Abstract
Coastal urbanization requires spatial models that distinguish historical change from conditional projections. This study reconstructed land-use and land-cover dynamics in Makassar, Indonesia, from 2005 to 2025 and projected built-up expansion to 2035 using an integrated artificial neural network–cellular automata (ANN–CA) framework and Bayesian Weights of Evidence (WoE). Five Landsat-derived maps were classified into seven classes using Random Forest, with reported overall accuracy ranging from 62.83% to 87.61%. Ten demographic, accessibility, economic, topographic, and environmental variables were used to estimate transition potential and interpret factor-class associations. Model performance was evaluated using Kappa and the Figure of Merit. Built-up land increased from 8,199.20 ha in 2005 to 9,880.06 ha in 2025, a net gain of 1,680.86 ha (20.50%), while its share of the study area rose from 46.61% to 56.17%. Expansion decelerated sharply: 93.48% of the observed net increase occurred during 2005–2015. Cropland recorded the largest net decline (861.81 ha), whereas wetlands and water bodies jointly declined by 607.87 ha. Population growth and land value showed the strongest positive summarized contrasts, while the flood-exposure variable showed the strongest negative contrast. Kappa reached 0.768, but the Figure of Merit was 21.01%, indicating stronger whole-map agreement than locational accuracy for changed cells. Built-up land is projected to reach 10,394.23 ha (59.09% of the study area) by 2035, with 93.38% of projected growth occurring after 2030. The results provide a conditional basis for spatial screening of development pressure, but mapping uncertainty, uneven temporal allocation, and limited locational accuracy require cautious planning interpretation.