Comparing Feature-Driven and Data-Driven Models for GNSS-Based ZTD Reconstruction

Zenith Total Delay (ZTD) derived from Global Navigation Satellite System (GNSS) observations is a critical parameter for precise positioning and atmospheric studies. ZTD is continuously estimated at GNSS stations, forming high-resolution temporal time series that reflect the dynamic behavior of the troposphere. In recent years, deep learning approaches have been increasingly applied to ZTD estimation due to their ability to model nonlinear relationships. However, particularly for time-series reconstruction problems, it remains an open question whether the added architectural complexity of such models is always necessary.

In this study, feature-driven classical machine learning models and a data-driven neural network approach are systematically compared for reconstructing missing ZTD values in GNSS time series. The analysis is based on ZTD observations from six International GNSS Service (IGS) stations covering the period from February 2023 to January 2024. All models are trained using an identical feature set comprising lagged ZTD values, tropospheric gradients, ZTD variances, station coordinates, and temporal attributes. This design ensures a fair and interpretable comparison between different modeling approaches.

Linear Regression is considered as a baseline model, while Random Forest represents a nonlinear yet interpretable machine learning approach, and a Fully Connected Neural Network (FNN) is employed as a deep learning model. Model performance is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²), with a leave-one-station-out validation strategy applied to assess generalization capability.

The results indicate that the Random Forest model achieves accuracy comparable to that of the FNN, while exhibiting greater stability and consistency across stations. The results highlight that incorporating physically meaningful features into the input space can be as effective as increasing model complexity for ZTD reconstruction. The study provides methodological and practical insights for selecting appropriate modeling strategies in tropospheric delay estimation.