Cellular Automata–Deep Learning Modelling of Urban Growth in Windhoek, Namibia

This repository contains a Google Colab Jupyter notebook for cellular automata (CA) and deep learning-based urban growth modelling in Windhoek, Namibia. The notebook supports the geospatial analysis and simulation of urban expansion using a Landsat-derived land use/land cover (LULC) time series spanning 1990–2025. The research compares three deep learning-based transition rules—a convolutional neural network (CA-CNN), an attention-gated U-Net (CA-U-Net), and a bidirectional LSTM-CNN—with conventional cellular automata benchmarks, including logistic-regression CA. Model performance is assessed through a hindcast experiment, with additional simulations exploring conditional urban growth scenarios to 2050. The findings indicate that deep learning models did not outperform the conventional logistic-regression CA benchmark in the evaluated hindcast. Proximity to and density of existing urban land were the strongest predictors of urban expansion, while all models had difficulty reproducing leapfrog and corridor development. The study highlights the importance of reliable reference data, planning information, and urban growth demand assumptions when modelling urban expansion in data-scarce, semi-arid cities. The notebook is shared to promote research transparency, computational reproducibility, and further methodological development in geospatial urban growth modelling. The simulations represent conditional scenarios rather than definitive predictions of future urban expansion. Study area: Windhoek, NamibiaStudy period: 1990–2025Scenario horizon: 2050Methods: Cellular automata, deep learning, logistic regression, Landsat-derived LULC analysis, and urban growth simulationEnvironment: Google Colab and Python