Companion Data for "Mapping the Intelligence Gap: GPU-Accelerated Machine Learning and Hyperparameter Optimization for the Temporal Analysis of Dutch Colonial Cartography of the Isthmus of Panama, 1609 to 1714"

Companion data deposit for the article “Mapping the Intelligence Gap,” submitted to the Journal of Data Mining and Digital Humanities (JDMDH, submission #18266). This deposit contains the corpus inventory, annotation protocol, extraction and training scripts, complete Optuna hyperparameter optimization study, evaluation metrics, trained model weights, and figures for the proof-of-concept fortification-detection analysis reported in Section IV.A of the article. Corpus: 195 individually catalogued early modern maps of the Isthmus of Panama and the Tierra Firme coast (1609–1714), assembled from six repositories (David Rumsey Map Collection, Nationaal Archief, British Library, Bibliothèque nationale de France, Archivo General de Indias, National Archives and National Library of Scotland), with 20 metadata fields per entry including cartobibliographic references and verification status. An 87-map subset is flagged for the proof-of-concept analysis. Model: ResNet-50 fine-tuned on 20,182 manually annotated 256×256 pixel patches (9,214 positive, 10,968 negative) drawn from a 392-image pool. Hyperparameter optimization ran 30 Optuna trials (19 completed, 11 pruned) on an NVIDIA RTX A4000 in approximately ten hours. The best configuration (trial 13) achieved F1 = 1.000 on a held-out test set of 2,018 patches. Note on the perfect score: the train/test split was performed at the patch level rather than the map level, allowing patches from the same map to appear in both sets. The model therefore likely exploits map-level stylistic signatures rather than fortification-specific features. This is discussed transparently in the article; a map-level split is planned for the dissertation phase. See README. Map images are not redistributed; they remain available from the source repositories identified in the corpus inventory under their respective access terms.