Geospatial decision support for prefabricated construction screening: A reproducible comparison of explicit decision rules, weighted MCDA and machine learning

Australia faces a persistent housing-supply shortfall, increasing interest in prefabricated construction as an alternative delivery approach. However, its practical application depends on site-specific terrain, accessibility and ground conditions that are often represented by incomplete regional evidence during early planning. Translating such evidence into automated screening therefore requires not only predictive models, but explicit consideration of how engineering decision knowledge and missing information are represented. This exploratory study compares three computational treatments of screening knowledge for coastal New South Wales: a deterministic conjunctive rule, weighted multicriteria decision analysis (MCDA), and nonlinear learned approximations using LightGBM and a multilayer perceptron (MLP). Elevation, categorised roads and land-and-soil capability (LSC) mapping were integrated at 21,217 points on a 100 m grid. The proxy rule requires slope below 10°, road distance below 1,000 m and a composite ground score below 0.45. On 4,244 held-out sites, MCDA achieved F1 0.723 and Average Precision (AP) 0.791; LightGBM achieved F1 0.995 and AP 1.000; and the MLP achieved F1 0.981 and AP 0.999. These scores quantify fidelity to encoded screening knowledge rather than engineering suitability. Critically, 84.98% of sites lack valid ordinary-class values in the selected LSC fields, and imputation causes these sites to automatically satisfy the ground condition, including 58.29% of proxy-positive sites. The study shows that model disagreement reflects different decision structures and that missing-evidence treatment can become an implicit decision rule. Independent construction outcomes remain necessary before operational validation.