Ensuring the reliability of national antenna infrastructure records requires data-driven methods capable of handling the heterogeneity and regulatory complexity of the United States Federal Communications Commission (FCC) and Federal Aviation Administration (FAA) reporting ecosystem. The publicly available Antenna Structure Registration (ASR) dataset provides extensive geospatial, administrative, and regulatory information but introduces substantial analytical challenges, including mixed data types, coordinate-format inconsistencies, spatial redundancy, and strong geographic autocorrelation that can bias conventional learning evaluations. Rather than replacing authoritative coordinates, this work frames geospatial prediction as a validation-oriented proxy task for assessing internal consistency, spatial signal preservation, and robustness under strict leakage controls. To address these challenges, this study introduces an optimization-driven learning framework that integrates metaheuristic search with transformer-based modeling for large-scale registry data. A Dynamic Leader Sibha Algorithm (Siba)–driven pipeline is employed for both feature selection (FS) and hyperparameter optimization (HPO), where binary Siba identifies compact predictive attribute subsets and real-coded Siba tunes the architectural and training parameters of a Variable Adaptive Sequential Transformer (VAST) model. The novelty lies in the disciplined two-stage integration of wrapper-based FS and HPO under a geographically stratified, leakage-aware evaluation protocol, rather than in introducing new optimization or transformer architectures. Rigorous benchmarking is conducted against ten established metaheuristic optimizers using identical training, initialization, and data-partitioning conditions. Baseline results indicate that VAST achieves the strongest initial performance among the evaluated learning models. Following Siba-based feature selection, consistent error reductions are observed across geographically disjoint test regions, demonstrating that substantial attribute compression can be achieved without degrading spatial generalization. Subsequent real-coded Siba-based HPO yields further improvements, producing an optimized configuration with statistically stable gains across repeated runs and evaluation metrics. The framework is intended for registry data quality assessment, anomaly screening, and methodological evaluation of geospatial learning pipelines under realistic spatial constraints, rather than for operational coordinate replacement. Beyond ASR, the approach provides a transferable experimental and optimization foundation for scalable modeling of other large, structured geospatial infrastructure repositories.