Karathanasis, Andreas, Violos, John, Varlamis, Iraklis, Leivadeas, Aris et Tserpes, Konstantinos.
2026.
« Leveraging autoencoders for GeoAI: A survey of methods and applications ».
Neurocomputing, vol. 700.
Prévisualisation |
PDF
Leivadeas-A-2026-34063.pdf - Version publiée Licence d'utilisation : Creative Commons CC BY. Télécharger (3MB) | Prévisualisation |
Résumé
The rapid growth of spatio-temporal data from sources such as GPS, remote sensing, and Internet of Things (IoT) devices has established trajectory data as a core component of modern Geospatial Artificial Intelligence (GeoAI). However, the high dimensionality, noise, irregular sampling, and heterogeneous nature of trajectory data pose significant challenges for traditional analytical methods. In recent years, autoencoders have emerged as a powerful representation learning framework for modeling complex trajectory patterns by encoding spatio-temporal data into compact and expressive latent spaces. This survey provides a comprehensive review of autoencoder-based methods for trajectory representation learning within the GeoAI domain. Moving beyond the traditional focus on trajectory compression, we present a unified perspective that encompasses a broad range of tasks, including clustering, anomaly detection, similarity computation, recovery and reconstruction, prediction, classification, and trajectory generation. We systematically categorize existing approaches based on autoencoder architectures such as feed-forward, recurrent, convolutional, variational, transformer-based, graph-based, and hybrid models, and analyze their suitability for different trajectory representations and application scenarios. Furthermore, we examine commonly used evaluation metrics across tasks and highlight how latent representations serve as a unifying abstraction for diverse analytical objectives. The survey also reviews real-world applications in domains such as urban mobility, maritime navigation, aviation safety, and interaction-aware modeling. Finally, we identify critical challenges related to limited labeled data, robustness under noisy, incomplete, degraded, or heterogeneous geospatial observations, privacy and ethical data usage, scalability and deployment constraints, lack of standardized evaluation protocols, limited cross-region and cross-domain generalization, multimodal representation alignment, uncertainty estimation, and the interpretability of latent representations. We outline future research directions toward robust and uncertainty-aware autoencoders, degradation-aware reconstruction, physics-guided and foundation model-enhanced GeoAI, multimodal and cross-domain representation learning, privacy-aware methods, scalable deployment, and semantically meaningful and explainable latent spaces.
| Type de document: | Article publié dans une revue, révisé par les pairs |
|---|---|
| Chercheur(-euse): | Chercheur(-euse) Leivadeas, Aris |
| Affiliation: | Génie logiciel et des technologies de l'information |
| Date de dépôt: | 30 juill. 2026 18:07 |
| Dernière modification: | 26 sept. 2026 18:19 |
| URI: | https://espace2.etsmtl.ca/id/eprint/34063 |
Actions (Authentification requise)
 |
Dernière vérification avant le dépôt |