MAUPP Methodology


The MAUPP project will work at two different spatial resolutions: HR (~30 m) and VHR (< 5m). The Morphological Urban Area (MUA) and urban land uses will be delineated, mapped and predicted mainly using SAR and optical remote sensing as input data. In addition, spatially-detailed census data, historical maps and existing population distribution datasets will be used for modelling and predicting urban expansion and human population densities across Africa.

MUA delineation using HRRS data

At the continent scale, the MUA of 50 African cities will be delineated using combined high resolution (HR ~30 m) archived SAR and optical data. Information fusion extracted from the two datasets will allow the exploitation of redundancy, in order to increase global information and complementarity and to improve certainty and precision.

Multi-temporal processing will be conducted for the years: 1980, 1990, 2000, 2010 and 2015. Due to this large multi-temporal dataset of 50 African cities we propose to apply a rapid and entirely automatic approach to extract human settlement from the SAR and optical HRRS data.

Preparation of the data through pre-processing procedures will include speckle reduction of SAR data and co-registration of SAR and optical scenes of the same year. Further, the processing method developed by Dell’acqua et al. (60) and which combines co-occurrence matrix and semi-variogram analysis will be applied for mapping urban density using satellite SAR data. It has been shown that in terms of final classification accuracy, the joint use of those two features to optimize the texture window size can be nearly as effective as an exhaustive search (61). The texture features images will be classified using a fuzzy K-means clustering method (56). The main advantage of such a classification is that it provides, for each pixel, membership degrees for the urban cluster thus offering the opportunity to handle the output for further processing, namely for the information fusion process. The Landsat-TM data will be used for the creation of landcover maps (contain four main classes: settlements, roads, rivers and vegetation) using spectral based classifier described in (44, 62).

Information on urban areas extracted from SAR and optical data will be combined using a fuzzy framework. This features fusion approach is particularly flexible due to existence of a variety of fuzzy combination operators, which may deal with heterogeneous information (63). At the end of the processing chain, a final delineation of the urban area will be extracted based on the maximum of membership values after the combination step. For each city, feature based change detection will be applied for processing and analyzing the multi-temporal urban extent.

Mapping land use using VHRRS data

At the city scale, the land use will be mapped within the MUA using a robust methodology based on combined very high resolution (VHR) SAR and optical data in an Object-Based Image Analysis (OBIA) approach. Pleiades and RADARSAT-2 data will be acquired for three selected African cities. Pre-processing will include georeferencing, atmospheric correction, creation of 3D data and by-products (DTM and DSM) and speckle reductions for the SAR data.

Layover, shadowing and multiple-bounce effects are characteristic elements of VHR SAR scenes of urban landscapes. In dense built-up areas, multiple bounce mechanisms of individual construction elements lead to complex backscatter patterns that complicate the reliable detection of individual buildings. One way to overcome these limitations is to use polarimetric information in order to obtain height estimation of buildings and terrain and by determining the landcover scattering characteristics (64). Moreover, the polarimetric properties, and scattering mechanism can be extracted by applying several polarimetric decompositions which can be used directly as classification features in object-based classifiers such Support Vector Machine SVM (65). Several PolSAR decompositions will be applied (64) for extracting textural and morphological features from the scenes. These features will be directly used as input bands to the fusion classifier.

Once pre-processed, the fused data will be segmented using the versatile and popular OBIA software eCognition with its Server licence allowing to process large datasets. Considering the spatial complexity of the urban land cover and the increased spatial resolution of VHR sensors, OBIA is largely used to map urban land cover (66). Avoiding the “salt-and-pepper” effect, it allows the fusion of spectral, textural, morphological and contextual features during the classification process (67, 68). The segmentation will be constrained by the relevant OSM layers (building, roads, …) (69). For each region, eCognition will be used to extract features from optical (spectral, textural and morphological), VHR SAR features, and ancillary data (e.g. height). The relevant features will be chosen according to an ontologic approach (70). For each city, African experts in remote sensing will define land cover interpretation keys and illustrate them with located ground pictures. These interpretation keys formalising the expert knowledge will then be used to select quantitative features. The expert knowledge on the 3 cities will also be used to establish an urban standardised land cover scheme. The object classification itself will be performed in the R statistical software as it allows testing advanced classifiers such as Supported Vector Machines (SVM), Random Forest (RF), Decision Trees (DT) or K-nearest neighbour (k-NN).

To validate the classification results, a large set of stratified random points will be interpreted using Google Earth; ground survey will be limited to difficult objects and/or classes. Per-pixel accuracy assessment will be conducted at regular time intervals in order to follow-up the progress in obtained accuracies. Emphasis will be put on improvements in producer’s and user’s accuracies of the different artificial classes.

While documenting the urban land cover types, African experts will also identify and document the residential land use patterns encountered in each city. This knowledge will be centralised and validated with the literature in urban geography over African cities. A common typology of residential land use will be proposed. The residential land use types will be identified for each urban block by a quantitative discriminant analysis of landscape indices computed on the land cover map, complemented by other indicators such as the distance to the city centre or the environmental conditions (e.g. located close to a marsh or on steep slopes).

Modelling MUA changes

A generalizable urban expansion model will be developed at HR using the urban land cover change database generated for 50 cities and for the years 1980, 1990, 2000, 2010 and 2015. Boosted Regression Trees (BRT) models will be used to calculate a probability of rural to urban conversion for each non-urban 30m pixel based on a set a covariates such as the accessibility, the topography or the proportion of urban pixels in the neighborhood. Cross-validation techniques will be used to calibrate and validate the model. Then, using urban growth rates predicted by the UN (3) and the probability raster, the evolution of the spatial pattern of the MUA will be predicted for each African city. Methods described here follow approaches already developed before by LUBIES but that suffered from limited urban change training datasets (42). Here, the large database created using HRRS data will allow the production of Africa-wide urban predictions more spatially detailed than what has been done before. In addition, VHRRS data assembled for 3 cities will allow to refine the covariates used in the HR model and test other geographic determinants of the MUA evolution such as roads, markets, central business district or non-aedificandi zones. Moreover, the MUA of the 3 cities will be mapped through time at VHR using a collection of historical maps since colonial time and the contemporary VHRRS land use – land cover map of the city. African experts and relevant literature will complement this mapping effort; the aim being to document the history of the city and identify the geographic factors that contribute or limit its extension.

Mapping intra-urban population densities

The most recent and the most spatially detailed census data will be acquired for the 3 cities selected as case studies. Geo-statistical methods will be used to study the relationships between census-derived population densities and VHRRS-derived attributes of residential land use types. VHRRS-derived attributes will include (i) landscape metrics computed using the land cover map, (ii) spectral, textural and morphological features derived from VHRRS data and (iii) geographical variables susceptible to explain the distribution of populations such as distance to the city center, distance to main roads, presence of colonial non-aedificandi areas, presence of military camp, etc. Based on these relationships, population weights will be assigned to the different land use classes and populations will be redistributed within each administrative census unit according to these weights. A very spatially detailed population density map will then be created for the 3 cities.

Geo-statistical methods will then be used to explore relationships between population densities calculated using VHRRS data and HRRS-derived attributes such as (i) landscape metrics derived from the HR urban maps, (ii) spectral and textural features derived from HRRS data and (iii) geographical layers available for the whole African continent such as road and railway networks. The main HRRS-derived attributes related to population densities will be used to build a model to predict intra-urban variations in population density in any large African city. The model will be calibrated and validated using detailed population census data assembled by the AfriPop project.

Integration and dissemination

Throughout the project, efforts will be made to develop automatic approaches, especially at HR, in order to make a maximum of models and results generalizable to the rest of the African continent. The urban expansion model and the population density model will be integrated into the Random Forest algorithm recently developed by DGE-US to disaggregate census population counts from administrative units to 100 m pixels. In addition to classical dissemination tools as participation in dedicated workshop and conferences, publication through web-site and academic peer reviewed journals, the project aims to contribute to the AfriPop/WorldPop project (www.worldpop.org.uk).

The Afripop/Worldpop project is an international initiative that has developed over the last 5 years to produce 100-m resolution maps of human population at global scale, and linking DGE-US, the University of Florida and ULB-LUBIES. Dissemination through the AfriPop activities will ensure a rapid dissemination of the project deliveries to the main end-members and stakeholders. Management and coordination MAUPP will be coordinated by the Biological Control and Spatial Ecology (LUBIES) department of the ULB. The purpose of the project management is to ensure that the project is properly coordinated and that the co-operation amongst the consortium members works smoothly and efficiently.

References

  1. Dell’Acqua F et al. (2006) Semi-automatic choice of scale-dependent features for satellite SAR image classification. Pattern Recognit Lett 27:244–251.
  2. Dell’Acqua F, Gamba P (2010) in Radar Remote Sensing of Urban Areas (Springer), pp 49–68.
  3. Tatem AJ et al. (2004) Defining approaches to settlement mapping for public health management in Kenya using medium spatial resolution satellite imagery. Remote Sens Environ 93:42–52.
  4. Oussalah M (2000) Study of some algebrical properties of adaptive combination rules. Fuzzy Sets Syst 114:391–409.
  5. Shimoni M et al. (2009) Fusion of PolSAR and PolInSAR data for land cover classification. Int J Appl Earth Obs Geoinformation 11:169–180.
  6. Xiu N Multitemporal Spaceborne Polarimetric SAR Data for Urban Land Cover Mapping. Dissertation (Institute of Technology (KTH)).
  7. Blaschke T (2010) Object based image analysis for remote sensing. ISPRS J Photogramm Remote Sens 65:2–16.
  8. Carleer AP, Wolff E (2006) Urban land cover multi-level region-based classification of VHR data by selecting relevant features. Int J Remote Sens 27:1035–1051.
  9. Aubrecht C et al. (2013) Multi-level geospatial modeling of human exposure patterns and vulnerability indicators. Nat Hazards 68:147–163.
  10. Carleer AP, Wolff E (2008) in GEOBIA (Calgary, Canada).
  11. Kohli D et al. (2012) An ontology of slums for image-based classification. Comput Environ Urban Syst 36:154–163.