Mapping and modelling urban growth and its impact on the hydrology of urban watersheds with satellite imagery

EarthzineEarth Observation, Feature, Featured Article, In-Depth, Original, Water

Figure 3: Average daily rainfall (shaded and units of millimeters) from June-August (200-2006) composited for days with weak atmospheric forcing (lef). The 5 mm contour is shown in blue and the yellow boxes represent the mean upwind (left box), city (center box), and downwind (right box) regions. Lightning flash anomalies (May to September, 1995‰ÛÒ2003) on days dominated by weak atmospheric forcing (right) in Atlanta, Georgia. Image sources: [4] and [5], respectively.

Tim Van de Voorde1, Boud Verbeiren2, Yves Cornet3, Marc Binard3, Johannes van der Kwast4, Guy Engelen5, Frank Canters1 and Okke Batelaan2


1 Vrije Universiteit Brussel, Department of Geography, Brussels, Belgium;

2 Vrije Universiteit Brussel, Department of Hydrology and Hydraulic Engineering, Brussels, Belgium;

3 UniversitÌ© de Li̬ge, Geomatics Unit, Li̬ge, Belgium, ‰ÛÒ

4 UNESCO-IHE Institute for Water Education, Delft, the Netherlands;

5 Flemish Institute for Technological Research (VITO), Mol, Belgium

As cities play a central role in human-environment interactions, the concepts of sustainable development and good governance have become important topics in urban policy. One of the main challenges is to safeguard and improve the quality of life in cities, while mitigating the negative effects of urban growth on the functioning of natural ecosystems. Spatially explicit data and models for mapping, analyzing and forecasting changes in urban form and function are indispensable decision-support tools to develop and evaluate strategies aimed at ensuring sustainable urban development. Earth observation satellites are useful in this respect as they provide regular information on changing urban landscapes.

The most relevant technological development in urban remote sensing is without a doubt the increased spatial resolution of sensor systems, which allows a more detailed and accurate mapping of complex urban landscapes from space.åÊ The launch in 1999 of Ikonos, the first commercial satellite capable of acquiring images with high spatial detail, was an important milestone. Subsequent missions such as Quickbird, GeoEye-1, WorldView1-2 and Pleiades have further increased the spatial resolution to 50 centimeters per pixel. Besides the trend toward higher spatial detail, Earth observation satellites providing images of moderate spatial resolution (100 meters or less) have been collecting valuable data since the early 1970s and therefore provide a historic perspective on urban growth. Archives from sensors such as Landsat and SPOT offer images with a vast potential for monitoring, modeling and understanding urban dynamics and associated environmental impacts.åÊ Such “medium-resolution” sensors provide less spatial detail, but can observe extensive areas in a relatively limited timeframe. This is valuable for synoptic monitoring of urban growth at a regional or (inter)national scale. To ensure the continued availability of these types of images, NASA successfully launched the Landsat Data Continuity Mission on Feb. 11, 2013 [1].

A recently concluded research project called Measuring and Modelling Urban Dynamics (MAMUD), funded by Belspo, investigated how high and moderate resolution satellite imagery can be used for mapping and modeling urban growth and its impact on the hydrology of the urban and suburban environment. In this paper, some of the research methods and major findings of this project will be briefly discussed. We will focus on work that was carried out on the Greater Dublin Area in Ireland.

Prior to the global financial crisis, Dublin experienced a period of unprecedented economic growth and demographic change. This led to a relatively fast expansion of the city’s functional hinterland with a dispersed pattern of residential development around peri-urban areas [2]. One of the consequences of this urban sprawl is a reduction in soil permeability as previously vegetated areas are turned into pavement or housing. Soil sealing limits infiltrationåÊ[3] and generates faster runoff within the watershed [4], increasing the risk of flooding following storm events. A higher degree of soil sealing also negatively impacts the water quality in urbanized watersheds [5].åÊ One of the objectives of MAMUD was to derive a time series of sealed surface maps from medium-resolution Earth observation data. These maps formed the basis for developing a remote-sensing supported hydrological modeling methodology to assess the impact of urbanization on runoff generation and peak discharges. Another objective was to characterize urban morphology with spatial metrics based on the sealed surface maps, and to relate these metrics to broad land-use classes. While land-use maps that are derived from medium-resolution imagery in this way are not as detailed as visually interpreted maps, they nevertheless provide an abstraction of the spatial patterns and structure of growing cities at regular intervals. In MAMUD we used these remote-sensing derived land-use maps to improve the calibration of the MOLAND urban growth model that was åÊdeveloped for Dublin in the context of the Urban Environment Project.

Deriving a consistent time series of sealed surface maps at sub-pixel level

Although satellite data of medium spatial resolution seem ideally suited for monitoring urban growth, their relative coarseness often leads to low mapping accuracies because the sensor’s viewing field usually contains multiple types of land cover, especially in urban areas. Traditional classification algorithms that derive land-cover maps from digital images assign pixels individually to a single class, and will run into difficulties when dealing with such mixed pixels. Subpixel classification addresses this problem by estimating land-cover proportions from a mixed spectral signal. In MAMUD, we used this technique for estimating the fractional cover of sealed surfaces within pixels of medium spatial resolution. The remote sensing literature provides several approaches for sub-pixel classification, including linear spectral mixture analysis [6]åÊ[7], regression analysis [8]åÊ[9], regression trees [10], soft classification with neural networks [11]åÊ[12] or self-organizing maps [13]. A first objective of our research was therefore to define a suitable method for the Dublin study area. åÊTo this end, we compared several sub-pixel approaches based on the accuracy of the estimated proportions.

Defining and validating sub-pixel models requires reliable reference proportions, i.e., sample pixels for which the degree of sealed surface cover is known. We adopted a multi-resolution approach for this purpose, which involves the use of land-cover data derived from high-resolution imagery. This raised several new issues. A first issue was that such data are not available for the same moment at which the medium-resolution images were acquired. If land cover has changed in the mean time, the supposed relationship between the spectral values of a sample pixel and its sealed surface cover may not be valid. For this reason, a procedure for temporally filtering the samples was developed [12].

A second issue was that land-cover maps of urban areas derived from images with a high spatial resolution are distorted due to shadows that hide useful information, due to classification errors caused by spectral confusion and due to classification noise (salt-and-pepper effect). We therefore improved the accuracy of the land-cover maps derived from high-resolution imagery with a post-classification approach [14].åÊ Central to this approach are a set of rules that change a pixel’s initial class label to a new class label based on the size of the land-cover patch to which it belongs and based on the adjacency to neighboring patches of a certain class.

Finally, a third issue was that the spectral information present in broadband medium-resolution imagery is insufficient to fully distinguish some types of built-up land cover from non-built land cover. This was for instance the case for red tiled roofs, which were often identified as bare soil due to spectral similarities. As this problem can only be avoided by using hyperspectral imagery [15], which was unavailable for the time-series of Dublin, we applied a method that consisted of three steps.

First, the medium resolution images that were part of the time-series were classified into urban and non-urban areas with an iterative application of an advanced unsupervised classifier [16], followed by post-classification enhancement. Given the fact that the urban masks are subject to classification errors, irrational changes between urban and non-urban land cover could not be excluded. By applying a rule-based rationality evaluation [17] adapted to masks as a second step, we were able to detect irrational land-cover changes in the time-series. Change trajectories were declared as ‰ÛÏunstable‰Û if the ‰ÛÏurban‰Û status for one year was followed by a ‰ÛÏnon-urban‰Û status in one of the following years. Irrational change trajectories were corrected by assuming that once a pixel is labelled as urban, all subsequent labels for that pixel should also be urban. In a third and final step, all pixels that were ‰ÛÏnon-urban‰Û according to the improved urban masks were considered as 0 percent sealed, while sub-pixel classification was applied to the ‰ÛÏurban‰Û pixels in order to estimate sealed surface fractions.

Figure 1. Sealed surface proportion maps for 1988 and 2006, derived by sub-pixel classification.

We compared variants of three types of sub-pixel classification models: linear spectral mixture analysis, linear regression analysis and neural networks. The non-linear models based on neural networks performed better than the linear models at the level of individual pixels, but the differences were small and often statistically insignificant.åÊ The prediction errors did, however, demonstrate a distinct pattern in function of the sub-pixel method we applied and according to the actual sealed surface proportions present within the pixels. This led to different results when the proportions were aggregated to a higher level of spatial abstraction. Linear regression was slightly more accurate in that case due the absence of an estimation bias, which proved to be higher for the non-linear methods. We therefore used a linear regression approach to produce sealed surface proportion maps (figure 1) for each medium resolution image in a time series (1988, 1994, 1997, 2001 and 2006). The mean absolute error, which represents error magnitude, was about 10 percent for most maps while error biases were close to zero (table 1).

Table 1: Mean absolute errors (MAE) and mean errors (ME) of estimated sealed surface proportions
















0.129 (30m resolution)

0.165 (10m resolution)

0.021 (30m resolution)

0.023 (10m resolution)

Assessing urbanization effects of rainfall-runoff

Hydrological models are indispensable to describe and study hydrological conditions in urban catchments. Fully distributed hydrological modeling enables spatial and temporal analysis of various water balance components, but demands spatially distributed input data that are often difficult and/or expensive to collect. The complexity and heterogeneity of urban environments make parameterization and simulation even more challenging.åÊ In MAMUD, we used remote sensing data to obtain the detailed spatio-temporal information required for hydrological parameterization and modeling of the Tolka basin, an urban watershed in Dublin (figure 1).

Figure 2: Remote sensing parameterization for hydrological modelling.

We applied a remote sensing supported hydrological modelling method to assess the impact of urban dynamics on hydrology. The methodology integrates the time series of remote sensing derived maps into WetSpa, a physically-based distributed rainfall-runoff model (figure 2). The fully-distributed grid-based approach of WetSpa makes the sub-pixel sealed surface proportion maps ideally suited to be included as spatially distributed hydrological parameter maps instead of the more commonly used but subjective expert estimations of average sealed surface cover based on land use classes. In our research, we compared the subpixel approach with the class-based approach.

Figure 3: Impact of urbanization on the discharge for the period 1988-2006. Hydrographs for a major peak event (meteorological conditions for September 2000) have been considered.

The class and subpixel based parameterization approaches show comparable calibration and validation results. We observed major differences, however, in runoff production within the urban zone and in the spatial patterns of the generated surface runoff.åÊ The subpixel approach yields considerably higher surface runoff rates and clearly shows a spatial variation within the urban fabric that is similar to that of the sealed surface cover. This confirms the key role that sealed surfaces play in urbanized catchments regarding surface runoff. The subpixel approach also yields higher peak discharges, especially for minor events (< 20må_/s). This can be explained by the fact that, during high precipitation events, most soils get saturated and the whole catchment is contributing to the total discharge. During lower precipitation events, the high runoff production zones (sealed surfaces) become relatively more important. As the sealed surface proportion maps produced by subpixel classification describe the spatial connectivity of these zones better, higher discharges can be expected when this data is used as input to the hydrological model. The impact of urbanization on peak discharge for the period 1988-2006 was assessed using the time series of five remote sensing-derived sealed surface proportion maps. Figure 3 shows the hydrographs for a major peak event for each of the five time steps. We can conclude that the urbanization that took place in the Tolka basin after 1995 had a considerable impact on peak discharge as there is an average increase of 9 percent every five years for major events [18].

Inferring land use to improve the calibration of the MOLAND model for Dublin

Analyzing urban land use change is a key element in studyingåÊ urban dynamics. Contrary to land cover, which can be directly derived from remote sensing measurements as it refers to physical properties of the Earth’s surface, land use is linked to socio-economic activities and can therefore not be directly inferred from spectral information alone. Previous studies, however, have demonstrated a strong relationship between the spatial structure of the built-up environment and its functional characteristics [19]. In the MAMUD project, we relied on this premise for evaluating the capabilities of two methods that use spatial metrics for inferring land use from remote sensing derived land-cover maps: a kernel-based approach (OSPARK)åÊ[20] and a region-based approach [21]. While previous research on mapping land use from remotely sensed imagery relies on categorical land cover data, sealed surface proportions obtained by sub-pixel classification provide an interesting alternative. Our methods thus differ from previous approaches in the sense that they rely on continuous thematic data rather than on categorical maps or raw image data.

The kernel-based approach builds on the SPARK algorithm originally proposed by Barnsley and Barr [22], by implementing adaptive kernel sizes. Kernel size is important for spatial reclassification because it determines the heterogeneity of the land-cover patterns that can be detected. As spatial heterogeneity is different for each land use type, allowing the kernel to change in size will improve the overall accuracy of the land-use classification. Our results showed that the optimized SPARK approach indeed led to a higher classification accuracy for most classes.åÊ Although a kernel-based method can be applied in the absence of ancillary data, using additional information may lead to better results. For this reason, we also developed the region-based approach, which requires recent land-use data and a road network to define areas that are homogeneous in terms of land use composition. Good accuracies were obtained in that case when the land-use classes were generalized to residential versus non-residential areas.

Figure 4. Morphological/functional maps of Dublin derived from the time-series of medium-resolution images.

Although both approaches were successful for the classification of several important classes, we were not able to produce land-use maps with a high level of thematic detail as high accuracies could only be obtained by merging classes. The reason for this is that the morphologies of some classes are not unique. Distinguishing 6 classes (non-urban land, low-density residential, medium-density residential, low-density employment, medium-density employment, and dense urban fabric) was the maximum level of thematic detail we could obtain with a reasonable accuracy from medium resolution satellite images (figure 4). Even though the resulting maps are probably too generalized to be directly useful for an end-user, they nevertheless show urban morphology and represent urban functions (residential and employment) that are important drivers of land-use changes induced by urban growth processes. As such, our morphological/land-use maps can be used for evaluating the spatial patterns present in simulated land-use maps produced by the MOLAND urban growth model. This is helpful for improving the calibration of the model as such maps can be derived from satellite images at dates for which no existing land-use maps are available.

Figure 5. MOLAND Land-use change model for Dublin.

We developed a calibration framework for land-use change modelling that is based on comparing spatial metrics calculated from the remote sensing derived maps with metrics calculated from simulated land use. Each spatial metric describes a particular aspect of urban form and structure. In our approach, we tune the parameters of the simulation model in such a way that the simulated patterns of urban growthåÊ described by the metrics match the patterns observed in the remote sensing-derived land-use maps. We tested and evaluated our framework for the historic calibration of the MOLAND model of Dublin (figure5). The various metrics selected for calibration proved to be sensitive to different components of urban growth and succeeded well in explaining the distinct urban patterns resulting from different urban planning scenarios.


The results from the MAMUD project have shown that time-series of medium-resolution images from sensors like Landsat and SPOT are valuable for monitoring urban dynamics, for calibrating land-use change models and for estimating the potential impacts of urban growth on the hydrology of watersheds. A challenge for future research is to incorporate spatial metrics in automatic calibration algorithms, which will result in more objective models. This is the subject of an ongoing spin-off project called ASIMUD, which is also financed by the Belgian Science Policy Office (BELSPO). ASIMUD investigates the possibility of automatically calibrating land-use change models by incorporating the calibration method proposed in MAMUD in a data assimilation framework. To evaluate the applicability of this framework in a local as well as in a regional context, we are testing the approach on Dublin as well as on a part of Flanders.

The research carried out for MAMUD could also be integrated into existing hydraulic models for the Tolka River, which would make it possible to produce spatially-explicit future flood-risk maps for the area. This is highly relevant to society as it would allow assessing åÊfuture infrastructure requirements and enabling adequate planning and management of water resources in the Tolka catchment.


Funding for the MAMUD project was provided by Belgian Science Policy in the frame of the STEREO II programme – project SR/00/105.


[1] J. Holden, ‰ÛÏOn the Successful Launch of the Landsat Data Continuity Mission,‰Û Earthzine, 2013.
[2] D. McInerney and C. Walsh, ‰ÛÏSpatial distribution of urban land-use change in the Dublin city-region: 1990-2006,‰Û Irish Geography, vol. 42, p. 207‰ÛÒ223, 2009.
[3] A. I. Moglen and G. E. Mejia, ‰ÛÏImpact of the spatial distribution of imperviousness on the hydrologic response of an urbanizing basin,‰Û Hydrological Processes, vol. 24, pp. 3359-3373, 2010.
[4] R. Scalenghe and F. A. Marsan, ‰ÛÏThe anthropogenic sealing of soils in urban areas,‰Û Landscape and Urban Planning, vol. 90, pp. 1-10, 2008.
[5] M. Dougherty, R. L. Dymond, S. Goetz, C. A. Jantz and N. Goulet, ‰ÛÏEvaluation of impervious surface estimates in a rapidly urbanizing watershed,‰Û Photogrammetric Engineering and Remote Sensing, vol. 70, pp. 1275-1284, 2004.
[6] F. van der Meer, ‰ÛÏImage classification through spectral unmixing,‰Û in Spatial statistics for remote sensing, Dordrecht, The Netherlands, Kluwer Academic Publishers, 1999, p. 185 ‰ÛÒ 193.
[7] C. Wu and M. A. T. , ‰ÛÏEstimating impervious surface distribution by spectral mixture analysis,‰Û Remote Sensing of Environment , vol. 84, p. 493 ‰ÛÒ 505, 2003.
[8] X. Yang and Z. Liu, ‰ÛÏUse of satellite-derived landscape imperviousness index to characterize urban spatial growth,‰Û Computers, Environment & Urban System, vol. 29, p. 524 ‰ÛÒ 540, 2005.
[9] M. E. Bauer, B. C. Loffelholz and B. Wilson, ‰ÛÏEstimating and mapping impervious surface area by regression analysis of Landsat imagery Group,‰Û in Remote Sensing of Impervious Surfaces, Boca Raton FL, CRC Press Taylor & Francis, 2008, p. 3‰ÛÒ19.
[10] L. Yang, G. Xian, J. M. Klaver and B. Deal, ‰ÛÏUrban land-cover change detection through subpixel impervious surface mapping using remotely sensed data,‰Û Photogrammetric Engineering & Remote Sensing, vol. 69, p. 1003 ‰ÛÒ 1010, 2003.
[11] W. Liu, K. Seto, S. Gopal and C. Woodcock, ‰ÛÏART-MMAP: A neural network approach to subpixel classification,‰Û IEEE Transactions on Geoscience & Remote Sensing, vol. 42, p. 1976 ‰ÛÒ 1983, 2004.
[12] T. Van de Voorde, T. De Roeck and F. Canters, ‰ÛÏA Comparison of Two Spectral Mixture Modelling Approaches for Impervious Surface Mapping in Urban Areas,‰Û International Journal of Remote Sensing, vol. 30, p. 4785 ‰ÛÒ 4806, 2009.
[13] S. Lee and R. G. Lathrop, ‰ÛÏ Subpixel analysis of Landsat ETM+ using self-organizing map (SOM) neural networks for urban land cover characterization,‰Û IEEE Transactions on Geoscience & Remote Sensing, vol. 44, p. 1642 ‰ÛÒ 1654, 2006.
[14] T. Van de Voorde, W. De Genst and F. Canters, ‰ÛÏImproving pixel-based VHR land-cover classifications of urban areas with post-classification techniques,‰Û Photogrammetric Engineering and Remote Sensing, vol. 73, p. 1017 ‰ÛÒ 1027, 2007.
[15] L. Demarchi, F. C. C.-W. Canters and T. Van de Voorde, ‰ÛÏMultiple Endmember Unmixing of CHRIS/Proba imagery for mapping impervious surfaces in urban and suburban environments,‰Û IEEE Transactions on Geoscience and Remote Sensing, vol. 50, pp. 3409 – 3424, 2012.
[16] T. Kohonen, Self-Organzing Maps, 3rd ed., Berlin: Springer, 2001.
[17] H. Liu and Q. Zhou, ‰ÛÏAccuracy analysis of remote sensing change detection by rule-based rationality evaluation with post-classification comparison,‰Û International Journal of Remote Sensing, vol. 25, p. 1037 ‰ÛÒ 1050, 2004.
[18] B. Verbeiren, T. Van de Voorde, F. Canters, M. Binard, Y. Cornet and O. Batelaan, ‰ÛÏAssessing urbanisation effects on rainfall-runoff using a remote sensing supported modelling strategy,‰Û International Journal of Applied Earth Observation and Geoinformation, vol. 21, pp. 92-102, 2013.
[19] M. J. Barnsley and S. L. Barr, ‰ÛÏDistinguishing urban land-use categories in fine spatial resolution land-cover data using a graph-based, structural pattern recognition system,‰Û Computers, Environment and Urban Systems, vol. 21, p. 209 ‰ÛÒ 225, 1997.
[20] J. van der Kwast, T. Van de Voorde, F. Canters, I. Uljee, S. Van Looy and G. Engelen, ‰ÛÏInferring urban land use using the optimised spatial reclassification kernel,‰Û Environmental Modelling & Software, vol. 26, p. 1279 ‰ÛÒ 1288, 2011.
[21] T. Van de Voorde, W. Jacquet and F. Canters, ‰ÛÏ Mapping form and function in urban areas: an approach based on urban metrics and continuous impervious surface data,‰Û Landscape and Urban Planning 102, vol. 102, p. 143 ‰ÛÒ 155, 2011.
[22] M. J. Barnsley and S. L. Barr, ‰ÛÏInferring urban land use from satellite sensor images using kernel-based spatial reclassification,‰Û Photogrammetric Engineering & Remote Sensing, vol. 62, p. 949 ‰ÛÒ 958, 1996.