Conference PaperPDF Available

Urban land use extraction from very high resolution remote sensing images by bayesian network

Mengmeng Li, Alfred Stein, Wietske Bijker
Faculty of Geo-Information Science and Earth Observation (ITC), University of Twente, P.O. Box 217,
7500AE Enschede, The Netherlands,,
This study aims to characterize the spatial arrangement of
land cover features and integrate the spatial arrangement
information with other commonly used land use indicators.
The characterization is conducted at object level,
corresponding to land cover objects. At the local urban level,
a VHR image is dominated by buildings. Therefore, the
characterization of spatial arrangement of land cover
elements is mainly conducted on building objects. Since
building objects have different functional properties in urban
areas, we classify them into a set of different types according
to their geometrical, morphological, and contextual
properties. The spatial arrangement is characterized by
considering the composition of different building types.
Regarding the integration of land use indicators and spatial
arrangement information, we construct a Bayesian network,
in which the spatial arrangement is served as a latent variable,
and the land use indicators calculated according to existing
studies are treated as the nodes of this Bayesian network. This
is followed by urban land use classification. We applied our
proposed method to a subset of a Pleiades image for an urban
area of Wuhan, China. We conclude that our proposed
method can provide an effective means for urban land use
Index Terms— Urban land use, spatial arrangement
characterization, building types, Bayesian network
Urban land use information plays an important role in many
urban-related applications. Remote sensing images have the
potential of extracting land use and monitoring land use
changes at local, regional, and global scales [1]. In particular
at the local level, the growing availability of very high
resolution (VHR) remote sensing images, e.g., QuickBird,
GeoEye, WorldView and Pleiades images, has caused an
increase in extracting urban land use at local scale. Remote
sensing images record the physical properties of the earth
surface, i.e. land cover, whereas land use refers to the
corresponding functional aspects, i.e. how the land cover is
used by human beings. Contrary to land cover extraction, the
low-level spectral, textural, and geometrical information of
VHR images fails to effectively characterize different types
of land use.
Traditional land use extraction from VHR images relies
on landscape metrics calculated at well-defined land use
units, such as street blocks [2, 3]. These commonly-used
landscape metrics, however, fail to effectively characterize
urban structures in complex urban areas, thus leading to poor
extraction results. Earlier studies have emphasized that the
use of spatial arrangements would improve the performance
of land use extraction [4, 5].
In the context of urban land use extraction from VHR
images, this study provides a statistical characterization to
model the spatial arrangement of land cover elements, and a
Bayesian framework to integrate the modeled spatial
arrangement information with the commonly used land use
features. It explores novel ways to characterize urban land
use from VHR imagery.
The remainder of this paper is organized as follows:
Section 2 describes the study area and used data, and Section
3 illustrates the proposed method for urban land use
extraction. Section 4 presents experimental results, followed
by discussion and conclusions in Section 5.
The proposed method was tested over an urban area located
in Wuhan, China. Wuhan is the capital of Hubei province,
China, and is one of the most populous cities in China. It has
a population of more than 10 million, which still increases,
resulting in rapid land use changes, particularly in newly built
urban zones. For this study area, we acquired a subset of
Pleiades-1B image on 11 July 2013, with four multispectral
bands of 2 m spatial resolution and a panchromatic of 0.5 m,
and having an area of 25 square kilometers (i.e. 10,000 ×
10,000 pixels) (see Figure 1). This 0.5 m panchromatic band
was resampled from the 0.7 m raw panchromatic band by the
image provider using spline resampling. In addition, a road
map of this study area is used to partition the study image into
homogenous street blocks.
3334978-1-5090-3332-4/16/$31.00 ©2016 IEEE IGARSS 2016
The proposed workflow for extracting urban land use from
VHR images consists of four main components: (1) land
cover classification, (2) spatial arrangement characterization,
(3) land use indicators calculation, and (4) Bayesian network
modeling and land use inference.
The land cover classification adopts our previous work [6, 7],
which contains four main classes: vegetation, shadow,
buildings, and others. Briefly, we first obtain the vegetation
from a VHR image using Otsu's threshold method based on
the Normalized Difference Vegetation Index (NDVI). To
extract the shadow, a Hue-Saturation-Intensity (HSI) color
space converted from a Near-Infrared-Red-Green color space
is used to derive an index ሺܵ െ ܫሻ ሺܵ ൅ ܫ
Τ, upon which the
Otsu's threshold is conducted for shadow extraction.
Buildings are obtained by means of directional relationships,
modeled by a fuzzy landscape, between buildings and
shadows. The remaining regions are assigned as others.
Before proceeding to the characterization of spatial
arrangement for land use extraction, we should define the
land use classes and building types. The land use
classification system applied in this study, is a modified
version of the Chinese National Standard Land Use
Classification System, adapted to the characteristics of the
study area and the image. They are:
xHighdensityresidential, is used for living, mainly for
low income residents, and has high building density,
little open space and vegetation coverage. The buildings
inside a high-density residential area usually have small
size and low building height. These buildings are
commonly found to be adjacent, and are easily delineated
xLowdensity residential, consists of a group of high-rise
building for high income residents. In addition, one or
more detached buildings served as ancillary facilities
might be found.
xCommercial, is used for business and services, and
contains wholesale and retail land, accommodation and
restaurant, business and finance, and other commercial
land. Within a commercial street block, multi-functional
halls, which have large size, complex shapes and
adjacent parking lots, are commonly found close to
xResidentialcommercialcomplex, mixes the residential
and commercial land use. Within such a street block,
several different types of buildings can be found.
xIndustrial, is used for industrial manufacture. The
spatial arrangement of buildings with respect to an
industrial region, is similar to residential area. But
industrial areas usually have less vegetation coverage
and lower building height.
xPublicmanagementservices, contains government
agencies, organizations land, science and education land,
entertainment land, public facilities land, and parks and
green space. The corresponding spatial arrangement of
buildings might be characterized by few detached halls,
and detached houses.
xTransportation, mainly contains roads and the
corresponding stations.
The building types are defined by observing the
characteristics of the study area and image. Accordingly, this
paper defines five different building types: densely clustered
buildings, pairwise similar buildings, detached buildings,
along-street halls and detached halls.
3.2.1. Building type’s classification
After we have extracted buildings from a VHR image, we
classify buildings into the five defined building types by
exploiting the geometrical, morphological and contextual
properties of buildings. In this paper, eight geometrical
features are used, namely, length/width, border index,
compactness, shape index, roundness, rectangular fit,
solidity, convexity [7]. Two morphological features, i.e. the
mean and variance of morphological profiles, are used to
quantify the structural information of an image region [6].
Particularly, for these two morphological features, the
corresponding morphological profiles are created by a set of
path opening and closing operations, which is contrast with
the traditional morphological opening and closing operations.
In addition, for each image object, the proximity to roads is
computed as the inverse distance between this image object
to the border of its nearest streets.
Figure 1: Pleiades image of the study area in Wuhan, China.
3.2.2. Spatial arrangement characterization based on
building types
Given a street-block with a set of building objectsܴ, ݅ൌ
ͳǡ ǥ ǡ ݊, we assume that these building objects are classified
into five defined building types ܴ௜௝, ൌͳǡǥǡͷ. Let ݊ be the
number of building objects for the  building type, and
σ݊ൌ݊. Let ܨ
ௌ஺ be the feature vector representing the
spatial arrangement of this street-block, the feature vector ܨ
can be defined as ܨ
ௌ஺ ൌሾݔ
, where an entry ݔ refers
to a variable associated with the  building type. In this
paper, the variable ݔ represents the proportion of building
pixels belonging to the  building type within the street
block. In short, the spatial arrangement of a street block is
characterized by a feature vector, which is derived from the
proportion and types of different building objects inside this
In this paper, a set of commonly used land use indicators are
used for characterizing land use classes [2].
A Bayesian Network (BN) is a probabilistic graphical model
that encodes joint probability distribution over a set of
random variables. A BN is a directed acyclic graph (DAG),
where each node represents a random variable (or a group of
random variables), and the links express probabilistic
relationships between these variables. The graph then
captures the way in which the joint distribution over all of the
random variables can be decomposed into a product of factors
each depending only on a subset of the variables. Thus, for a
BN with nodes, the joint distribution is given by
௞ୀଵ (1)
where ݔ is a random variable, ݌ܽ denotes the set of parents
of ݔ, and ܠൌሼݔ
In this paper, the structure of the adopted BN is
constructed based on the expert knowledge on land use
extraction, corresponding to the integration of commonly
used land use indicators and spatial arrangement of land
cover elements. Figure 2 gives the structure of the proposed
BN for land use extraction from VHR images. Based upon
our BN, we classify the unlabeled land use units into different
land use.
Four basic land cover, namely building, vegetation, shadow
and unclassified (i.e. others) were extracted, see Figure 3. In
this map, the shadow accounted for approximately 10% of the
total image area was automatically extracted. It helped to
extract building objects with a number of 10273 (see red in
Figure 3). From a visual inspection with the original VHR
image, the buildings were extracted with a comparable
accuracy, by using the directional relationships between
buildings and their shadow. At object level, a homogenous
land use unit (i.e. street block) is heterogeneous in terms of
land cover types, their coverage and their spatial
arrangement, particularly for buildings.
By exploiting the geometrical, morphological, and spatial
properties of extracted building objects, we classified these
building objects into five pre-defined building types (see
Figure 4). It is evident that, within a homogenous street block,
land use classes differed from various composition of
Figure 2: The proposed Bayesian network for land use extraction
from VHR images. The root node ࡸࢁrefers to a land use unit (i.e.
a street block). The group of nodes ǡǥǡࡲ
refers to the used
land use indicators. The node ࡿ࡭ refers to the spatial arrangement
of a land use unit, and is represented by a group of nodes ࡮ࢀ
corresponding to building types. The nodes ࢈࢒ࢊ
ࡸ࡯ , ࢜ࢋࢍ
ࡸ࡯ ǡࡾ
ࡸ࡯ ǡࡾ
denote the land covers with respect to building, vegetation, shadow
and others.
Figure 3: The classified land cover map of the study image.
building types. For example, a residential block was generally
dominated by a large proportion of pairwise similar
buildings, while a commercial block was often occupied with
a large coverage ratio of along-street halls and detached halls.
At this level, we can see that for a homogenous land use unit,
it is still heterogeneous in terms of building types. On the
other hand, the spatial arrangement of land cover features
characterized in this paper by the composition of different
building types, however, is distinguishable for different land
use classes.
At street-block level, we classify unlabeled street blocks
into pre-defined land use classes (see Figure 5). The land use
information in this level is useful for a variaty of urban-
related applications.
In this study we proposed a novel way for characterizing the
spatial arrangement of land cover elements for urban land use
extraction from VHR images. We characterize the spatial
arrangement by considering the composition of a set of
predefined building types. Moreover, we construct a
Bayesian network to integrate the commonly used land use
indicators and spatial arrangement information. In this study,
we exploited multiscale information obtained from VHR
imagery. At object level, individual buildings can be
extracted. They can be further categorized into a set of
building types according to their functions by making use
their geometrical, morphological, and spatial properties. At
this level, i.e. building types’ level, isolated building objects
are aggregated to characterize the spatial arrangement of a
land use unit. Last, we obtain a land use map at street-block
level. Furthermore, by using a tree-shaped Bayesian network,
the multiscale information can be well transmitted. Future
work can be conducted to investigate urban land use changes
from multi-temporal VHR images by exploiting multiscale
information, and by using Bayesian networks.
[1] E. Banzhaf and M. Netzband, "Monitoring Urban Land use
Changes with Remote Sensing Techniques," Applied Urban
Ecology, ed: John Wiley & Sons, pp. 18-32, 2011.
[2] M. Herold, X. H. Liu, and K. C. Clarke, "Spatial metrics and
image texture for mapping urban land use," Photogrammetric
Engineering and Remote Sensing, vol. 69, pp. 991-1001, 2003.
[3] J. H. Lowry and M. B. Lowry, "Comparing spatial metrics that
quantify urban form," Computers, Environment and Urban Systems,
vol. 44, pp. 59-67, 2014.
[4] L. Gueguen, "Classifying Compound Structures in Satellite
Images: A Compressed Representation for Fast Queries," IEEE
Transactions on Geoscience and Remote Sensing, vol. 53, pp. 1803-
1818, 2015.
[5] J. van der Kwast, T. Van de Voorde, F. Canters, I. Uljee, S. Van
Looy, and G. Engelen, "Inferring urban land use using the optimised
spatial reclassification kernel," Environmental Modelling &
Software, vol. 26, pp. 1279-1288, 2011.
[6] M. Li, A. Stein, W. Bijker, and Q. Zhan, "Region-based urban
road extraction from VHR satellite images using Binary Partition
Tree," International Journal of Applied Earth Observation and
Geoinformation, vol. 44, pp. 217-225, 2016.
[7] M. Li, W. Bijker, and A. Stein, "Use of Binary Partition Tree and
energy minimization for object-based classification of urban land
cover," ISPRS Journal of Photogrammetry and Remote Sensing,
vol. 102, pp. 48-61, 2015.
[8] H.-T. Lin, C.-J. Lin, and R. Weng, "A note on Platt’s
probabilistic outputs for support vector machines," Machine
Learning, vol. 68, pp. 267-276, 2007.
Figure 4: The classified land cover map of the study image.
Figure 5: The classified land use map of the study image.
... Land use is categorized commonly into urban and rural land use, the distinction being made on the basis of population density (M. Li, Stein, Bijker, & Zhan, 2016). Urban land use, in turn, is categorized into residential land use, commercial land use, transportation-land use, industrial land use, open space and recreational land use. ...
Full-text available
In this digital world, there is a large requirement of high resolution satellite image. Images at a low resolution may contain relevant information that has to be integrated with the high resolution image to obtain the required information. This is being fulfilled by image fusion. Image fusion is merging of different resolution images into a single image. The output image contains more information, as the information is integrated from both the images Image fusion was conducted with two different algorithms: regression kriging and the LULU operators. First, regression Kriging estimates the value of a dependent variable at unsampled location with the help of auxiliary variables. Here we used regression Kriging with the Hyperion image band as the response variables and the LISS III image bands are the explanatory variables. The fused image thus has the spectral variables from Hyperion image and the spatial variables from the LISS III image. Second, the LULU operator is an image processing methods that can be used as well in image fusion technique. Here we explored to fuse the Hyperion and LISS III image. The LULU operators work in three stages of the process, viz the decomposition stage, the fusion and the reconstruction stage. Quality aspects of the fused image for both techniques have been compared for spectral quality (correlation) and spatial quality (entropy). The study concludes that the quality of the fused image obtained with regression kriging is better than that obtained with the LULU operator.
... Dlamini (2016) proposed a BNs approach to identify high deforestation risk areas and its underlying causes, while Mayfield et al. (2017) used several statistical and machine learning techniques, including BNs, to predict deforestation. In turn, Li et al. (2016) used BNs to infer about urban land use. Additionally, BNs have increasingly been used as decision support tools for environmental management. ...
In the last decade, Brazil has successfully managed to reduce deforestation in the Amazon forest. However, continued increases in annual deforestation rates call for environmental modeling to support short-term decision-making. This paper presents the functioning of a stepwise spatio-temporal Bayesian Network approach for spatially explicit analysis of deforestation risk based on observation data. The study area comprises a deforestation expansion frontier located in the southwest of Pará state, Brazil. The proposed approach has been successful in estimating deforestation risk over the years. Among the selected variables to compose the Bayesian Network model, distance from hot spots and distance from degraded areas present the highest contribution, while protected areas variable present a significant mitigation effect on the phenomenon. Accuracy assessment indices corroborate the agreement between deforestation events and predictions.
Full-text available
Identifying urban functional zones is of great significance for understanding urban structure and urban planning. The rapid growth and open accessibility of multi-source big data, including remote sensing imagery and social sensing data, lead to a new way for dynamic identification of urban functional zones. Here, we propose an SOE (scene-object-economy) based learning framework which integrates the scene features from remote sensing imagery, the object features from building footprints and the economy features from POIs (points of interest). From these three perspectives, the rich information hidden in the urban zone is excavated for function identification. CNNs (convolutional neural networks) are used to extract high-level scene information from remote sensing images with different resolutions, which is distinguish from spectral and texture features of the urban zone. The object features comprising building indicators are constructed by measuring the area, perimeter, floor number, year of the building. Moreover, we extract socio-economic characteristics from POIs, which reflect the types of human activities in the urban zone. Lastly, a RF (random forest) classifier is utilized to fuse the SOE features for identifying functional zones. The SOE-based framework is applied to the case of Shenzhen city, and the accuracy reaches 90.8\% in the case of using remote sensing images with a resolution of 0.3-meter. The experimental results show that the prediction performance of SOE-based framework is significantly better than other traditional methods, and the quantitative contribution of SOE factors is also revealed in determining the functionality of urban zone.
Full-text available
With the increased spatial resolution of current sensor constellations, more details are captured about our changing planet, enabling the recognition of a greater range of land use/land cover classes. While pixeland object-based classification approaches are widely used for extracting information from imagery, recent studies have shown the importance of spatial contexts for discriminating more specific and challenging classes. This paper proposes a new compact representation for the fast query/classification of compound structures from very high resolution optical remote sensing imagery. This bag-of-features representation relies on the multiscale segmentation of the input image and the quantization of image structures pooled into visual word distributions for the characterization of compound structures. A compressed form of the visual word distributions is described, allowing adaptive and fast queries/classification of image patterns. The proposed representation and the query methodology are evaluated for the classification of the UC Merced 21-class data set, for the detection of informal settlements and for the discrimination of challenging agricultural classes. The results show that the proposed representation competes with state-of-the-art techniques. In addition, the complexity analysis demonstrates that the representation requires about 5% of the image storage space while allowing us to perform queries at a speed down to 1 s/ 1000 km2/CPU for 2-m multispectral data.
Full-text available
The arrival of new-generation, high-spatial-resolution satellite imagery (e.g., Ikonos) has opened up new opportunities for detailed mapping and analysis of urban land use. Drawing on the traditional approach used in aerial photointerpretation, this study investigates an “object-oriented” method to classify a large urban area into detailed land-use categories. Spatial metrics and texture measures are used to describe the spatial characteristics of land-cover objects within each land-use region as derived from interpreted aerial photographs. In assessing how land-use categories vary in their spatial configuration, spatial metrics were found to provide the most important information for differentiating urban land uses. A detailed land-use map with nine categories was derived for the Santa Barbara South Coast Region area. Results from our work suggest that the region-based method exploiting spatial metrics and texture measurements is a potential new avenue to extract detailed urban land-use information from high-resolution satellite imagery.
Full-text available
Land use changes and their consequences for urban ecologyUrban remote sensing (URS) and geographical information systems (GIS) for research in urban ecologyMeasuring physical characteristics of urban areas with remote sensing technologyGlobal initiatives to measure urban expansion and land use changeRegional urban monitoring activitiesSynthesis and outlookReferences
Two main challenges are faced when classifying urban land cover from very high resolution satellite images: obtaining an optimal image segmentation and distinguishing buildings from other man-made objects. For optimal segmentation, this work proposes a hierarchical representation of an image by means of a Binary Partition Tree (BPT) and an unsupervised evaluation of image segmentations by energy minimization. For building extraction, we apply fuzzy sets to create a fuzzy landscape of shadows which in turn involves a two-step procedure. The first step is a preliminarily image classification at a fine segmentation level to generate vegetation and shadow information. The second step models the directional relationship between building and shadow objects to extract building information at the optimal segmentation level. We conducted the experiments on two datasets of Pléiades images from Wuhan City, China. To demonstrate its performance, the proposed classification is compared at the optimal segmentation level with Maximum Likelihood Classification and Support Vector Machine classification. The results show that the proposed classification produced the highest overall accuracies and kappa coefficients, and the smallest over-classification and under-classification geometric errors. We conclude first that integrating BPT with energy minimization offers an effective means for image segmentation. Second, we conclude that the directional relationship between building and shadow objects represented by a fuzzy landscape is important for building extraction.
In the 1990s, promising results in land-use classification were obtained by kernel-based contextual classification algorithms. Soon, however, it was recognised that kernel-based reclassifiers have important shortcomings and research instead focused on object-based image analysis. This study proposes a solu-tion to two of the most important drawbacks of kernel-based reclassifiers: (1) the use of kernels tends to smooth boundaries between discrete land-use/land-cover parcels, and (2) it is difficult to determine a priori the optimum kernel size of the classifier. The Spatial Reclassification Kernel (SPARK) algorithm has been adapted in order to automatically optimise the kernel size depending on the spatial variation found in the neighbourhood of a pixel to be classified, resulting in the Optimised SPARK (OSPARK) algorithm. The performance of SPARK and OSPARK for land-use classification has been evaluated for the Dublin urban area (Ireland), using a Landsat TM image. The MOLAND land-use map of 1990 was used as a reference. Results show that the use of optimal kernel sizes instead of fixed kernel sizes reduces the omission and commission errors for most land-use classes.
Platt’s probabilistic outputs for Support Vector Machines (Platt, J. in Smola, A., et al. (eds.) Advances in large margin classifiers. Cambridge, 2000) has been popular for applications that require posterior class probabilities. In this note, we propose an improved algorithm that theoretically converges and avoids numerical difficulties. A simple and ready-to-use pseudo code is included.