Identifying underutilized land by eXplainable artificial intelligence and geographic similarity ensemble model with limited samples

As cities globally confront the dual challenges of spatial resource scarcity and aging urban fabrics, the precise identification of underutilized land emerges as a critical pathway toward sustainable urban regeneration. However, persistent methodological gaps hinder precise identification due to three unresolved scientific problems: (1) multifactorial spatial complexity obscuring determinant interactions, (2) limited sample availability constraining machine learning efficacy, and (3) opaque decision-making processes in conventional algorithms. This study resolves these through an eXplainable Artificial Intelligence-Geographic Similarity Reasoning (XAI-GSR) model integrating three innovations: a multidimensional indicator system quantifying land-use efficiency across morphology, economic, social, and ecological dimensions; XGBoost-SHAP interpretation elucidating nonlinear factor contributions; and geospatial analogical reasoning overcoming sample scarcity. Applied to Shenzhen, the model achieved 82.9 % accuracy, identifying 9668 underutilized blocks (25.44 % total) with distinct typological distribution - Type 1 (6.99 %) reflecting central district efficiency versus Type 2 dominance (56.17 %) revealing suburban improvement potential, while Type 3 (27.18 %) and Mixed-type (9.67 %) clusters predominantly occupy eastern/northern low-density zones. Compared to existing methods, our framework advances underutilized land detection by simultaneously resolving sample limitations through geospatial similarity reasoning and enhancing reliability via uncertainty-quantified similarity metrics, providing urban planners with an empirically validated decision-support tool for targeted regeneration strategies.