Urban land cover mapping based on medium spatial resolution multispectral satellite imagery

Global urbanization has been happening for decades and will continue in the future, converting substantial amounts of natural land cover types into urban surface types. This human-caused global land cover change has induced environmental issues that can affect the quality of urban life, such as heat waves and air pollution. Accordingly, preventing and regulating these urban environmental problems has become a common concern for scientists and policy makers. To this end, obtaining spatially explicit information on the distribution and change of urban land covers may be the key to solving this problem. With global coverage, free access, and frequent observation intervals, medium spatial resolution multispectral satellite imagery has been one of the primary datasets used for mapping urban land covers at various geographical scales. However, distinguishing between spectrally and spatially heterogeneous urban surface types in these image data is challenging due to their limited spectral and spatial resolutions, thus limiting urban land cover mapping accuracy. For the same reason, all continental-scale land cover products developed from medium spatial resolution satellite imagery cannot characterize the spatial distribution of specific urban land covers, such as the urban green space. In this PhD dissertation, we tried to exploit the potential of medium spatial resolution multispectral satellite imagery in urban remote sensing by (i) improving the capability of the image data to distinguish between urban land covers spatially and spectrally and (ii) producing reliable maps of the spatial distribution of urban green spaces at the Europe continental scale. In Chapter-2, we tried to reduce the presence of mixed pixels in urban areas by increasing the spatial resolution of satellite images, thus improving the capability of the image data to distinguish between urban land covers spatially. In particular, given the spatial textures provided by the four Sentinel-2 10 m bands, we proposed an improved unmixing-based image fusion approach (UnFuSen2) to sharpen the six Sentinel-2 20 m bands to 10 m resolution. Compared to traditional unmixing-based image fusion methods, UnFuSen2 can self-adapt to the spectral variability of varying land covers and improve the image fusion accuracy by constraining the unmixing equations on the basis of spectral mixing models and the correlation between spectral bands of coarse and fine spatial resolution, respectively. In parallel, the objective of Chapter-3 was to improve the capability of satellite imagery to distinguish between urban surface covers spectrally. Given that Fisher Discriminant Analysis (FDA) can enhance the interclass spectral separability between samples of different classes and the spectral similarity of samples of the same category, Chapter-3 integrated FDA and Multiple Endmember Spectral Mixture Analysis (MESMA) (F-MESMA) for more accurate mapping of urban land cover. Our experiments demonstrated that compared to other state-of-art data transformation methods, the ratio of within- vs between-class spectral variability of urban land covers was most strongly reduced after applying the FDA. Consequently, F-MESMA consistently provided the most accurate impervious surface fraction estimates across five urban areas (RMSE F-MESMA = 0.13 vs. RMSE alternative approaches = [0.16-0.17]). Furthermore, in Chapter-4, we analyzed the individual and the combined effect of UnFuSen2, FDA, and multi-temporal observed image data on Sentinel-2-based urban land cover mapping accuracy. Our results demonstrated that the classification accuracy of UnFuSen2-processed single-date imagery, FDA-processed single-date imagery, and Sentinel-2 image time series (ITS) was higher than that of the original single-date Sentinel-2 imagery. The classification of the ITS that consists of UnFuSen2-processed single-date images showed the highest average Kappa coefficient (0.7225) compared to the classifications of other datasets. Finally, in Chapter-5, we used a machine learning-based subpixel classification approach to map urban green spaces across Europe from Landsat images in 1990, 2000, and 2015, filling a gap in the accurate extraction of urban green space information at a continental scale using medium spatial resolution multispectral satellite imagery. Our results showed that the modeled urban green area fractions yielded low RMSE values ranging from 0.09 to 0.16 across ten validation urban areas. Meanwhile, our modeled urban green space maps were validated to outperform other land cover products such as CORINE and the Urban Atlas. Based on the obtained urban green space maps, we found: (i) urban green spaces in Western European countries are more spatially concentrated, while those in Eastern and Southern Europe are relatively sparsely distributed; (ii) the green area in urban core areas (the urbanized areas before 1990) remained almost constant between 1990 and 2000 but started to increase noticeably between 2000 and 2015 throughout Europe; (iii) recent urban expansions (the urbanized areas after 1990) contain more urban green space than the increased urban green area in urban core areas from 1990 to 2015; (iv) the urban green area per capita has been increasing in Western, Eastern, and Northern Europe from 1990 to 2015, but has been declining in Southern Europe.

Publication year:2022

Accessibility:Closed