You are here

Warning message

Attention! This event has already passed.

Mapping, monitoring and modeling urban areas with medium-resolution satellite imagery: A multi-resolution approach for characterizing the dynamics of urban form and function

Thursday, 10 November, 2011 - 16:30
Campus: Brussels Humanities, Sciences & Engineering campus
Faculty: Science and Bio-engineering Sciences
auditorium P. Janssens
Tim Van De Voorde

The majority of the world's population lives in cities, which together generate over half of the
global GDP. Safeguarding and improving the quality of urban environments has become an
important aspect of sustainable development strategies, and has kindled interest in monitoring
cities in terms of quality of life and socio-economic development. Effective urban management
and planning strategies are also required to temper the negative consequences of urban growth
on ecosystem functions and human society. To develop and evaluate such strategies it is
necessary to assess their spatial impact, and this requires data for mapping, analysing and
forecasting changes in the urban fabric. Data from earth observation satellites provide regular
information on urban development and could in that way contribute to mapping and monitoring
the morphology of expanding cities. Sensors of the new generation produce images with a high
level of spatial detail, but they also have a small coverage, a low spectral resolution, a high price
and a limited historic archive. Mid-scale images from older missions, on the other hand, are
cheap, often available for free and there are archives dating back to the early seventies.

The general objective of this thesis is to exploit these advantages by developing methods for
mapping, monitoring and modelling urban areas based on satellite images of medium spatial
resolution. A large part of the thesis is dedicated to mapping two elementary land cover
components of cities: built-up land (i.e. "impervious surfaces") and vegetation. The
disadvantage of using data with a lower spatial resolution for this purpose is that a single sensor
observation usually covers multiple land-cover types and therefore represents a composite
spectral signal. Conventional classifiers allocate each measurement to a single land-cover class,
and this inevitably leads to classification errors and wrong estimations of areal coverage. We
addressed this problem with subpixel classification, which allows the estimation of fractional
land-cover composition inside an image pixel. Models for inferring proportions from a mixed
spectral response were built and validated in a multi-resolution framework. This involves the
use of land-cover maps derived from high-resolution imagery as reference data. In order to
ensure the reliability of these maps, three post-classification techniques were used to assign
shadow pixels to meaningful land-cover classes, to correct wrongly classified pixels and to
reduce high frequency structural clutter. This procedure was successfully applied to pixel-based
classifications of high spatial resolution images, improving the thematic accuracy and increasing
the information content. For deriving subpixel proportions of built-up area or vegetation, we
applied variants of three types of models: linear spectral mixture analysis, linear regression
analysis and neural networks. Non-linear models performed better than the linear models at the
level of individual pixels, but the differences were small and often not significant. The
prediction errors did, however, demonstrate a distinct pattern in function of the actual landcover
proportions within the pixels. This led to different results when the fractions were
aggregated to a higher level of spatial abstraction. Linear regression was slightly more accurate
in that case due the absence of estimation bias, which was more pronounced for the non-linear

The subpixel approach was applied in a case study on Brussels for which changes in impervious
surface and vegetation cover were mapped at 4 dates between 1978 and 2008. First,
conventional classification methods were used to derive land-cover maps with broad classes:
built-up land, open vegetation, dense vegetation and water. A linear regression model was then
used to estimate vegetation proportions for each pixel belonging to the "built-up" class. At least
4500 hectares of land were consumed by urban expansion in the observed period, and this has
led to an increased fragmentation of the landscape. Urban growth also appeared to accelerate
with time as most changes occurred in the most recent interval (2001-2008). Not all of this
newly urbanised land consists of impervious surfaces, however, given that the subpixel
classification indicated a more moderate decline in vegetation cover of around 1900 hectares.
This is due to the morphology of the new developments, which is mainly residential land use
with a low built-up density.

Land use is linked to socio-economic activities and can therefore not be directly inferred from
spectral information. As previous studies have indicated that urban form is related to function,
we have examined the potential for inferring broadly defined land-use categories from maps of
impervious surface cover obtained by subpixel classification, and tested our approach on the
Greater Dublin Area. The distribution of impervious surface cover was described at the scale of
predefined spatial entities by three urban metrics: the average impervious surface cover, shape
characteristics of the cumulative frequency distribution of impervious surface proportions and
spatial variance of these proportions. Morphology was related to land use by applying
supervised classification with the metrics as variables and with generalised land-use categories
as target classes. Combined with built-up densities derived from the impervious surface data,
the classification produced morphological/functional maps that clearly show the urban
dynamics in Dublin between 1988 and 2001. A good overall accuracy was achieved when a
distinction was made between two broadly defined classes representing employment-related
and residential land-use. Changes in the spatial pattern of these two classes are important in the
context of calibrating urban growth models as these classes represent important drivers of
urban land-use change.

Urban growth models are useful tools for assessing and comparing the environmental impact of
alternative policy scenarios. Their increasing popularity as spatial planning instruments also
poses new scientific challenges, such as correctly calibrating the model. The challenge in model
calibration is twofold: obtaining a reliable and consistent time series of land-use information
and finding suitable measures to compare model output to reality. Both these issues were
addressed in the final chapter of this thesis, in which we propose a model calibration framework
that is supported by information on urban form and function derived from medium-resolution
remote sensing. The remote sensing derived maps were compared to model output of the same
date for two growth scenarios, using well-known spatial metrics as goodness-of-fit measures.
The analysis indicated that many of the selected metrics were sensitive enough to detect the
differences between a reference scenario from a previous calibration exercise and a deliberately
unrealistic urban growth scenario. Most of these sensitive metrics also indicated a good fit
between the reference scenario and the remote sensing derived maps, which allowed a
tentative suggestion of a suitable set of spatial metrics for assessing the goodness-of-fit
between model output and remote sensing derived information in the calibration of urban
growth models.