From Classical Urban Growth Models to Data-Driven Methods: Predicting Urban Expansion and Built-Up Intensity in Accra

Understanding historical and future patterns of urbanization is essential for anticipating demographic change, guiding sustainable development, and managing climate and hazard risks. Although Shared Socioeconomic Pathways (SSPs) incorporate urbanization conceptually, few settlement projections have been adapted to fine spatial scales that capture intra-urban heterogeneity. Because most growth in developing-country cities occurs via horizontal expansion at the peri-urban fringe, improving spatially explicit models of land conversion and build-up dynamics remains a key methodological need.

We evaluate alternative open and reproducible modeling approaches for the Accra metropolitan area (Ghana), a rapidly growing and spatially uneven urban region. Using satellite-derived land use/land cover and built-up layers (2001, 2005, 2009), we compare (i) established urban growth modeling (UGM) toolchains focused on binary expansion (MOLUSCE, FUTURES, SLEUTH) against a flexible statistical learning baseline (LEARN), and (ii) newer approaches that model the continuous built-up surface directly. For the expansion-focused setup, models are trained on 2001–2005 and evaluated by predicting 2009 transitions using a shared covariate set (e.g., prior urban extent/LULC, distance to roads and waterways, protected areas, elevation, and distance to existing development).

Performance is assessed using the Figure of Merit (FoM), a change-focused accuracy measure that avoids inflated scores under rare-change conditions typical of urban expansion. In the expansion-only comparison, the statistical learning framework LEARN provides the strongest baseline performance (FoM ≈ 0.20), exceeding MOLUSCE (0.07), FUTURES (0.01), and SLEUTH (0.10).

We then extend the task from binary land conversion to predicting the continuous build-up surface. A random forest baseline that models built-up change directly achieves FoM ≈ 0.50 in Accra. Building on this, we implement a two-head U-Net that jointly estimates (i) the likelihood of expansion and (ii) the magnitude of build-up increase, with constraints to keep predicted change non-negative and spatially plausible. This neural approach performs best overall (FoM ≈ 0.65), improving substantially on both classical UGM baselines and the random-forest model.

Overall, results indicate that modeling build-up as a continuous surface—and explicitly coupling expansion with magnitude via neural networks—can markedly improve change-prediction skill in fast-growing cities, while remaining compatible with scenario-consistent urban forecasting frameworks.