Mapping global annual urban land cover fractions (2001–2020) derived with multi-objective deep learning

• Proposing a NNMRT model to unmix 6 categories ULC fractions. • Proposing a GAULCF dataset from 2001 to 2020. • Discovering the different ULC change patterns across different countries worldwide. Changes in urban land cover (ULC) provide critical evidence of urbanization including both urban expansion and internal structural renewal. Existing global urbanization research focused on urban expansion and neglected the dynamic ULC changes occurring inside urban areas. This study addresses this issue by developing a Global Annual Urban Land Cover Fraction (GAULCF) dataset, which encompasses six ULC categories (Non-crop Vegetation-Cropland-Building-Non-build Impervious-Surface-Soil-Water, V-C-B-I-S-W) in global urban areas and measures their dynamics from 2001 to 2020. However, V-C-B-I-S-W six kinds of ULC extraction is much more difficult than classical V-I-S extraction, especially for distinguishing non-crop vegetation and cropland, buildings and non-build impervious. Accordingly, this study proposes a novel deep learning unmixing algorithm, Normalized Non-negative Multi-objective Residual T-ConvLSTM (NNMRT) model, with strong encoding and recognition capacities for extracting GAULCF. GAULCF has undergone rigorous accuracy assessment and third-party validation, whose standard error of regression, root mean square error, and mean absolute error are 0.127, 0.113, and 0.061, respectively. The GAULCF dataset reveals significant global ULC changes over the past 20 years: the impervious surface area nearly doubled, while building areas increased from 124,589 km2 to 206,603 km2; and vegetation and cropland were the predominant land cover types lost to urban expansion. GAULCF’s intensive monitoring of ULC changes during urbanization aids in identifying potential issues in urban development, providing crucial insights and supporting data for addressing global challenges and achieving sustainable urban development.