Document Type : Original Article
Graphical Abstract
Highlights
· A back-propagation multilayer-perceptron ANN trained on eight GIS criterion layers derived directly from a municipal land-use dataset predicted neighborhood facility-location suitability with a coefficient of determination of 0.974 and a root-mean-square error of 0.040 on held-out validation data, demonstrating that a learning-based approach can model the complex, non-linear interactions among urban spatial criteria with high fidelity and minimal over-fitting.
· Connection-weight importance analysis identified residential fabric density (0.156), land-use compatibility (0.151), and distance to existing services (0.142) as the three most influential determinants of suitability, together accounting for approximately 45% of total normalized criterion importance and confirming that the alignment of supply with accessible demand constitutes the primary design objective for neighborhood facility location in heterogeneous metropolitan contexts.
· The ANN–GIS model outperformed a conventional Analytic Hierarchy Process (AHP) weighted linear-combination baseline by a substantial margin, reducing root-mean-square error by approximately 44% (from 0.071 to 0.040) and raising the coefficient of determination from 0.842 to 0.974, while also capturing non-linear criterion interactions and threshold effects that the additive model systematically suppressed.
· Application of the trained network to the full raster of the eastern zone of Tabriz classified only 38.7% of the study area as suitable or highly suitable for new neighborhood-facility development, with recommended sites clustering in identifiable high-suitability corridors in the northern and south-central parts of the zone, providing municipal planners with a transparent, spatially explicit, and defensible priority map for investment allocation.
· The spatial residual diagnostic confirmed the absence of systematic spatial bias across the study area a result that distinguishes the ANN–GIS framework from conventional overlay methods and that, together with its interpretable connection-weight importance ranking, addresses the principal objection that neural models are 'black boxes' unsuitable for transparent public planning practice in Iranian metropolitan contexts.
