IARA: An Underwater Acoustic Database

Recent advancements in artificial intelligence have been driven by the availability of diverse and well-structured public databases; however public databases for passive sonar classification remain scarce. To address this gap, this paper presents a new public underwater acoustic labeled database collected during the Santos Basin Soundscape Monitoring Project carried out by PETROBRAS in compliance with conditions required by the federal environmental licensing process conducted by IBAMA. This database expands public resources, containing 129 h 34 min of audio across 1825 recordings (approximately 50 GB). In contrast, existing public databases include ShipsEar (3 h 9 min, 90 recordings) and DeepShip (47 h 4 min, 613 recordings). The dataset provides several hours of general-purpose recordings; however, in this work, only classification was explored, as this is a typical use of such databases in the literature. Baselines were established using Random Forest, MLP, and CNN classification models, with the MLP achieving the best performance: a balanced accuracy of $67.48~\pm ~1.24$ % and an F1-score of $66.72~\pm ~1.17$ %, evaluated using $5\times 2$ cross-validation. The primary contribution of this work is the introduction of the database and its associated baselines for underwater acoustic target recognition tasks, with the aim of providing a valuable resource for the research community. The study also emphasizes the importance of proper data selection, including the use of exclusion regions to enhance signal confidence. Generalization experiments highlight the critical role of dataset diversity, demonstrating that models often struggle to generalize to other datasets. These findings underscore the necessity of diverse, well-structured databases to improve model robustness across varied acoustics scenarios. The IARA an Acoustical Recordings Archive (IARA) is available at doi.org/10.5281/zenodo.15777429 and the code used in this paper is available at github.com/labsonar/IARA