< Back to previous page

Project

Mapping urban composition and green infrastructure using remote sensing in support of urban ecosystem service assessment

Urbanization presents one of the major challenges to humankind in the current century. Our cities are true drivers of global environmental change, but at the same time also represent the most susceptible areas to be suffering from the local impacts of these ongoing changes (e.g. through heat waves, flooding and air pollution). Sustainable urban management and development therefore focuses on safeguarding the local quality of life by reducing both the local and global impacts of urbanization. Urban green is increasingly recognized as a valuable means in this respect, given the many benefits (or ecosystem services) it may provide to human society. Mapping these services in a quantitative and spatially-explicit way, would enable urban planners and managers to identify those zones within a city’s boundaries that should be prioritized in future urban development and would allow for a critical and objective evaluation of different urban planning scenarios. Generating such ecosystem service maps however requires detailed information about urban land cover composition in general, and concerning the specific type, properties and state of urban green in particular. By measuring the interaction of solar radiation with the earth’s surface in high spatial and spectral detail, airborne hyperspectral remote sensing in theory allows for the detailed characterization of the urban environment. However, the high spatial and spectral complexity of urban areas still actively impedes such detailed assessments. In this PhD dissertation, we therefore looked into the concept of data fusion to overcome specific remaining issues in this respect. Our study area comprises the eastern part of the Brussels Capital Region (Belgium), which has been covered by 2 m resolution hyperspectral data in Summer 2015.

The highly heterogeneous nature of our cities lead to the phenomenon of mixed pixels in remotely sensed imagery. Several spectral unmixing approaches have therefore been proposed in the past, aiming to determine the true composition of individual image pixels. Multiple Endmember Spectral Mixture Analysis (MESMA) is amongst the most commonly used techniques and relies on a spectral library (i.e. a collection of pure material spectra, or endmembers), ideally covering all endmember variability present within a given scene. Despite the advent of several automated endmember extraction techniques, building such libraries represents a time-consuming, yet crucial task. In Chapter 2, we therefore proposed fusing already existing urban spectral libraries as the basis for an alternative solution. As this generic urban spectral library is expected to contain a high proportion of irrelevant spectra with regard to any particular image to be processed, we developed an automated endmember selection technique (AMUSES) allowing to refine a given spectral library in function of a given image. Several experiments on simulated and real hyperspectral imagery with libraries of increasing generic nature confirmed the potential of AMUSES in this respect and have additionally shown a significant increase in subsequent mapping accuracies compared to more traditional library pruning techniques. Despite these improvements, considerable classification errors were still observed, which could be attributed to the high spectral similarity between land cover classes (e.g. roof versus pavement and grass versus tree). In Chapter 3, we therefore integrated height information extracted from airborne LiDAR data into the MESMA algorithm and successfully reduced confusion between spectrally similar, but structurally different land cover classes. In particular, height distribution information within single image pixels was employed as an additional endmember selection constraint and as fraction constraints during the unmixing procedure, thereby also reducing computation times by up to 85 %. Band selection (i.e. using different, relevant subsets of spectral bands for each individual land cover class) did not further improve classification results, but resulted in an additional decrease in computation times by 50 %. The added value of the proposed techniques for processing imagery featuring lower spectral and/or spatial resolution has been demonstrated for both the library pruning technique and the integration with LiDAR. Given the increasing availability of (hyper-)spectral data and the associated need for highly automated and efficient processing algorithms, both chapters are expected to contribute towards the development of more universal urban mapping workflows.

In order to facilitate the mapping of ecosystem services provided by urban green, a functional urban green typology, covering 23 distinct types, was established in Chapter 4. Given the high spectral and structural similarities between the proposed types, object-based image analysis, combined with Random Forest classification, was employed as a more advanced image fusion technique to further explore the complementarity between airborne hyperspectral and LiDAR data. Height and intensity derived from LiDAR data were found to be the most important features overall, but required additional spectral information to accomplish good classification results at a thematically detailed level. In this respect, hyperspectral data was found to be more useful compared to multispectral data, although the latter did feature a higher spatial resolution. Despite these encouraging results, spatially continuous mapping of urban green still was severely impeded by shadow and adjacency effects, resulting in class-wise kappa values below 0.5 for detailed shrub and herbaceous vegetation types. Additionally incorporating phenological information and adopting multi-scale segmentation approaches, is expected to further increase the potential of remote sensing for detailed urban green mapping.

Finally, in Chapter 5, chlorophyll concentration and Leaf Area Index (LAI) of urban trees were determined using hyperspectral and LiDAR data, and were subsequently combined into an objective estimation of tree health. Similar to the findings in Chapter 4, mixed pixel effects significantly complicated the analysis. As a result, Partial Least Squares regression, being able to learn from local calibration data and employing the full hyperspectral signal, highly outperformed existing spectral indices. As our tree health assessment showed good agreement with visual tree assessment data obtained on the ground, the proposed workflow could be used as a basis for further research focusing on revealing the drivers of urban tree health. Further efforts should additionally be devoted to the early detection of stress in urban trees in order to optimize the utility of this tool for urban green managers.

Although the concept of ecosystem services has certainly increased the awareness on the importance of urban green for safeguarding the quality of life in our future cities, implementation of these concepts into urban management and design is still lagging behind. The research conducted throughout this PhD dissertation has confirmed the potential of remote sensing data and technology to contribute to the detailed assessment of urban ecosystem services, as such providing an important stepping stone towards their operational use. Due to the typically high spatial and spectral complexity of our cities, the urban remote sensing community is highly encouraged to continue the search for complementary information derived from both new (e.g. social media, sensor networks) and existing data sources in order to optimize the workflows proposed here.

Date:1 Jan 2015 →  17 Oct 2019
Keywords:Hyperspectral remote sensing, LiDAR, Urban green
Disciplines:Ecology, Environmental science and management, Other environmental sciences, Forestry sciences, Physical geography and environmental geoscience, Communications technology, Geomatic engineering, Landscape architecture, Art studies and sciences
Project type:PhD project