30-m Urban Land Expansion Projections for Sub-Saharan Africa (2022–2050)

This dataset provides projections of urban land expansion in Sub-Saharan Africa from 2022 to 2050 at 30-m resolution. Pixel values indicate the projected year of urban conversion. Specifically, the pixel value plus 2000 gives the year in which the pixel is projected to become urban. For example, a pixel value of 23 indicates urban conversion in 2023, and a pixel value of 50 indicates urban conversion in 2050. The projections were generated using the deep-learning-based spatial-temporal fusion model for urban growth simulation (STUrban), with 2022 urban land derived from the Annual Global Land Cover (AGLC) dataset as the baseline and future urban land increments estimated from Africapolis-derived country-level growth rates. STUrban, originally proposed by Zhou and Chen (2025), separately models the spatial and temporal effects of urban growth and then fuses them to improve simulation performance. The model consists of three major components: Random Forest models the spatial effects of urban growth from driving factors, ConvLSTM captures temporal growth trends from annual urban land sequences, and 3D-STF-CNN fuses these effects to generate annual urban development probability maps, which are then combined with annual urban land increments to produce the simulation results. The production of this dataset followed a validation-before-projection workflow, in which STUrban was first evaluated by comparing simulated historical urban expansion with observed urban land maps, and then applied to generate future annual urban land maps. The model was implemented in two stages. In the validation stage, the RF was trained using urban land expansion from 2006 to 2014, while ConvLSTM and 3D-STF-CNN were trained in an end-to-end manner using annual urban land maps from 1999 to 2006 as input and maps from 2007 to 2014 as the reference output. The trained models were then tested by using maps from 2007 to 2014 to simulate urban expansion from 2015 to 2022. The simulated results were compared with observed urban land maps. In the simulation stage, the RF was trained using urban land expansion from 2014 to 2022, while ConvLSTM and 3D-STF-CNN were trained using maps from 2011 to 2016 as input and maps from 2017 to 2022 as the reference output. The trained models were then applied iteratively to generate annual urban land maps by 2050. Model performance was evaluated using RF AUC, cell-level Figure of Merit (FoM), and neighborhood-level agreement within 1.5 km × 1.5 km moving windows. RF AUC values reached 0.955–0.976 in training and 0.943–0.965 in application. FoM values exceeded 0.2 in most training regions and approached or exceeded 0.35 in some regions. Compared with the FLUS model, STUrban achieved higher FoM values in almost all training regions, with an advantage greater than 0.05 in nearly half of them. At the neighborhood level, STUrban explained 60%–62% of observed urban growth across sub-regions, compared with 46%–59% for FLUS, indicating stronger spatial agreement. The 2050 projections in this dataset were compared with Africapolis and low-resolution urban expansion datasets under the SSP2 scenario. Note that the dataset from Chen et al. (2020) was generated using FLUS, the model for accuracy comparison. Three comparison figures are provided: 1-Binary-Comparison.png compares binary urban expansion maps across multiple periods, 2-Aggregated-Comparison.png compares the STUrban results after aggregation to 1 km resolution with urban-fraction datasets, and 3-Africapolis-Comparison.png highlights differences between STUrban and Africapolis. These comparisons show that the STUrban projections provide finer spatial detail than all other datasets. Compared with Africapolis, STUrban avoids overpredicting urban land in some regions, which may result from the relatively loose urban definition used in Africapolis. Compared with binary low-resolution datasets, STUrban reduces the widespread omission of small and medium-sized urban areas. Compared with urban-fraction datasets, STUrban results provide a more reliable scenario of urban expansion and better capture the dynamics of small and medium-sized cities. Reference Chen, G., Li, X., Liu, X., Chen, Y., Liang, X., Leng, J., Xu, X., Liao, W., Qiu, Y., Wu, Q., & Huang, K. (2020). Global projections of future urban land expansion under shared socioeconomic pathways. Nature Communications, 11(1), 537. Gao, J., & O’Neill, B. C. (2020). Mapping global urban land for the 21st century with data-driven simulations and Shared Socioeconomic Pathways. Nature Communications, 11(1), 2302. He, C., Liu, Z., Wu, J., Pan, X., Fang, Z., Li, J., & Bryan, B. A. (2021). Future global urban water scarcity and potential solutions. Nature Communications, 12(1), 4667. He, W., Li, X., Zhou, Y., Shi, Z., Yu, G., Hu, T., Wang, Y., Huang, J., Bai, T., & Sun, Z. (2023). Global urban fractional changes at a 1 km resolution throughout 2100 under eight scenarios of shared socioeconomic pathways (SSPs) and representative concentration pathways (RCPs). Earth System Science Data, 15(8), 3623–3639. Heinrigs, P. (2020). Africapolis: Understanding the dynamics of urbanization in Africa. Field Actions Science Reports. The Journal of Field Actions, (Special Issue 22), 18–23. Zhou, Z., & Chen, Y. (2025). STUrban: A novel spatial-temporal deep learning model to simulate long-term urban growth. Information Geography, 1(1), 100004.