Journal of Intelligent Computing Maleerat Sodanil 17 4 2026 https://doi.org/10.6025/jic/2026/17/4/223-245 https://www.dline.info/jic/fulltext/v17n4/jicv17n4_3.pdf Automated construction monitoring increasingly relies on geospatial metadata and machine learning to track urban development, yet the true generalizability and reliability of metadata driven models remain underexplored. This study evaluates the predictive capabilities of geospatial metadata including polygon geometry, temporal duration, and annotation attributes using Random Forest, Extra Trees, and Logistic Regression models on a large scale GeoJSON dataset of construction sites. We conducted binary (site presence) and multiclass (construction type) classification tasks, revealing a stark contrast: while Random Forest achieved near perfect binary accuracy (99.9%), multiclass construction type recognition remained highly ambiguous (macro-F1 = 0.610). Crucially, permutation feature importance analysis exposed that the binary predictive signal was almost entirely concentrated in a single annotation derived variable (n_features), with all physical and geometric features registering zero importance. This indicates that the near perfect performance is an artifact of the annotation workflow and potential target leakage rather than genuine physical site detection. We conclude that metadata only models are insufficient for fine grained semantic typing and issue a critical methodological warning for the GeoAI community: rigorous spatial cross validation and annotation leakage audits are mandatory to prevent spatially and procedurally inflated performance metrics.