On the improvement of earth orientation parameters predictions using deep learning methods

• Autores: Sonia Guessoum
• Directores de la Tesis: Santiago Belda Palazón (dir. tes.) , José Manuel Ferrándiz Leal (dir. tes.)
• Lectura: En la Universitat d'Alacant / Universidad de Alicante ( España ) en 2025
• Idioma: inglés
• Número de páginas: 251
• Tribunal Calificador de la Tesis: Jolanta Nastula (presid.) , Alberto Escapa García (secret.) , Alexander Kehm (voc.)
• Enlaces
  • Tesis en acceso abierto en: RUA
• Resumen

  • The rotation of the Earth is governed by complex and irregular behavior influenced by a range of geophysical processes. Understanding these dynamics is essential, as Earth’s orientation in space plays a critical role in numerous applications, including satellite orbit determination, climate modeling, and natural disaster mitigation. This thesis explores the temporal evolution of Earth’s rotation through the analysis and prediction of Earth Orientation Parameters (EOP), introducing advanced methodologies to enhance predictive accuracy. In recent decades, climate change has significantly altered Earth’s ecosystems and atmospheric behavior, posing new challenges for societal safety and infrastructure. Accurate forecasting of EOP is increasingly important in this context. This research critically examines the agreement between theoretical models and observational EOP data, with the goal of improving the reliability and precision of prediction methods. To advance beyond traditional analytical approaches, this study adopts modern machine learning (ML) techniques—particularly deep learning (DL)—to model and forecast EOP. With access to large-scale datasets from geodetic observation systems such as Very Long Baseline Interferometry (VLBI), Global Navigation Satellite Systems (GNSS), Gravity Recovery and Climate Experiment (GRACE), and Satellite Laser Ranging (SLR), deep learning methods offer a powerful framework for uncovering hidden patterns and nonlinear dependencies that conventional models often fail to capture. The proposed approach involves training neural networks on historical EOP records and auxiliary geophysical variables—including Atmospheric Angular Momentum (AAM), Oceanic Angular Momentum (OAM), and Hydrological Angular Momentum (HAM)—to learn the complex relationships that govern Earth’s rotational dynamics. By leveraging DL’s capacity to process high-dimensional and temporally correlated data, the study aims to significantly enhance prediction accuracy. However, the integration of deep learning into geodetic prediction models presents several practical challenges. Data quality is a primary concern: geodetic time series often contain noise, discontinuities, and gaps stemming from sensor limitations, measurement errors, or preprocessing inconsistencies. These factors can adversely affect model training and performance. Moreover, the computational cost associated with training deep neural architectures—such as CNNs, LSTMs, and Transformers—can be substantial. These models often demand powerful hardware, large labeled datasets, and considerable processing time, potentially limiting their applicability in real-time or resource-constrained environments. Another critical issue is model interpretability. Unlike physical models with transparent mechanisms, deep learning models tend to behave as “black boxes,” making it difficult to trace predictions back to meaningful geophysical phenomena. This limits the scientific insight that can be drawn from model outputs and poses challenges in validation and trustworthiness. To address these issues, the study emphasizes rigorous data preprocessing, including normalization, outlier handling, and time-series alignment. Additionally, explainability techniques such as SHAP values and attention mechanisms are explored to improve the interpretability of the models and ensure consistency with physical understanding. In conclusion, this thesis presents a novel framework for predicting Earth Orientation Parameters by integrating data-driven methodologies with domain- specific knowledge. The research not only improves prediction performance but also contributes to a deeper understanding of Earth’s rotational behavior. While the study leverages DL techniques to advance EOP prediction, future work could benefit from incorporating Physics-Informed Neural Networks (PINNs), which embed physical laws—such as those governing Earth’s rotation—into the learning process. This integration may enhance physical consistency, generalization, and robustness, particularly in data-sparse or noisy scenarios. These findings have significant implications across fields such as space geodesy, environmental monitoring, and Earth system science, laying the groundwork for future interdisciplinary collaboration and innovation.