1, S. Vanhuysse1, E. Wolff1
1 Université libre de Bruxelles (ULB), Department of Geosciences, Environment and Society (DGES-IGEAT)
ABSTRACT
Very-High-Resolution (VHR) Remote Sensing (RS) data are
crucial for deriving essential geospatial information on
cities, e.g. for urban planning, population estimation and
socioeconomic assessments with particular merit in sub-
Saharan Africa (SSA) due to the scarcity or absence of
reference data. One of the cornerstones of information that
can be produced from RS is classified Land Use and Land
Cover (LULC) maps. For VHR imagery, Object Based
Image Analysis (OBIA) is the most efficient methodology to
produce such outputs. A crucial intermediate step in OBIA
is the selection of a suitable segmentation scale. However,
for large, heterogeneous areas (e.g., at city level), little effort
has been made to optimize OBIA algorithms. Supervised
methods to optimize segmentation parameters are subjective
and time consuming while Unsupervised Segmentation
Parameter Optimization (USPO) techniques, assume spatial
stationarity for the whole image. This is problematic for
geographically large heterogenous areas and does not
capture intra-urban variations due to building size, materials
and fractions of LULC intrinsically varying in space. In this
study, we employ a novel framework named Spatially
Partitioned Unsupervised Segmentation Parameter
Optimization (SPUSPO) that optimizes segmentation
parameters locally for two SSA cities, Dakar and
Ouagadougou. The framework employs the open access
GRASS GIS software that is suitable for large scale
computing. Our results suggest that SPUSPO is an efficient
way to optimize segmentation parameters for large and
heterogeneous urban areas.
Index Terms— unsupervised segmentation parameter
optimization, GRASS GIS, VHR imagery, land cover
1. INTRODUCTION
In sub-Saharan African cities, in the absence of adequate
data for urban planning, one of the most important pieces of
information we can derive through Very-High-Resolution
(VHR) Remote Sensing (RS), is detailed and accurate Land
Use and Land Cover (LULC) maps. These products are
often used as input for epidemiological, population and
socio-economic models, among others. Nonetheless, to
achieve adequate classification results in these challenging
landscapes, Object Based Image Analysis (OBIA) [1] is
frequently employed over pixel-based approaches due to
superior performance [2]. Nonetheless, OBIA can be a very
tedious and computationally demanding technique,
particularly for large-scale applications (e.g., at city level).
Moreover, the classification of relevant images requires
several preparatory tasks, ranging from the selection of an
appropriate segmentation (object-creating) algorithm, to the
effective calibration of the parameters of the algorithm itself
[3]. Such parameters control the shape and size of the
created segments. Since the segmentation quality is crucial
to the results of the classification itself, the optimization of
these parameters is a critical methodological facet.
Region-growing segmentation algorithms are very
frequent in the OBIA literature, mainly due to their
effectiveness and ease of implementation. The selection of
the parameters of the segmentation algorithm is most
commonly achieved through a time-consuming, user
dependent, trial and error process in which the quality of the
segmentations is assessed visually or through alternative
methodologies which rank different segmentations based on
reference data. However, supervised calibration approaches
are untenable for large heterogeneous regions if maximizing
classification accuracy and segmentation quality in an
automated approach is one of the aims. Consequently,
research efforts have been directed towards the development
of objectively defined Unsupervised Segmentation
Parameter Optimization (USPO) techniques, that evaluate
individual segmentations based on geostatistical metrics,
and do not require reference data and user interference, thus
allowing for automated processes [4]. The optimization of
USPO metrics has been attempted mainly by the use of
global methods, either at single or at multiple scales. This
means that the segmentation over the whole region of
interest is optimized by using one set of parameters. This
global approach has recently been shown to be inferior to
local approaches in urban and agricultural environments
either through evaluation metrics such as classification
accuracy and detailed visual examination [5],[6]. Thus, in
order to treat the growing amount of VHR remote sensing
data available in a way that is both time-efficient and
accurate, an approach based on the hypothesis that the
segmentation optimal parameters would intrinsically and
significantly vary across heterogeneous regions due to local
variations in data structure, is more appropriate. However,
there is no established automated framework for applying
local USPO for very large VHR urban datasets. In general,
little effort has been made to identify and quantify the
degree of spatial non-stationarity between the algorithm
parameters, especially for large heterogeneous areas such as
cities, where neighborhood structure and consequently,
optimal segmentation parameters might vary due to different
building materials, size, fractions and types of vegetation

and road network. As such, our aim is to present a
methodological framework to optimize segmentation results
for large heterogenous areas based on the following
premises: i) allowing segmentation parameters to vary,
which can provide significant increases in classification
quality and ii) implement the methods through an open
source semi-automated processing chain, suitable for large-
scale computing utilizing mainly GRASS GIS, in order to
make the previous task manageable.
2. DATA AND STUDY AREA
The classification scheme is implemented for Dakar and
Ouagadougou, the capitals of Senegal and Burkina Faso,
respectively (Figure 1). They are both major Sahelian cities,
which have been facing an extensive urban growth since the
last decades. Pansharpened tristereo Pleiades imagery
(VNIR, 0.5m, 400 km2) acquired in 2015 was used for
Dakar while Worldview-3 stereo imagery acquired in the
same year was used for Ouagadougou (VNIR, 0.5m, 630
km2). Normalized Digital Surface Models (nDSM) were
produced by photogrammetry from both datasets to aid in
built-up fabric identification.
Figure 1. True color composites of Ouagadougou (top) and
Dakar (bottom).
3. METHODS
In this paper, we present a methodological framework
named Spatially Partitioned Unsupervised Segmentation
Parameter Optimization (SPUSPO) in which
optimization of segmentation parameters is performed
locally. We followed a different approach for the spatial
partitioning of each city, a fully unsupervised process for
Dakar and a supervised one for Ouagadougou (Figure 2).
For Dakar, we split the image into 1700 500m by 500m grid
cells. For Ouagadougou, an expert-based visual delineation
of local morphological zones (LMZ) was performed by
applying the following ruleset:
• LMZs should be homogeneous, both in terms of
building size and density, and should be visibly
different from their neighboring LMZs.
• LMZs boundaries should follow, as far as possible,
man-made or natural linear elements, e. g., roads, paths,
rivers, streams, railways.
• Built-up LMZs should be larger than 1.5 hectares (ha).
Both of these spatially partitioning methods were compared
to a global approach in which segmentation parameters were
optimized in a small subset of the study areas and the
suggested value was applied to the whole study area.
Figure 2. Different methods of spatial partitioning a)
automatic delineation of grid cells in Dakar and b) manual,
expert delineation of LMZ in Ouagadougou.
Afterwards, we made extensive use of the semi-
automated processing chain developed by Grippa et al. [7],
that combines GRASS GIS [8], Python and R programming
languages along with PostGIS support in a Jupyter
Notebook implementation. The chain allows for a complete
analysis, from input of the initial image datasets to the
production of final LC maps. An excerpt example of the
Jupyter chain is given in Figure 3, while the general
workflow is described in Figure 4. In detail, the
segmentation was optimized in each of the spatial subsets
(grid cells for Dakar and LMZs for Ouagadougou) using the
i.segment.uspo module of GRASS [9] with a region growing
algorithm. The module uses geospatial metrics to perform
the optimization, such as Moran’s I and the variance to find
the best compromise between intra- and inter-segment
heterogeneity. Moreover, it allows for parallelization and is
thus computationally efficient. Once layers of features
calculated on the segmentation layers coming from different
techniques were produced, a random forest classifier was
implemented to evaluate their efficacy [10]. To train and
evaluate the models we visually identified roughly 3000
objects for both cities. As shown in Table 1, we used a
detailed classification scheme to draw more robust
conclusions about the merits of using local optimization on

large regions.
Figure 3. Excerpt of the Jupyter notebook consisting of a
sequence of code and a descriptive text that documents the
different processing steps that can be executed directly from
the notebook.
Figure 4. General workflow of the processing chain.
Table 1. Classification scheme for Ouagadougou and Dakar
Ouagadougou
Dakar
Buildings
Building type 1
Swimming Pools
Building type 2
Asphalt
Asphalt
Brown/Red Bare Soil
Bare soil, dusty
concrete
White/Grey Bare Soil
Trees
Trees
Mixed Bare
Soil/Vegetation
Low vegetation
(e.g. grass)
Dry Vegetation
Bushes
Other Vegetation
Inland Waters
Inland waters
Shadow
Shadow
4. RESULTS
The results indicated that a local approach can be of merit.
Figure 5 demonstrates the variability of optimized threshold
parameter for each grid cell across Dakar (Figure 5). It is
evident that the local method suggests spatial non-
stationarity of the USPO parameter values. The patterns of
the deviation from the single parameter of the global
approach follow variations in the landscape such as the
types of vegetation and built-up fabric. Supporting our
hypothesis, the results from evaluating the classification
against a reference set implied that undertaking a local
optimization path constructs segments of a higher quality.
The overall accuracy (OA) for Dakar reached 88.20% and
89.50% for the global and local approaches, respectively. In
Ouagadougou, similar results were observed with an OA of
84.77% and 85.45%. Being an urban application, it is also
relevant to examine how well the built-up classes are
predicted. By using the F-score as an indicator to evaluate
per class performance, the results indicate an increase of 2%
and 1% for Dakar and Ouagadougou, respectively. Although
the improvements are not very strong, a detailed visual
examination implied that the local approaches predict built-
up and asphalt classes, in a more consistent fashion. This
can be important given that these products are frequently
used for population and epidemiological applications where
built-up information is the most important feature.
Figure 5. Deviation of the local threshold parameter from
the global one for each grid cell in Dakar. The threshold
parameter controls the shape and size of the segments and
thus, the quality of the segmentation.
There are several reasons that could explain these
improvements. A single segmentation parameter can over-
segment and under-segment objects of the same class in
different regions (Figure 6). For example, industrial areas
consist of large buildings while unplanned areas contain
very small housing units. By allowing the threshold
parameter to vary, SPUSPO can fit these changes in the data
allowing for better classification results.
5. CONCLUSIONS
In this paper, we assessed a new method for optimizing
segmentation parameters in large and highly heterogeneous

Figure 6. Segmentation and classification results for Dakar.
Top left: Classified segments resulting from the global
approach. Top right: Classified segments resulting from the
local approach. Bottom left: True color composite. Bottom
right: segments boundaries coming from the global (red) and
local (green) approaches.
Table 2. Classification accuracy for Ouagadougou and
Dakar with a global approach and SPUSPO, respectively.
Local
Global
Ouagadougou
Dakar
Ouagadougou
Dakar
OA%
85.45
89.50
84.77
88.20
Kappa%
0.840
0.878
0.833
0.863
Built-up%
0.930
0.960
0.920
0.940
urban areas with the prospect of deriving high quality
LU/LC maps. We call this method Spatially Partitioned
Unsupervised Segmentation Parameter Optimization
(SPUSPO). We employed two different ways to partition the
images i.e., automatic partition by grid cells and expert
based delineation using morphological criteria. The
methodological framework was realized through an open
source processing chain that mainly exploited the artillery of
GRASS GIS for large scale computing. Our results
suggested that spatial partition methods have merit as
increased classification accuracies were observed in
comparison to a global approach, along with better
prediction of artificial surfaces. Future work includes more
testing of the method in different imageries and refining
techniques of automatic delineating useful spatial subsets to
optimize the data. Finally, a comparison between the two
approaches demonstrated in this study– automated vs expert
based, will be attempted to assess if a completely automated
approach can be as efficient with a focus in built-up areas
identification as they are the most challenging elements to
classify in SSA cities.
AKNOWLEDGEMENTS
The research presented in this paper is funded by BELSPO
(Belgian Science Policy Office) in the frame of the
STEREO III program – project REACT (SR/00/337).
6. REFERENCES
[1] T. Blaschke, “Object based image analysis for remote sensing,”
ISPRS J. Photogramm. Remote Sens., vol. 65, no. 1, pp. 2–16,
2010.
[2] T. Blaschke et al., “Geographic Object-Based Image Analysis -
Towards a new paradigm,” ISPRS J. Photogramm. Remote Sens.,
vol. 87, pp. 180–191, 2014.
[3] A. Rasanen, A. Rusanen, M. Kuitunen, and A. Lensu, “What
makes segmentation good? A case study in boreal forest habitat
mapping,” Int. J. Remote Sens., vol. 34, no. 23, pp. 8603–8627,
2013.
[4] B. Johnson, M. Bragais, I. Endo, D. Magcale-Macandog, and P.
Macandog, “Image Segmentation Parameter Optimization
Considering Within- and Between-Segment Heterogeneity at
Multiple Scale Levels: Test Case for Mapping Residential Areas
Using Landsat Imagery,” ISPRS Int. J. Geo-Information, vol. 4, no.
4, pp. 2292–2305, 2015.
[5] T. Grippa, S. Georganos, M. Lennert, S. Vanhuysse, and E.
Wolff, “A local segmentation parameter optimization approach for
mapping heterogeneous urban environments using VHR imagery,”
in Remote Sensing Technologies and Applications in Urban
Environments II, 2017, vol. 10431, p. 104310G.
[6] F. Cánovas-García and F. Alonso-Sarría, “A local approach to
optimize the scale parameter in multiresolution segmentation for
multispectral imagery,” Geocarto Int., vol. 30, no. 8, pp. 937–961,
2015.
[7] T. Grippa, M. Lennert, B. Beaumont, S. Vanhuysse, N.
Stephenne, and E. Wolff, “An Open-Source Semi-Automated
Processing Chain for Urban Object-Based Classification,” Remote
Sens., vol. 9, no. 4, p. 358, 2017.
[8] M. Neteler, M. H. Bowman, M. Landa, and M. Metz, “GRASS
GIS: A multi-purpose open source GIS,” Environ. Model. Softw.,
vol. 31, pp. 124–130, 2012.
[9] M. . Lennert and G. D. Team, “Addon i.segment.uspo,” Geogr.
Resour. Anal. Support Syst. SoftwareVersion 7.3, 2016.
[10] S. Georganos, T. Grippa, S. G. Vanhuysse, M. Lennert, and E.
Wolff, “Optimizing classification performance in an object-based
very-high-resolution land use-land cover urban application,” in
Remote Sensing Technologies and Applications in Urban
Environments II, vol. 10431.