Fast Unsupervised Multi-Scale Characterization of Urban Landscapes Based on Earth Observation Data

Abstract

Most remote sensing studies of urban areas focus on a single scale, using supervised methodologies and very few analyses focus on the “neighborhood” scale. The lack of multi-scale analysis, together with the scarcity of training and validation datasets in many countries lead us to propose a single fast unsupervised method for the characterization of urban areas. With the FOTOTEX algorithm, this paper introduces a texture-based method to characterize urban areas at three nested scales: macro-scale (urban footprint), meso-scale (“neighbourhoods”) and micro-scale (objects). FOTOTEX combines a Fast Fourier Transform and a Principal Component Analysis to convert texture into frequency signal. Several parameters were tested over Sentinel-2 and Pleiades imagery on Bouake and Brasilia. Results showed that a single Sentinel-2 image better assesses the urban footprint than the global products. Pleiades images allowed discriminating neighbourhoods and urban objects using texture, which is correlated with metrics such as building density, built-up and vegetation proportions. The best configurations for each scale of analysis were determined and recommendations provided to users. The open FOTOTEX algorithm demonstrated a strong potential to characterize the three nested scales of urban areas, especially when training and validation data are scarce, and computing resources limited.

Keywords

• remote sensing
• multi-scale
• unsupervised
• urban landscapes
• texture