Inferring urban land use using the optimised spatial reclassification kernel
Highlights
► Major shortcomings of kernel-based reclassification algorithms were solved. ► An algorithm that optimises the kernel size for each pixel has been developed. ► Using optimal instead of fixed kernel sizes improves the reclassification results.
Introduction
Planners and policy makers need tools to anticipate and assess the impact of their decisions on the spatial systems they are to manage. For this purpose many land-use change models are currently available (Agarwal et al., 2002). These models are typically calibrated using a historic calibration procedure, which requires time series of land-use maps (Engelen et al., 2007). Often, however, time series of land-use maps are lacking. Even when time series exist, inconsistencies in mapping methodologies, legends and scales result in measured land-use changes that are caused by mismatches in the mapping procedures rather than real land-use changes (Van der Kwast et al., 2009, Pontius and Petrova, 2010).
On the other hand, large archives of medium resolution earth observation data are available (e.g. LPDAAC (2009)), with images dating back to the early 1970s. During the 1980s and 1990s spectral classifiers applied to multispectral and hyperspectral images gained popularity. Spectral classifiers have been successfully used in deriving thematic land-cover maps, using the often simple and direct relationship between land-cover type and reflectance measured by the sensor (Barnsley et al., 2001). Land use, however, is more difficult to derive from remote sensing images using spectral classifiers (Eyton, 1993, Barnsley and Barr, 1996), since it refers to human activities taking place at the earth’s surface, which often have a complex and indirect relation with the reflectance signal received by the sensor (Barnsley and Barr, 1997).
In the beginning of this millennium, the coarse spectral resolution of new fine resolution images, the overlap between spectral signatures from different classes and the demand for land-use maps sparked interest in classifiers that use the spatial domain. The concept of these so-called contextual classifiers is based on the idea that information captured in neighbouring cells or information about patterns surrounding the pixel of interest may provide useful supplementary information in the classification process (De Jong and Van der Meer, 2004). Previous studies have demonstrated a strong relationship between the spatial structure of urban areas and its functional characteristics (Barnsley and Barr, 1997), even at medium resolution (Barnsley and Barr, 1996).
Early experiments with classifications using the spatial domain mainly focused on moving window, i.e. kernel-based techniques (Wharton, 1982, Gong and Howarth, 1990, Eyton, 1993, Barnsley and Barr, 1996). Most kernel-based reclassifiers have the advantage of conceptual simplicity and are therefore easy to implement. Although the results were promising, important shortcomings of kernel-based techniques were recognised (Barnsley et al., 2001):
- 1.
Boundaries between discrete land-use/land-cover parcels are smoothed;
- 2.
It is difficult to determine a priori the optimum kernel size;
- 3.
A regular, rectangular window is an artificial shape used to classify irregularly-shaped land-cover/land-use parcels (Dilworth et al., 1994, Herold et al., 2003).
These limitations of kernel-based reclassifiers have recently moved the focus of research in contextual classification from kernel-based reclassifiers to object-based image analysis (Blaschke, 2010). Object-based image analysis, known as OBIA or GEOBIA, aims at extracting and classifying objects from remote sensing images. Most of these methods use image segmentation algorithms to divide an image into homogeneous, continuous and contiguous objects. However, because of the heterogeneity of the urban landscape and its spatial variability, the segmentation of urban scenes is a challenging task (Pesaresi and Bianchin, 2001) and requires manual tuning (Nikfal and Samadzadegan, 2008). Furthermore, the detection of real world objects is not possible using medium resolution images for which the pixel size is much larger than the size of the objects of interest.
This study tries to overcome the first two problems related to the use of kernel-based reclassifiers, using a recommendation suggested by Barnsley and Barr (1996) and Barnsley et al. (2001). They suggest that the use of optimal kernel sizes might help to take into account the different spatial scales of variation in land-cover characteristics of different land-use categories. Kernels that are too large in respect of the land-use objects will increase the smoothing of edges, while too small kernels possibly do not include all spatial variation needed to classify the pixel. By varying the kernel size so that it matches the typical scale of variation of each land-use type, the spatial signature of different land uses can be better captured. By avoiding the use of too large kernels for land-use types that can be characterised well with small kernel sizes, the impact of edge smoothing may be substantially reduced.
Many studies have been done on optimising the kernel size for classifiers using image texture (e.g. Franklin et al. (1996)), mostly based on semivariogram analysis (Woodcock et al., 1988). Far less research has been done on optimising the kernel size for contextual reclassification algorithms, since semivariogram theory cannot be applied to discrete thematic data. Contextual reclassification algorithms refine previously classified images using spatial variation in the neighbourhood of each pixel (Sluiter et al., 2004, Van de Voorde et al., 2007). Examples of reclassification algorithms include the local frequency distribution method (Wharton, 1982), the Spatial and Spectral Classification (SSC) method (De Jong et al., 2001) and the Spatial Reclassification Kernel (SPARK) method (Barnsley and Barr, 1996).
In this paper we present a novel approach to contextual reclassification of medium resolution images yielding an important improvement of the classification of urban areas. For this purpose the Spatial Reclassification Kernel (SPARK) algorithm (Barnsley and Barr, 1996) is adapted for the automatic optimisation of kernel sizes. Results for using a variable kernel size for each pixel or a fixed kernel size for all pixels are compared. The approach is evaluated for medium resolution images, because time series of archived remote sensing data are still dominated by medium resolution images that can be potentially used as additional data for the historic calibration of land-use change models.
Section snippets
The concept of SPARK
The Spatial Reclassification Kernel (SPARK, (Barnsley and Barr, 1996)) is a contextual reclassification method. The conceptual idea behind SPARK is that land-use types of interest can be characterised by the spatial arrangement and size of objects within the landscape. For urban areas, for example, this can be the spatial arrangement of objects like streets, buildings, bare soil areas, shade, trees and grass. SPARK examines the local, spatial patterns of land cover in a square kernel or moving
The Optimised SPARK algorithm
The Optimised SPARK (OSPARK) algorithm is based on SPARK, which was extended to automatically adapt the kernel size to the spatial variation detected around the pixel to be classified, i.e. the centre pixel of the kernel, in order to overcome the first two disadvantages of kernel-based reclassifiers discussed in the introduction of this paper. Conceptually, the approach is similar to the recently published A-SPARK approach (Alimohammadi and Shirkavand, 2010). The flowchart of the OSPARK
Study area
The study area for this research is Dublin, the political, economical and cultural capital of Ireland and home to over 40% of the country’s population. Dublin experienced rapid urban expansion in the 1980’s and 1990’s. While the Greater Dublin area as a whole experienced only a moderate population growth of 3.6% between 1986 and 1996, population in the urban periphery increased more rapidly with as much as 9.6% in South Dublin and 21.1% in Fingal, to the north. This has resulted in a hollowing
Results
Fig. 4(c) and (e) show respectively the SPARK (9 × 9 kernel) and OSPARK classification results. Visual inspection of the results shows us more spatial variability in the remote sensing derived land-use maps than in the reference MOLAND land-use map (Fig. 4(a)), especially in the agricultural areas.
The SPARK and OSPARK classification results were evaluated by comparison with the MOLAND land-use map. Goodness-of-fit statistics (Pontius and Millones, 2011) were derived from the analysis of the
Discussion
The results show that OSPARK improves the classification output for most classes by using adaptive kernel sizes. A disadvantage of incorporating large kernel sizes in the OSPARK algorithm, however, is the reduction of the area that can be classified. This is caused by the choice to assign missing values to the pixel under consideration if one or more missing values exist in the kernel defined around the pixel. In the SPARK application, where a fixed kernel of 9 × 9 pixels was used, this
Conclusions
A novel approach has been presented as a solution to the most important limitation of kernel-based contextual classification algorithms, which is the choice of an a priori kernel size that is fixed for the entire image and not necessarily optimal for all land-use classes. The proposed OSPARK algorithm replaces the fixed kernel size of the SPARK algorithm with an automatically adapting kernel, as a function of the spatial variability around each pixel to be classified. The algorithm has been
Acknowledgements
The research presented in this paper is funded by the Belgian Science Policy Office in the frame of the STEREO II programme - project SR/00/105. The land-use data used for Dublin were made available by the MOLAND project of the EU - Joint Research Centre in Ispra, Italy. In this study the implementation of the original SPARK algorithm made by the PCRaster research and development group at Utrecht University, the Netherlands was used in the implementation of the OSPARK algorithm.
References (43)
- et al.
Distinguishing urban land-use categories in fine spatial resolution land-cover data using a graph-based, structural pattern recognition system
Computers, Environment and Urban Systems
(1997) Object based image analysis for remote sensing
ISPRS Journal of Photogrammetry and Remote Sensing
(2010)- et al.
An integrated spatial and spectral approach to the classification of Mediterranean land cover types: the SSC method
International Journal of Applied Geosciences
(2001) Urban land use classification and modeling using cover-type frequencies
Applied Geography
(1993)- et al.
Automated derivation of geographic window sizes for use in remote sensing digital image texture analysis
Computers & Geosciences
(1996) - et al.
Assessing a predictive model of land change using uncertain data
Environmental Modelling & Software
(2010) The landsat ETM+ spectral mixing space
Remote Sensing of Environment
(2004)- et al.
Design and analysis for thematic map accuracy assessment: fundamental principles
Remote Sensing of Environment
(1998) A contextual classification method for recognising land use patterns in high resolution remotely-sensed data
Pattern Recogition
(1982)- et al.
The use of variograms in remote sensing: I. Scene models and simulated images
Remote Sensing of Environment
(1988)
A Review and Assessment of Land-use Change Models: Dynamics of Space, Time, and Human Choice
Adaptive spatial reclassification kernels for urban mapping from remotely sensed data: the A-SPARK approach
International Journal of Remote Sensing
ISODATA, a Novel Method of Data Analysis and Pattern Classification
The greater Dublin region: planning for its transformation and development
Inferring urban land use from satellite sensor images using kernel-based analysis and classification
Photogrammetric Engineering & Remote Sensing
Inferring urban land use by spatial and structural pattern recognition
Monitoring trends in urban growth and surveying city quarters in Ouagadougou, Bourkina Faso using SPOT-XS
Geocarto International
Measuring landcape structure using geographic and geometric windows
Photogrammetric Engineering and Remote Sensing
The MURBANDY and MOLAND models for Dublin
The Moland modelling Framework for urban and regional land-use dynamics
Cited by (18)
Kernel-based granulometry of textural pattern measures on satellite imageries for fine-grain sparse woodlands mapping
2021, Computers in Earth and Environmental Sciences: Artificial Intelligence and Advanced Technologies in Hazards and Risk ManagementA Bayesian characterization of urban land use configurations from VHR remote sensing images
2020, International Journal of Applied Earth Observation and GeoinformationCitation Excerpt :It is thus difficult to distinguish urban land use types at fine categories, e.g., residential, commercial, and industrial land use, from VHR images. To improve the separability between different urban land use types, previous research has recognized the importance of spatial arrangement for urban land use extraction using VHR images (van der Kwast et al., 2011; Li et al., 2016). For example, van der Kwast et al. (2011) characterized the arrangement of land cover pixels within a spatial kernel for pixel-based land use classification.
Urban land use extraction from Very High Resolution remote sensing imagery using a Bayesian network
2016, ISPRS Journal of Photogrammetry and Remote SensingCitation Excerpt :We used a well-established urban land use map to segment a VHR image for obtaining homogenous land use units. This differs from other studies on land use, e.g. van der Kwast et al. (2011) using regular grids, Comber et al. (2012) using sub-graphs partitioned from a planar graph, and Hu and Wang (2013) using parcels. An existing urban land use map however may not be available for some areas.
An ontological system for interoperable spatial generalisation in biodiversity monitoring
2015, Computers and GeosciencesCitation Excerpt :To automatically aggregate this class, a spatial reclassification is needed. This work proposes a spatial reclassification approach, which is able to extend existing generalisation methods (Thoonen et al., 2010; van der Kwast et al., 2011) by using semantic relations and inference to generate comparability of the outcomes of different regions with regard to its content. It represents a further enhancement and detailed evaluation of the methodology developed by Nieland et al. (2014).
A land-use systems approach to represent land-use dynamics at continental and global scales
2012, Environmental Modelling and SoftwareCitation Excerpt :Most time series of historical land-use data are based on backward simulations as in Goldewijk (2001) and can therefore not be regarded as independent data suitable for calibration and validation. The preferred data source for evaluating spatial patterns of land-use change would be time series of satellite products, and method have been developed to infer land-use data from existing remotely sensed images (van der Kwast et al., 2011). However, these data do only cover a short time period (Gibbs et al., 2010; Hansen et al., 2008), and differences between time steps may rather show differences in accuracy of reference data due to measuring and processing methodologies than differences in land use (Pontius and Petrova, 2010).
A free and open source programming library for landscape metrics calculations
2012, Environmental Modelling and SoftwareCitation Excerpt :Many applications of landscape metrics come from landscape ecology, but other geographical applications are being developed by taking advantage of the properties of these metrics. For example, the Geographic Object-Based Image Analysis (GEOBIA) paradigm continues to prove its efficacy in remote sensing image analysis (Jacquin et al., 2008; van der Kwast et al., 2011) and very related to this discipline the field of Knowledge Discovery in Databases (KDD) offers many possibilities for solving many complex environmental problems (Pereira dos Santos Silva et al., 2008; Gibert and Sànchez-Marrè, 2011). The most important conclusion of this literature is that landscape metrics enable diverse information to be incorporated into a geographic information system (GIS), help to select areas with high ecological values, and allow researchers to use appropriate management measures (Musacchio and Coulson, 2001).