Are urban material gradients transferable between areas?

Abstract

Urban areas contain a complex mixture of surface materials resulting in mixed pixels that are challenging to handle with conventional mapping approaches. In particular, for spaceborne hyperspectral images (HSIs) with sufficient spectral resolution to differentiate urban surface materials, the spatial resolution of 30 m (e.g. EnMAP HSIs) makes it difficult to find the spectrally pure pixels required for detailed mapping of urban surface materials. Gradient analysis, which is commonly used in ecology to map natural vegetation consisting of a complex mixture of species, is therefore a promising and practical tool for pattern recognition of urban surface material mixtures. However, the gradients are determined in a data-driven manner, so analysis of their spatial transferability is urgently required. We selected two areas—the Ostbahnhof (Ost) area and the Nymphenburg (Nym) area in Munich, Germany—with simulated EnMAP HSIs and material maps, treating the Ost area as the target area and the Nym area as the well-known area. Three gradient analysis approaches were subsequently proposed for pattern recognition in the Ost area for the cases of (i) sufficient samples collected in the Ost area; (ii) some samples in the Ost area; and (iii) no samples in the Ost area. The Ost samples were used to generate an ordination space in case (i), while the Nym samples were used to create the ordination space to support the pattern recognition of the Ost area in cases (ii) and (iii). The Mantel statistical results show that the sample distributions in the two ordination spaces are similar, with high confidence (the Mantel statistics are 0.995 and 0.990, with a significance of 0.001 in 999 free permutations of the Ost and Nym samples). The results of the partial least square regression models and 10-fold cross-validation show a strong relationship (the calculation-validation R2 values on the first gradient among the three approaches are 0.898, 0.892; 0.760, 0.743; and 0.860, 0.836, and those on the second gradient are 0.433, 0.351; 0.698, 0.648; and 0.736, 0.646) between the ordination scores of the samples and their reflectance values. The mapping results of the Ost area from three approaches also show similar patterns (e.g. the distribution of vegetation, artificial materials, water, and ceremony area) and characteristics of urban structures (the intensity of buildings). Therefore, our findings can help assess the transferability of urban material gradients between similar urban areas.

Keywords

• Hyperspectral image
• Urban mapping
• Gradient analysis
• Transferability
• Imaging spectroscopy