A new spectral library developed to map urban fabric

A novel new spectral library is developed using high spatial and spectral resolution measurements for a wide range of European urban wall, roof and ground materials. Identifying urban surface materials can contribute to urban weather and climate modelling, inform urban planning processes, and support the evaluation of climate resilience strategies. However, given the vast variations in material composition, the three-dimensional structure of the urban canopy and its spatial variability, mixed pixels exist even in high-resolution products (e.g., 1.2 m WorldView-3). Existing spectral libraries do not cover the diversity of urban materials needed for image-based surface cover classification. Using a HySpex Mjolnir VS-620 hyperspectral camera mounted on a drone or ground rotation stage allows observations under a wide range of shading, weathering states, and view angles to be gathered across a spectral range of 410 to 2500 nm. The images are calibrated to reflectance using a Spectral Evolution RS-3500 spectroradiometer. The variations of materials are analysed and labelled based on their colour, roughness, weathering and orientation, with multiple samples per class. The library includes spectra aligned (using the spectral response functions) for Sentinel-2, Landsat-8/9, WorldView-2/3, EnMAP, PRISMA and Planet SuperDove bands, enabling direct training of models for the individual sensors. To assess the library’s utility for identifying materials of an entire city, we classify a WorldView-3 image of Heraklion (Greece), employing a machine learning model trained using only the library. Given WorldView-3’s 3:1 ratio of VNIR and SWIR spatial resolutions, we develop a new co-registration and resolution harmonisation procedure to align all bands to the VNIR, producing a 16-band cube. The library accurately identifies urban surface materials with an overall accuracy of 80.3%, outperforming image-based approaches. This approach eliminates labour-intensive, image-specific endmember collection and requiring less training samples. The new library’s high spectral resolution and vast number of samples enables multiple applications using various sensors, avoiding limitations imposed by using image-derived endmembers from specific scenes and platforms.