Urban mapping

Introduction

Urban ecosystem analysis requires information on physical composition and structure of the urban environment. Multispectral satellite images with medium spatial resolutions between 10 and 30 m are useful for quantitative sub-pixel mapping. While such satellite imagery is generally available at regional to global scale and at high temporal resolution, the coarser spatial resolution largely fails to preserve the detail of the urban mosaic. Hyperspectral airborne images with high spatial resolution of 5 m and below are another important data source for urban mapping, particularly for detailed classification of construction materials, low vegetation types or tree species. Altimetric airborne sensors can produce a detailed 3D model of the urban environment and are increasingly used to complement 2D optical imagery. Despite high spatial detail, airborne imagery is often limited to sporadic acquisition of a small number of narrow flight lines. No sensor is ideally suited for mapping every aspect of the urban environment. In UrbanEARS we therefore investigate the potential of combining different sensors in operational mapping workflows that produce accurate land cover maps with an optimal trade-off between thematic detail, resolution and coverage.

Objectives

  1. Explore innovative unmixing approaches to operationalize urban LC mapping using hyperspectral imagery.

  2. Assess the transferability of information from high to low spectral-spatial resolution imagery.

  3. Assess the transferability of spectral information and models between different types of urban environments.

Our Approach

BandClust Dimensionality Reduction

Hyperspectral data can be cumbersome to manage due the high number of bands involved. High data dimensionality may also cause collinearity among bands and decrease mapping performance. Although widely used, Principal Component Analysis is not suited for hyperspectral dimensionality reduction due to its inability to retain all important features. We therefore used the BandClust algorithm, which uses Mutual Information (MI) as a criterion to merge adjacent bands into a small number of easy to interpret clusters. We show that BandClust retains or even slightly improves mapping performance compared to using all bands, while drastically decreasing computational demand.

Shadow Mapping

Urban high-resolution remote sensing imagery is often strongly affected by shadow cover, which reduces its information content and quality of derived maps. To assess the seriousness of shadow cover, we use LiDAR derived Digital Surface Models (DSM). Shadow maps are produced using iterative projection of the DSM in the direction of cast shadow, based on the solar azimuth and height at time of image acquisition. Shadow maps are also useful for targeted mapping and validation of shaded areas in an image.

Synergistic Use of Hyperspectral and LiDAR Data

Airborne high-resolution hyperspectral imagery, such as APEX, has proven to be useful for detailed urban material mapping, particularly when performed with strong machine learning classifiers. Such mapping does remain susceptible to spectral confusion between land cover classes with similar material composition. By combining geometric information with spectral information, much of this confusion can be avoided. We therefore use LiDAR derived height and slope features to increase material mapping accuracy in post-processing.

Universal Mapping Workflow

A universal mapping workflow is developed within the frame of UrbanEARS. The workflow flexibly adapts to the thematic requirements of the end-user while considering spatial, spectral and temporal characteristics of available remote sensing data. A generic spectral library, representing a general database on the diversity and variability of urban cover types, represents a key input for the workflow. The library consists of endmembers with pure cover materials at high spectral resolution and detailed hierarchic thematic labels. The other main input is a pool of available images with varying characteristics. Depending on user requirements, the spectral library is optimized for the appropriate image. Optimization firstly entails harmonization of library and image spectra. Selection or so-called pruning of library subsets is then used to enhance identification and separation of urban surface types present in the image. Optimized library spectra are used as input to develop mapping models.

SVR, Synthetic Mixing and Sentinel-2

Support Vector Machines (SVM) is a machine learning algorithm that is widely used for mapping of urban areas. It allows modelling of complex non-linear decision boundaries through higher dimensional transformation of the image feature space. Support Vector Regression (SVR) can be used for sub-pixel fraction mapping, but requires quantitative mixed data for model training, which is typically derived from expensive high-resolution land cover maps. Spectral libraries derived from hyperspectral imagery are detailed sources of training data but are discretely labelled. These spectra can however be converted to quantitative training data by means of synthetic mixing, i.e. library-based training. We explore use of an APEX derived hyperspectral library, synthetic mixing and SVR to produce Vegetation – Impervious – Soil fractions maps from 20 m Sentinel-2 imagery. Comparison is made with classic map-based training based on high-resolution land cover data. We show that library-based training performs similar compared to map-based training while having the advantage of being less user and data demanding.

Publications

 

Okujeni, A., van der Linden, S., Tits, L., Somers, B. and Hostert, P. (2013). Support vector regression and synthetically mixed training data for quantifying urban land cover. Remote Sensing of Environment, 137, 184-197.

Okujeni, A., van der Linden, S., & Hostert, P. (2015). Extending the vegetation–impervious–soil model using simulated EnMAP data and machine learning. Remote Sensing of Environment, 158, 69-80.

Degerickx J., Iordache, M.D., Okujeni, A., Hermy, M., van der Linden, S. & Somers, B. (2016). Spectral unmixing of urban land cover using a generic library approach. Remote Sensing Technologies and Applications in Urban Environments. Proc. of SPIE. Edinburgh 2016, p. 13. DOI: 10.1117/12.2241189

Priem F. & Canters F. (2016). Synergistic use of LiDAR and APEX hyperspectral data for high-resolution urban land cover mapping. Remote Sensing, 8, pp.1–22, DOI: 10.3390/rs8100787.

Priem F., Okujeni A., van der Linden S. & Canters F. (2016). Use of multispectral satellite imagery and hyperspectral endmember libraries for urban land cover mapping at the metropolitan scale. Remote Sensing Technologies and Applications in Urban Environments. Proc. of SPIE. Edinburgh 2016, p. 13. DOI: 10.1117/12.2240929

Canters, F., Priem, F., Okujeni, A. & van der Linden, S. (2017). Machine learning based unmixing of urban land cover from Sentinel-2 using synthetically mixed training data (abstract + oral presentation). International Cartographic Conference 2017, ICA, Washington DC

 

Degerickx, J., Okujeni, A., Iordache, M.D., Hermy, M., van der Linden, S., Somers, B. (2017). A novel spectral library pruning technique for spectral unmixing of urban land cover. Remote Sensing, 9(6), 565.

 

Okujeni, A., van der Linden, S., Suess, S. & Hostert, P. (2017). Ensemble learning from synthetically mixed training data for quantifying urban land cover with support vector regression. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.

Okujeni, A.; van der Linden, S., Priem, F., Canters, F., Degerickx, J., Somers, B. (2017). Towards The Development Of Universal Support Vector Regression Models For Quantifying Land Cover Across Cities. Abstract submitted to the 10th EARSeL SIG Imaging Spectroscopy Workshop, Zurich, Switzerland.

Priem, F., Okujeni, A., Van der Linden, S. & Canters, F. (2017). Optimizing mixed spectra generation for regression-based unmixing of land cover in urban areas. IEEE Conference proceedings (abstract + oral presentation). JURSE 2017 Dubai.

Rabe, A.; Jakimow, B.; Held, M.; Okujeni, A.; Leitão, P.; Hostert, P.; van der Linden, S. (2017). EnMAP-Box 3.0 - concept for a QGIS-Python toolbox for imaging spectroscopy data processing. Abstract submitted to the 10th EARSeL SIG Imaging Spectroscopy Workshop, Zurich, Switzerland.

 

Roberts, D.A., Alonzo, M., Wetherley, E., Dudley, K., and Dennison, P. (2017). “Multiscale Analysis of Urban Areas Using Mixing Models.” In D.A. Quattrochi, E. Wentz, N.S. Lam & C. W.  Emerson (ed.) Integrating Scale in Remote Sensing and GIS, 247.

 

Small, C., Okujeni, A., van der Linden, S., Waske, B. (2017). Remote Sensing of Urban Environments. In Reference Module in Earth Systems and Environmental Sciences.

Wetherley, E., Roberts, D., McFadden, J. (2017). Mapping spectrally similar urban materials at sub-pixel scales. Remote Sensing of Environment, 195, 170-183.

Degerickx, J., Roberts, D.A., Somers, B. (2019). Enhancing the performance of Multiple Endmember Spectral Mixture Analysis (MESMA) for urban land cover mapping using airborne lidar dataand band selection. Remote Sensing of Environment, 221, 260-273.

This project is funded by the Belgian Federal Science Policy (Belspo) within the RESEARCH PROGRAMME FOR EARTH OBSERVATION - “STEREO III”.

© 2015 UrbanEARS