Session assets
Big data is one of the most important phenomena of the 21st century, creating unique opportunities and challenges in its processing and interpretation. Machine learning (ML), a subset of artificial intelligence (AI), has become a foundation of data science, which enables applications ranging from computer vision, geoscience, aviation and medicine. ML becomes important when establishing mathematical models that connect explanatory variables to predicted variables is impossible due to complexity. Deep learning (DL), a subset of ML, has revolutionized fields such as speech recognition, email filtering, and time series analysis. However, DL methods face challenges such as high data demand, overfitting, and the “black box” problem.
We review least-squares-based deep learning (LSBDL), a framework that combines the interpretability of linear least squares (LS) theory with the flexibility and power of deep learning (DL). LS theory, widely used in engineering and geosciences, provides powerful tools for parameter estimation, quality control, and reliability through linear models. DL, on the other hand, deals with modelling complex nonlinear relationships where the mapping between explanatory and predicted variables is unknown. LSBDL bridges these approaches by formulating DL within the LS framework: training networks to establish a design matrix, an essential element of linear models. Through this integration, LSBDL enhances DL with transparency, statistical inference, and reliability. Gradient descent methods such as steepest descent and Gauss-Newton methods are used to construct an adaptive design matrix. By combining the transparency of LS theory with the data-driven adaptability of DL, LSBDL addresses challenges in different fields including geoscience, aviation, and data science. This approach not only improves the interpretability of DL models, but also extends the applicability of LS theory to nonlinear and complex systems, offering new opportunities for innovation and research.
By embedding statistical foundations in the DL workflow, LSBDL offers a three-fold advantage: i) Direct computation of covariance matrices for predicted outcomes allows for quantitative assessment of model uncertainty. ii) Well-established theories of hypothesis testing and outlier detection facilitate the identification of model misspecifications and outlying data, and iii) The covariance matrix of observations can be used to train networks with statistically correlated, inconsistent, or heterogeneous datasets. Incorporating least squares principles increases model explainability, a critical aspect of interpretable and explainable artificial intelligence, and bridges the gap between traditional statistical methods and modern DL techniques. For example, LSBDL can incorporate prior knowledge using soft and hard physics-based constraints, a technique known as physics-informed machine learning (PIML).
The approach is illustrated through three illustrative examples: Surface fitting, time series forecasting, and groundwater storage downscaling. Beyond these examples, LSBDL offers opportunities for various applications including geoscience, inverse problems, aviation, data assimilation, sensor fusion, and time series analysis.
How to cite: Amiri-Simkooei, A.: Theory and implementation of least-squares-based deep learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9878, https://doi.org/10.5194/egusphere-egu25-9878, 2025.
Signal separation is a general problem in many geodetic datasets representing the superposition of various sources. We investigate global, temporal gravity data such as measured by the satellite missions Gravity Recovery and Climate Experiment (GRACE) and GRACE-Follow on (GRACE-FO). As gravity is an integral quantity, these data include signal components related to all kinds of geophysical processes involving a redistribution of masses in the Earth’s system. Examples for the latter are water mass redistribution processes such as seasonal hydrological variations or extreme events, as e.g. floods and droughts, but also Earthquakes or mass changes of ice sheets.
For optimally exploiting temporal gravity data regarding geophysical downstream applications, algorithms splitting up the data into the contained sub-signals are required. We attempt solving this signal separation task by training a neural network-based algorithm to recognize the individual signal components based on their typical patterns in space and time. Thereby, prior knowledge on the spatio-temporal behavior of the individual signals is introduced via forward-modeled time-variable gravity data for each of the components, as well as additional constraints.
Our algorithm is based on a multi-channel U-Net architecture which takes the sum of several signals as input and gives the retrieved individual sub-components as output. For the supervised training and subsequent testing of our software, we use a closed-loop simulation environment, working with the time-variable gravity signals given by the Updated ESA Earth System Model. The latter includes separate datasets for temporal gravity signals caused by mass change processes in the atmosphere and oceans (AO), the continental hydrosphere (H), the cryosphere (I) and the solid Earth domain (S).
For converting the global, temporal gravity data depending on the axes latitude, longitude and time to a 2-d image data format fitting the input and output layers of the U-Net, we split the data along one of its three axes to obtain latitude-longitude, latitude-time or time-longitude samples.
In a test example, we investigate the task of separating the above-mentioned AO, H, I and S components from their sum. The resulting relative RMS test errors being between 19% and 67% demonstrate that our network successfully separates the four considered signals from their sum at signal-to-noise ratios larger than 1.
In our contribution, we describe the functionalities of our software and possibilities to adapt it to any task of interest, including methods for introducing additional physical knowledge on the behavior of specific signals. In general, the described framework is applicable for signal separation in any dataset that depends on three axes (e.g., two spatial and one temporal, or three spatial axes). For the real data application of the framework, we suggest to use representative forward-modeled signals for training, and to subsequently test the trained separation model on real observational data.
How to cite: Heller-Kaikov, B., Pail, R., and Werner, M.: Neural network-based framework for gravity signal separation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1481, https://doi.org/10.5194/egusphere-egu25-1481, 2025.
Since 2002, the GRACE and GRACE Follow-On (GRACE/-FO) missions monitored total water storage changes (TWSC) – the amount of water stored on continents – with an unprecedented accuracy. However, there is a latency of a few months in generating the standard GRACE TWSC product while operational services that aim at such as forecasting drought or flood potentials require near-real time or even forecasted total water storage maps. In this study, we use machine learning to forecast the global GRACE-like total water storage change at long-lead times (up to one year ahead), where our approach works with lagged/forward-moving hydrometeorological variables/indices (e.g., precipitation) as predictors. We evaluate these data-driven TWSC forecasted from a hindcast experiment over 2010-2024 using GRACE observations and compare them to the existing model forecast products from the European Centre for Medium-Range Weather Forecasts (ECMWF)'s new long-range forecasting system (SEAS5). We find that the presented approach performs overall better than the land surface model (HTESSEL) as used by the ECMWF SEAS5 in forecasting TWSC. We forecast global TWSC at a grid resolution of 1° using three GRACE mascon solutions, resulting in three forecasted GRACE-like TWSC datasets. Possible applications include large-scale drought and flood early warning, constraining and downscaling water storage in hydrological forecasting models, the sea level forecasting, interpreting total water storage changes during the latency period of GRACE/-FO data, and geodetic applications like forecasting Earth orientation parameters via hydrological angular momentum excitation or forecasting loading corrections in GNSS and altimetry data analysis.
How to cite: Li, F. and Kusche, J.: Global long-lead forecast of total water storage and evaluation for 2010-2024, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-19735, https://doi.org/10.5194/egusphere-egu25-19735, 2025.
Accurately monitoring and predicting the large-scale dynamic changes of water levels in coastal zones is essential for its protection, restoration and sustainable development. However, there has been a challenge for achieving this goal using a single radar altimeter and retracking technique due to the diversity and complexity of coastal waveforms. To solve this issue, we proposed an approach of estimating water level of the coastal zone in Beibu Gulf, China, by combination of waveform classifications and multiple sub-waveform retrackers. This paper stacked Random Forest (RF), XGBoost and CatBoost algorithms for building an ensemble learning (SEL) model to classify coastal waveforms, and further evaluated the performance of three retracking strategies in refining waveforms using Cryosat-2, SARAL, Sentinel-3 altimeters. We compared the estimation accuracy of the coastal water levels between the single altimeter and synergistic multi-altimeter, and combined Breaks for Additive Season and Trend (BFAST), Mann-Kendall mutation test (MK) with Long Short-Term Memory (LSTM) algorithms to track the historical change process of coastal water levels, and predict its future development trend. This paper found that: (1) The SEL algorithm achieved high-precision classification of different coastal waveforms with an average accuracy of 0.959, which outperformed three single machine learning algorithms. (2) Combination of Threshold Retracker and ALES+ Retracker (TR_ALES+) achieved the better retracking quality with an improvement of correlation coefficient (R, 0.089~0.475) and root mean square error (RMSE, 0.008∼ 0.029 m) when comparing to the Threshold Retracker & Primary Peak COG Retracker and Threshold Retracker & Primary Peak Threshold Retracker. (3) The coastal water levels of Cryosat-2, SARAL, Sentinel-3 and multi-altimeter were in good agreement (R>0.66, RMSE<0.135m) with Copernicus Climate Change Service (C3S) water level. (4) The coastal water levels of the Beibu Gulf displayed a slowly rising trend from 2011 to 2021 with an average annual growth rate of 8mm/a, its lowest water level focused on May-August, the peak of water level was in October-November, and the average annual growth rate of water level from 2022-2031 was about 0.6mm/a. These results can provide guidance for scientific monitoring and sustainable management of coastal zones.
How to cite: Qin, J., Li, Z., and Jiang, W.: Synergistic multi-altimeter for estimating water level in the coastal zone of Beibu Gulf using SEL, ALES + and BFAST algorithms, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15082, https://doi.org/10.5194/egusphere-egu25-15082, 2025.
Abstract: The Earth is subject to complex internal and external forces in space. External forces include the gravitational attraction from the sun, moon, and other planets, and the internal forces include mass loads and frictional forces from the atmosphere, oceans, ice, snow, water, as well as interactions between the crust and mantle. Earth rotation parameters (ERPs) are essential for transforming between the celestial and terrestrial reference frames, and for high-precision space navigation and positioning. Amongst the ERPs, Polar Motion (PM) is a critical parameter for analyzing and understanding the dynamic interaction between the solid Earth, atmosphere, ocean, and other geophysical fluids. To investigate the impact of effective angular momentum (EAM) on long-term ERPs prediction, this thesis conducts research on long-term ERPs prediction considering EAM. Taking into account the influence of EAM, a discrete Liouville equation related to polar motion and UT1-UTC was first established, and the corresponding geodetic angular momentum was obtained. Finally, the residual geodetic angular momentum was obtained and modeled. Taking into account the residual of geodetic angular momentum and the experimental results of EAM, it is shown that, compared with Bulletin A, the LS+LSTM model has improved the accuracy of PMX, PMY, and UT1-UTC in the mid-and long term.
Keywords: Earth rotation parameters, Polar Motion, UT1-UTC, Least squares, Long Short-Term Memory model, effective angular momentum
Funding: Part of this work is supported by the National Natural Science Foundation of China (NSFC) (Grant No.42174037, No. 42030105, No. 42204006, No. 42274011, No. 42304095) and the State Key Laboratory of Geo-Information Engineering and Key Laboratory of Surveying and Mapping Science and Geospatial Information Technology of MNR, CASM (Grant No. 2024-01-01), the China Postdoctoral Science Foundation under Grant No.2024M752480.
How to cite: Wang, C., Wang, J., Zhang, C., Zhang, P., and Sang, J.: High-precision Earth Rotation Parameters Prediction with Physical Excitation Factors and Deep Learning Methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4895, https://doi.org/10.5194/egusphere-egu25-4895, 2025.
Simulation studies are an essential tool in space geodesy. They support the optimization of ground and space networks, the exploration of innovative concepts, and the advancement of technological developments. Beyond measurement noise, realistic simulations must incorporate various error sources, including atmospheric effects, clock drift, and technology-related issues. For many space geodetic techniques, accurate simulations of tropospheric effects, particularly incorporating spatio-temporal correlations, are essential in this context.
Traditionally, high-quality tropospheric simulations popular in the Very Long Baseline Interferometry (VLBI) technology rely on Kolmogorov turbulence theory combined with the frozen flow assumption. These simulations are parameterized using the refractive index structure constant (Cn), alongside auxiliary parameters like wind velocity and troposphere height. Although Cn values can be derived from Global Navigation Satellite System (GNSS) observations, most existing studies assume a generalized average troposphere, largely independent of specific locations or seasonal variations. Only a few incorporate location-based conditions, often through a simplistic latitude-based interpolation. Furthermore, reliance on GNSS data limits the ability to test potential network extensions when such observations are not available.
This work enhances tropospheric simulations by introducing a global, three-dimensional (latitude, longitude, time) Cn model. The model is trained using zenith wet delay (ZWD) estimates from 21,000 globally distributed GNSS stations (2000–2023) and leverages meteorological data from the ERA5 reanalysis, including specific humidity across 11 pressure levels (1000 to 300 hPa) and wind velocity as features. Utilizing the XGBoost algorithm, the model supports short-term prediction scenarios with HRES weather forecasts and provides model uncertainty through an ensemble strategy. The proposed model is able to effectively capture the spatio-temporal patterns in the input data and provides high accuracy, allowing for enhanced simulations of space geodetic observations operating at radio frequencies such as VLBI and GNSS. Additionally, global monthly average estimates on a 0.25° x 0.25° latitude, longitude grid can be derived, offering a practical solution with sufficient accuracy for most simulation studies.
How to cite: Schartner, M.: A global refractive index structure constant model for enhanced tropospheric simulations of space geodetic observations , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6936, https://doi.org/10.5194/egusphere-egu25-6936, 2025.
Water vapor, the most significant greenhouse gas, plays a critical role in the Earth's climate system, influencing the hydrological cycle, energy distribution, and atmospheric dynamics. Integrated Water Vapor (IWV) serves as a vital parameter for understanding these processes. While traditional methods, including ground-based instruments and satellite observations, provide IWV measurements, they are often limited by spatial resolution, coverage, and accuracy. Advances in numerical weather models (NWM) and remote sensing have improved large-scale IWV estimations but still face challenges in capturing high accuracy. To address these limitations, this study introduces a novel approach using a Gaussian Mixed Long Short-Term Memory (GM-LSTM) deep learning model to generate high-resolution water vapor fields (WVF) with enhanced spatial and temporal resolution. The GM-LSTM integrates numerical weather models (NWM) and GNSS data to create an adaptive mapping for zenith wet delay (ZWD) estimation, which is then converted to IWV. By utilizing a bidirectional Long Short-Term Memory (Bi-LSTM) architecture and probabilistic density distribution sequences, the model not only improves ZWD estimation accuracy but also quantifies inherent uncertainties due to spatial heterogeneity. Compared with ERA5 and VMF3, the proposed GM-LSTM achieves an average RMSE reduction of 67.68% and 48.74%. The proposed WVF generation method was validated through an inter-comparison with MODIS and Fengyun satellite products, highlighting its superior accuracy and reliability. This study demonstrates the potential of deep learning models like GM-LSTM to overcome the limitations of traditional techniques, providing a transformative tool for high-resolution IWV estimation and supporting advancements in climate monitoring and weather prediction.
Keywords:
WVF, Inter-comparison, IWV, GM-LSTM, GNSS, MODIS, Fengyun satellite
How to cite: Wang, L., Wang, D., and Kutterer, H.: High-Resolution Water Vapor Field Generation Using the Gaussian Mixed Long Short-Term Memory Network: A Satellite-Based Inter-Comparison in Germany, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-6309, https://doi.org/10.5194/egusphere-egu25-6309, 2025.
Global Navigation Satellite System (GNSS) Ionospheric Seismology explores the ionospheric response to earthquakes and tsunamis, which generate Traveling Ionospheric Disturbances (TIDs) detectable through GNSS-derived Total Electron Content (TEC) observations. Real-time TID identification can offer a transformative approach to tsunami detection, enhancing tsunami early warning systems (TEWS) by extending coverage to open-ocean regions where traditional buoy-based systems are limited. Scalable and automated TID detection is therefore essential for augmenting TEWS capabilities.
In this work, we present a novel deep learning framework for real-time TID detection [3]. Leveraging Gramian Angular Difference Fields (GADFs), we transform TEC time-series data, retrieved via the VARION algorithm [1, 2], into images. Images were categorized based on ground truth: those overlapping labeled TID ranges were classified as TIDs, while others represented normal ionospheric TEC. This approach offers multiple advantages: (1) it preserves temporal dependencies through the sequential structure of GADFs, (2) allows reconstruction of original time-series data via its bijective nature, and (3) highlights temporal correlations within the data. Furthermore, GADFs encode temporal information directly in the images, making the method robust to missing data and suitable for image-based deep learning models in anomaly detection. Additionally, GADFs produce visually interpretable differences across classes, enhancing their utility.
We evaluated our framework using data from four tsunamigenic earthquakes in the Pacific Ocean: the 2010 Maule earthquake, the 2011 Tohoku earthquake, the 2012 Haida Gwaii earthquake, and the 2015 Illapel earthquake. The first three events were used for model training, while out-of-sample validation was performed on the Illapel earthquake to assess real-world applicability.
A single ResNet (50-layer) model was trained for TID detection, incorporating both anomalous (TID-containing) and normal data to ensure exposure to diverse scenarios. TEC data streams were processed chronologically in 60-minute windows, generating GADF images that the model classified as anomalous or normal. Predicted anomalies were concatenated into sequences and compared against ground truth. To further enhance performance, we integrated a false positive mitigation strategy, based on the likelihood of a TID at each time step, significantly reducing false positives. The model achieved an F1 score of 91.7% and a recall of 84.6%, demonstrating its strong potential for operational use in real-time applications.
By embedding deep learning into real-time GNSS-TEC analysis, this research represents a significant advancement in the use of TEC data for TEWS and underscores the potential of deep learning in geodetic time series analysis.
References
[1] Savastano, G. et al. (2017). “Real-time detection of tsunami ionospheric disturbances with a stand-alone GNSS receiver: A preliminary feasibility demonstration”. Scientific reports, 7(1), 46607.
[2] Ravanelli, M., et al. "GNSS total variometric approach: first demonstration of a tool for real-time tsunami genesis estimation." Scientific reports 11.1 (2021): 3114.
[3] Ravanelli, M. et al. "Exploring AI progress in GNSS remote sensing: A deep learning based framework for real-time detection of earthquake and tsunami induced ionospheric perturbations." Radio Science 59.9 (2024): 1-18.
How to cite: Ravanelli, M., Constantinou, V., Liu, H., and Bortnik, J.: Harnessing AI in GNSS remote sensing: A Deep Learning framework for Real-Time Detection of Ionospheric Perturbations triggered by earthquakes and tsunamis , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16106, https://doi.org/10.5194/egusphere-egu25-16106, 2025.
Radio Occultation (RO) using signals from Global Navigation Satellite Systems (GNSS) is one of the most promising remote sensing techniques for global atmospheric sounding. RO is a limb-sounding technique that uses GNSS signals, refracted during their propagation through the Earth’s atmosphere to a receiver on a low-Earth orbit (LEO) satellite. Over the last decades, RO products have been extensively used for data assimilation in Numerical Weather Prediction (NWP) as well as in climate science.
The RO retrieval of atmospheric profiles is based on accurately measuring phase deviations, which are induced by atmospheric bending of the signal. Over the past two decades, several improvements of the retrieval process have been achieved, but significant challenges remain, including the dependency of certain retrieval steps on external information or the assumption of spherical symmetry.
On the other hand, several RO missions such as the FORMOSAT-3/Constellation Observing System for Meteorology, Ionosphere and Climate (COSMIC) and its successor COSMIC-2 have been initiated over the last two decades. In addition, several commercial companies have launched their own RO payloads, which led to an enormous increase in data amounts in recent years. These large data amounts make it suitable for the application of machine learning (ML) models, which have not been used much by the RO community until now. Only few studies have tested the suitability of ML for replacing classic retrieval algorithms and despite already achieving promising results, they were not able to uncover the full potential of ML, mostly due to the small amounts of data used.
This study presents an initial assessment of the performance of ML-based RO retrievals of temperature, pressure and humidity, trained using RO data from e.g. COSMIC-2 and state-of-the-art reanalysis products such as ERA5. It explores the suitability of various experimental setups and evaluates the sensitivity of the results to different feature setups.
How to cite: Aichinger-Rosenberger, M.: Machine learning-based retrieval of thermodynamic profiles from GNSS-RO observations: Preliminary results , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11342, https://doi.org/10.5194/egusphere-egu25-11342, 2025.
Vegetation water content (VWC) is a critical parameter for understanding Earth’s ecological and hydrological systems, especially as climate change accelerates and extreme events become more frequent. Existing remote sensing methods for monitoring VWC face limitations due to restricted spatiotemporal coverage, soil moisture interference, and poor cloud penetration capability. To address these challenges, this study explores the synergy between unprecedentedly large datasets enabled by GNSS Reflectometry (GNSS-R) constellations and advanced deep learning algorithms for VWC estimation.
We propose a triple-collocated CGS dataset that integrates measurements from Cyclone GNSS (CYGNSS), Global Land Data Assimilation System (GLDAS), and Soil Moisture Active Passive (SMAP). Several deep learning models are benchmarked and evaluated over a three-year timespan. Validation results against ground truth measurements demonstrate robust performance with a minimum root mean square deviation (RMSD) of 1.099 kg/m2. Moreover, predictive uncertainty is quantified using the Monte Carlo dropout method, providing a trustworthy representation for timely applications where ground truth data are unavailable.
Our study highlights the potential of combining GNSS-R with deep learning to address vegetation monitoring gaps. By leveraging the proposed CGS scheme and its large-scale dataset, we aim to catalyze further algorithmic advancements in GNSS-R-based vegetation monitoring and enhance the utility of GNSS-R for environmental applications.
How to cite: Zhao, D., Asgarimehr, M., Heidler, K., Wickert, J., Zhu, X. X., and Mou, L.: Advancing Global Vegetation Water Content Estimation with GNSS-R and Deep Learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-11214, https://doi.org/10.5194/egusphere-egu25-11214, 2025.
Small bodies such as asteroids, comets and moons are primary targets for space exploration. To test guidance, navigation and control algorithms an accurate model of the target’s gravity field is crucial for mission success. However, creating an accurate model near the body’s surface is a challenging task due to the often highly irregular shape and the limited information about the internal mass distribution.
This study presents a Convolutional Neural Network (CNN) to determine the density distribution from accelerations at the surface using gravity inversion. We used our voxel-based mascon simulation environment VMC to generate over 100k realistic density distributions (labels) for a cube and calculated the corresponding acceleration (features) for a fixed grid of positions. We selected this simplistic toy problem due to the perfect shape reconstruction and optimal data representation for the deep learning architectures. To investigate the effect of the shape mismatch between a voxel-reconstructed object and the real object we trained an additional Neural Network (NN) to extract the mass distribution for a triaxial ellipsoid using a similar amount of data.
We used a train/test ratio of 80/20 and trained both models using multiple hyperparameter sets for a maximum of 200 epochs using the Adam optimizer. After convergence, all models conserve the total mass. The best-performing architectures are able to determine the general trend of the mass distribution. For the ellipsoid, it can be observed that the model’s prediction is strongly influenced by the contribution of the body’s shape to the gravitational field.
Our results show that NNs are a promising candidate to extract the density distribution using gravity inversion.
How to cite: Haser, B., Andert, T., and Förstner, R.: Gravity Inversion using Convolutional Neural Networks, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-1723, https://doi.org/10.5194/egusphere-egu25-1723, 2025.
Recent advancements in Global Navigation Satellite System Reflectometry (GNSS-R) have led to significant progress in retrieving soil moisture (SM) and other land parameters such as Above Ground Biomass (AGB). The 2016 launch of NASA’s Cyclone GNSS (CYGNSS) mission provided high spatiotemporal resolution GNSS-R data, enabling more accurate soil moisture estimation. This is a key factor influencing the dielectric constant of scattering surfaces. Recent studies have demonstrated the effectiveness of Artificial Neural Networks (ANN) and Deep Learning (DL) models, including convolutional neural networks (CNNs), in addressing the non-linear complexities of soil moisture retrieval [1], [2]. In this study, CyGNSSnet, which was originally designed for global ocean wind speed estimation [3], is being adapted and optimized for global soil moisture estimation.
The research utilizes CyGNSS Level 1 version 3.2 data, specifically Delay Doppler Maps (DDMs), as the primary input. Each reflection point includes additional parameters like measurement geometry and reflectivity. Auxiliary datasets, including topography, soil texture, vegetation indices, and climate-related variables, are incorporated alongside SMAP soil moisture data as the target variable. These datasets, covering August 2018 to July 2021, are matched with CyGNSS specular points using the nearest neighbor method. Data division for training, testing, and validation follows a year-based approach. The CyGNSSnet architecture includes three main components: a map feature extractor using CNNs, an ancillary feature extractor, and a target regressor. Hyperparameters are fine-tuned to achieve optimal performance, with training conducted on HAICORE servers using PyTorch Lightning and an early-stop scheme to minimize training time.
To evaluate the model’s performance across diverse climates and land covers, the global map is stratified by intersecting land cover data from the Climate Change Initiative (CCI) with temperature regimes from the FAO's Global Agro-Ecological Zones (GAEZ v4). This stratification ensures a comprehensive assessment of CyGNSSnet's soil moisture estimation capabilities under varying environmental conditions. The study highlights the potential of advanced DL models like CyGNSSnet to address complex geospatial challenges, enabling more accurate and efficient global soil moisture retrieval.
References:
[1] M. M. Nabi, V. Senyurek, A. C. Gurbuz, and M. Kurum, “Deep Learning-Based Soil Moisture Retrieval in CONUS Using CYGNSS Delay-Doppler Maps,” 2022, doi: 10.1109/JSTARS.2022.3196658.
[2] T. M. Roberts, I. Colwell, C. Chew, S. Lowe, and R. Shah, “A Deep-Learning Approach to Soil Moisture Estimation with GNSS-R,” 2022, doi: 10.3390/RS14143299.
[3] M. Asgarimehr, C. Arnold, T. Weigel, C. Ruf, and J. Wickert, “GNSS reflectometry global ocean wind speed using deep learning: Development and assessment of CyGNSSnet,” Remote Sens Environ, vol. 269, p. 112801, Feb. 2022, doi: 10.1016/J.RSE.2021.112801.
How to cite: Izadgoshasb, H., Xiao, T., Zhao, D., Pierdicca, N., Wickert, J., and Asgarimehr, M.: Adapting and Evaluating CyGNSSnet: A Deep Learning Approach to estimate Global Soil Moisture using GNSS Reflectometry, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-3607, https://doi.org/10.5194/egusphere-egu25-3607, 2025.
Slow slip events are typically associated with seismic activity, but the internal interactions and their relationship with earthquakes are still not well understood. With the rapid increase in Global Navigation Satellite System (GNSS) data, it has become possible to study the subtle slow slip signals in GNSS displacement time series using deep learning. In this study, GNSS displacement time series data from the Cascadia subduction zone are preprocessed using the Variational Bayesian Independent Component Analysis method to effectively remove non-slow-slip signals. Additionally, we developed a deep learning model that includes a multi-layer bidirectional Long Short-Term Memory neural network and an attention mechanism, which can effectively detect slow slip events from complex data. Through this deep learning model, we successfully detected 56 slow slip events in the Cascadia region from 2012 to 2022. The start times, durations, spatial distribution, and propagation patterns of these 56 events were consistent with earthquake catalogs, providing new insights into the slow slip behavior of the Cascadia subduction zone. Overall, our work offers an effective framework for extracting subtle signals hidden in GNSS time series.
How to cite: Wang, J. and Chen, K.: Detecting slow slip events in the Cascadia subduction zone from GNSS time series using deep learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4976, https://doi.org/10.5194/egusphere-egu25-4976, 2025.
Detecting and analyzing spatiotemporal features of surface deformation signals caused by anthropogenic activities remains a challenging task in areas facing multi-hazard risks (e.g., earthquakes, subsidence, sea level rise, and flooding), particularly in coastal regions. We use Global Navigation Satellite System (GNSS) displacement time-series and apply deep learning procedures to identify and characterize ground deformation patterns resulting from natural subsidence and human activities. The focus area is Northern Italy, specifically the North Adriatic coasts, a region with many gas and oil production and storage operations. Unlike production sites, where hydrocarbons are extracted continuously throughout the year, storage sites follow a seasonal cycle: gas/oil is injected in summer and extracted in winter. Our goals are to (1) identify spatial and temporal ground deformation patterns in GNSS time series linked to these anthropogenic activities, and (2) estimate key reservoir properties such as volume, depth, and spatial extent. We generate synthetic training datasets using 114 GNSS stations and simulated reservoirs modeled with the Mogi model, varying depth and volume-change characteristics over time. Weighted Principal Component Analysis (WPCA) is employed to handle missing GNSS data by assigning zero weights to gaps. We will discuss results relative to the application of Convolutional Neural Networks, AutoEncoders, and Graph Neural Networks. After training and calibrating these models on synthetic GNSS datasets, we apply them to real-world GNSS observations. A comparison will be carried out, discussing pros and cons of the various techniques.
How to cite: Vu, D. T., Gualandi, A., Pintori, F., Serpelloni, E., and Pezzo, G.: Deep Learning for detecting anthropogenic ground deformation signals from GNSS time series along the North Adriatic coasts of Italy, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7477, https://doi.org/10.5194/egusphere-egu25-7477, 2025.
Geodetic data analysis has traditionally relied on geophysical models and statistical methods to quantify Earth's deformation, correct for atmospheric effects, and refine measurement uncertainties. However, with the increasing volume and complexity of geodetic observations, machine learning (ML) may offer a better alternative for modelling non-linear motions of geodetic sites by capturing environmental effects and providing corrections for unmodelled influences.
ML has applications across various domains of geodesy, including coordinate time series analysis, geophysical deformation modelling, atmospheric and hydrological loading corrections, prediction of Earth orientation parameters, and tropospheric delay modelling. This study explores the use of ML techniques in geodetic data processing, focusing on modelling station height variations due to non-tidal loading (NTL) and other unmodelled effects in Very Long Baseline Interferometry (VLBI) data analysis using meteorological and land surface state variables.
Different ML approaches, including ensemble methods and neural networks, are examined to understand how well they can model displacement of geodetic sites due to meteorological and land surface state variables responsible for the redistribution of geophysical fluids on Earth. The study aims to compare these methods, highlighting their strengths and limitations in geodetic applications. By providing a broad perspective on ML integration in geodesy, this work contributes to the ongoing discussion on data-driven approaches for improving geodetic modelling and analysis.
How to cite: Singh, S., Böhm, J., Krásná, H., Böhm, S., Balasubramanian, N., and Dikshit, O.: Leveraging Machine Learning for Advanced Geodetic Data Analysis, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-7827, https://doi.org/10.5194/egusphere-egu25-7827, 2025.
Regional sea level prediction plays a vital role in understanding the impacts of climate change and guiding the design of coastal infrastructure. Sea level rise is mainly driven by two primary factors: barystatic sea level change, caused by the melting of ice sheets, glaciers, and run-off of terrestrial water, and by steric sea level change, resulting from the expansion of seawater due to temperature and salinity changes. The former can be monitored from the Gravity Recovery and Climate Experiment (GRACE) and GRACE Follow-On missions, and the latter is commonly calculated based on ocean salinity and temperature models. In this study, satellite altimetry was used to observe relative sea level changes spanning from May 2002 to April 2023. Specifically, barystatic sea level changes were derived using Mass Concentration (Mascon) solutions, while the steric height was estimated through the Ocean Physics Reanalysis model. According to the sea level budget equation, the total sea level change aligns closely with the combined contributions of barystatic and steric sea level components, validating the consistency of the data and methodology.
Machine learning has become increasingly significant in climate research in recent years. It enables the analysis of large and complex datasets that exceed human processing capabilities. Among the machine learning techniques, the Long Short-Term Memory (LSTM) model is particularly effective for the time series prediction due to its ability to capture long-term dependencies and patterns through its gated mechanisms. LSTM models excel at handling trends, seasonality, and noise in data, making them ideal for understanding the temporal dynamics of sea level changes and predicting future values.
In this research, we applied a CNN-LSTM model to predict total, barystatic, and steric sea level changes. The model leverages the feature extraction capabilities of convolutional neural networks (CNNs) combined with the sequential learning strengths of LSTM. The results of this study provide valuable insights into the contributions of mass and steric components to regional sea level changes. By predicting these signals, this research enhances our understanding of the mechanisms driving the sea level rise, offering the critical information for climate change mitigation and coastal adaptation planning.
How to cite: qiu, F., Gruber, T., and Pail, R.: Prediction of regional sea level change and its components using machine learning methods, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-8792, https://doi.org/10.5194/egusphere-egu25-8792, 2025.
Sea level forecasting is crucial for safeguarding coastal regions, enhancing marine infrastructure, ensuring navigational safety, and mitigating natural hazards. Traditional methods, which often rely on single data sources like tide gauges or merged altimeters, may struggle to provide accurate predictions in highly dynamic regions due to their limited spatial and temporal resolution. Multi-Sensor Data Fusion has the potential to address these limitations by integrating data from multiple sensors, thereby improving the consistency, quality, and accuracy of predictions. While this approach has been successfully applied to other oceanographic parameters such as sea surface temperature (SST), sea ice, and salinity, sea level prediction remains challenging due to the spatial constraints of tide gauges and the relatively low temporal frequency of satellite observations. New satellite altimetry missions, such as the Surface Water and Ocean Topography (SWOT) mission, can deliver high-resolution data. When combined with deep learning (DL) techniques, this has the potential to improve the accuracy of sea level forecasting.
This research presents a novel DL approach for sea surface height (SSH) forecasting based on SWOT satellite data. By incorporating data from various sources (e.g. tide gauges, hydrodynamic models, and meteorological variables, such as wind speed and atmospheric pressure), the model aims to provide accurate, instantaneous SSH maps referenced to the geoid surface (i.e. Dynamic Topography (DT)). The DL architecture also encompasses attention mechanisms to prioritize spatially and temporally critical information, potentially overcoming the inefficiencies seen in traditional data assimilation methods.
The goal is to generate DT maps for real-time, adaptable predictions by integrating spatio-temporal data sources. Unlike previous methods, this approach incorporates instantaneous DT directly into the modeling process rather than solely for validation. The approach will be evaluated in the Baltic Sea to assess its precision. This model has the potential to set a new benchmark for sea level forecasting, offering valuable contributions to environmental monitoring, climate research, operational oceanography, and decision-making by providing high-resolution and interpretable predictions.
How to cite: Rajabi-Kiasari, S., Delpeche-Ellmann, N., and Ellmann, A.: Integrating Satellite Altimetry and Deep Learning for Enhanced Sea Level Forecasting , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-9951, https://doi.org/10.5194/egusphere-egu25-9951, 2025.
Global ionospheric maps (GIMs) are widely used ionospheric products, especially in Global Navigation Satellite System (GNSS) applications. They allow for instance to correct for the ionospheric delays in single-frequency applications. The input data for generating GIMs is typically obtained from dual-frequency observations at globally distributed GNSS stations. However, the distribution of these stations is in-homogeneous and predominantly concentrated in continental regions, resulting in large data gaps over the oceanic regions. The absence of data reduces the accuracy of GIMs in these regions. There exist other space techniques, such as satellite altimetry and GNSS radio occultation (GNSS-RO) that can retrieve the state of the ionosphere over ocean areas and have the potential to address such data gaps. However, the vertical total electron content (VTEC) observations from GNSS, satellite altimetry, and GNSS-RO differ due to variations in orbital altitudes and instrumental biases. Another challenge is the sparsity of observations from satellite altimetry and GNSS-RO, particularly when only considering data from a single day.
In this study, we developed a framework that integrates GNSS, Jason-3 satellite altimetry, and COSMIC-2 GNSS-RO observations in a GIM, based on a neural network (NN). First, we calibrated the satellite altimetry and GNSS-RO VTEC to be more consistent with GNSS VTEC. To address data sparsity, we used VTEC observations from satellite altimetry and GNSS-RO for the entire year 2023 to build a background ionospheric model with XGBoost. This background model captures the general climatological characteristics of the ionosphere over the oceans. We then utilized the background model to generate VTEC samples for training the NN-based GIM in regions lacking GNSS station observations. For our three test regions (Hawaii, Southern Atlantic, Antarctic), we find relative improvements in MAE of 37%, 11%, and 37% over the year 2023 compared to GNSS-only GIMs
The results demonstrate that the proposed data fusion method can effectively improve the modeling accuracy in regions with missing data.
How to cite: Iten, M., Mao, S., and Soja, B.: Ionospheric data fusion with GNSS, GNSS-RO and satellite altimetry based on machine learning, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-5640, https://doi.org/10.5194/egusphere-egu25-5640, 2025.
As a guiding mechanism for protecting cultivated land, achieving intensive and economical land use, and improving the ecological environment, the potential analysis of comprehensive land consolidation in the whole area relies on the accuracy of land use classification. In karst regions, such as the Li River Basin, the unique topography and geomorphology often render traditional land-cover classification methods prone to uncertainties, thereby affecting the consistency of conclusions in potential analysis for comprehensive land consolidation in the whole region. This study developed an improved random forest classification algorithm optimized for land-cover classification in the Li River Basin. Based on this algorithm, a potential evaluation model of comprehensive land consolidation in the whole area was constructed, and the spatial distribution characteristics of areas requiring consolidation under different potential levels for various land use types in the Li River Basin were analyzed in depth. The results indicate that the improved random forest algorithm performs best in land use classification in the Li River Basin, with overall accuracy and the kappa coefficient enhanced by 3% and 4%, respectively, compared with the standard random forest algorithm. The spatial distribution of potential levels of comprehensive land consolidation in the whole area across various land use types in the Li River Basin shows significant differences, ranking as ecological land > agricultural land > construction land. Compared to traditional evaluation models, the new model identifies consolidation areas for various land use types, covering 94.97% of the total area in the Li River Basin. This study can facilitate a comprehensive consolidation of land use types in karst regions from all elements, dimensions, and periods, thus provides scientific basis for land use and ecological protection in such areas.
How to cite: Tang, Q., Jiang, W., and Li, Z.: Potential Analysis of the Comprehensive Land Consolidation for Entire Karst Region based on Improved Random Forest Method: A Case Study in the Li River Basin , EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15130, https://doi.org/10.5194/egusphere-egu25-15130, 2025.
The detection and characterization of ground deformations play a critical role in understanding and mitigating the risks associated with natural hazards such as earthquakes, volcanic activity, and land subsidence. These deformations can have significant impacts on human life, infrastructure, and the environment, making their timely and accurate detection essential for disaster management and planning. Remote sensing technologies, particularly those that offer global coverage and high temporal resolution, are indispensable in capturing ground motion over large areas.
Synthetic Aperture Radar (SAR) data, particularly from the Sentinel-1 mission, has revolutionized geodesy and remote sensing by providing high-resolution and frequent observations of Earth's surface. Interferometric SAR (InSAR) techniques allow for precise measurement of surface displacements at millimeter-scale accuracy, enabling the detection of subtle ground deformation patterns. Despite the availability of massive datasets from Sentinel-1, most of this data remains unlabeled, limiting the ability to directly apply supervised machine learning techniques for deformation classification. The need to manually label vast amounts of data is both time-consuming and resource-intensive, leaving a significant portion of the data underutilized for scientific discovery and practical applications.
Representation learning offers a promising approach to address these challenges by extracting meaningful features from large, unlabeled InSAR datasets. Self-supervised learning methods can leverage contrastive learning techniques to pretrain neural network encoders on deformation data, capturing patterns and structures inherent in the data without requiring labels. These learned representations can then be fine-tuned for downstream tasks, such as classifying deformation types (e.g., magma movements during volcanic eruptions or ground deformations coming from earthquakes) or detecting anomalies. By bridging the gap between vast, unlabeled data and the need for precise classification, representation learning enables more efficient use of InSAR datasets, advancing our ability to monitor and understand Earth's dynamic processes.
How to cite: Kaselimi, M. and Makantasis, K.: Learning InSAR Deformation Representations, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-12006, https://doi.org/10.5194/egusphere-egu25-12006, 2025.
Accurate short-term prediction of Celestial Pole Offsets (CPO), essential for applications such as satellite navigation and space geodesy, remains a significant challenge. We addressed this task using a 1D Convolutional Neural Network (1D CNN), a machine learning model designed to capture temporal patterns effectively. Our objective was to enhance the prediction accuracy of the key CPO components, dX and dY. To improve model interpretability, we employed SHapley Additive exPlanations (SHAP), which identifies the most influential input features, such as historical dX and dY values and data from models like the Free Core Nutation (FCN) model. This transparency is critical for scientific and operational contexts.
We evaluated our model using various input data types, including rapid Earth Orientation Parameters (EOPs), Bulletin A data from the International Earth Rotation and Reference Systems Service (IERS), and FCN-derived data. Our results demonstrated improved prediction accuracy across all data sources, with rapid EOPs emerging as the most effective for short-term forecasts, particularly for the initial day. Leveraging rapid EOPs, we simulated the conditions of the 2nd Earth Orientation Prediction Comparison Campaign (EOPPCC) to further test our approach. The model outperformed other machine learning methods used in the campaign for dX predictions, although dY proved more challenging due to its complex dynamics.
This study highlights the potential of 1D CNNs in advancing CPO forecasting, particularly when coupled with interpretable frameworks like SHAP and diverse, high-quality data sources. Our findings underscore the transformative role of deep learning in enhancing the precision and reliability of Earth Orientation Parameter predictions, thereby supporting critical scientific and operational applications.
How to cite: Guessoum, S., Belda, S., Ferrándiz, J. M., Modiri, S., and Karbon, M.: DL for CPO feature selection and forecasting, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-16061, https://doi.org/10.5194/egusphere-egu25-16061, 2025.
Posters virtual: Thu, 1 May, 14:00–15:45 | vPoster spot 1
EGU25-20320 | ECS | Posters virtual | VPS23
Deep learning in RTM gravity field modeling: A case study over Wudalianchi areaThu, 01 May, 14:00–15:45 (CEST) vPoster spot 1 | vP1.8
The Residual Terrain Modeling (RTM) technique is commonly used to recover short-wavelength gravity field signals. However, classical gravity forward modeling methods for RTM gravity field determination face challenges such as series divergence, inefficient computation, and errors induced by tree canopy in Digital Elevation Models (DEMs). In this study, deep learning methods are employed to enhance the quality of the computed RTM gravity field. Experiments are conducted at the Wudalianchi airborne gravity gradiometer test site, which provides a large volume of precise gravity measurements. The Random Forest method is used to estimate and correct tree canopy height errors in DEMs. A fully connected deep neural network (FC-DNN) is introduced to efficiently calculate the RTM gravity field. Additionally, to improve the network’s generalization capability, a novel terrain information fusion regularization method is applied to create an Improved FC-DNN with a refined loss function. The accuracy, computational efficiency, and generalization performance of the deep learning method are evaluated and compared in the Wudalianchi volcanic region. The results demonstrate a significant improvement in the accuracy of the RTM gravity field when based on tree canopy-corrected DEMs. The RTM gravity fields determined using both FC-DNN and Improved FC-DNN achieve mGal-level accuracy, with a remarkable 10,000-fold increase in computational efficiency compared to the classical Newtonian integration method. The Improved FC-DNN exhibits superior generalization, with accuracy enhancements ranging from 7% to 21% compared to the standard FC-DNN.
How to cite: Yang, M., Zhang, B., Wang, L., Feng, W., and Zhong, M.: Deep learning in RTM gravity field modeling: A case study over Wudalianchi area, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-20320, https://doi.org/10.5194/egusphere-egu25-20320, 2025.
EGU25-4633 | ECS | Posters virtual | VPS23
Deep Neural Networks for GNSS Coordinate Time Series Modeling and PredictionThu, 01 May, 14:00–15:45 (CEST) | vP1.9
High-precision GNSS coordinate time series modeling and prediction provide a critical reference for applications such as crustal deformation, structural safety monitoring, and regional or global reference frame maintenance. A Deep neural network framework based on Transformer was applied to 22 GNSS stations, each with 1000 days, in which data is preprocessed using a synchronization sliding window. The overall fitting and prediction trends exhibit a high degree of consistency with the original time series. The average fitting RMSE and MAE are 3.40 mm and 2.64 mm, respectively, while the corresponding average prediction RMSE and MAE are 3.54 mm and 2.77 mm. In comparison to the LSTM model, the proposed method achieved a redu78ction in RMSE and MAE by 20.7% and 19.6%, respectively. Furthermore, when benchmarked against the traditional least squares approach, the improvements were even more pronounced, with RMSE and MAE decreasing by 35.7% and 37.8%, respectively. The approach demonstrates robustness and effectiveness under conditions of discontinuous data. Therefore, it could be used as a convenient alternative to predict GNSS coordinate time series and will be of wide practical value in the fields of reference frame maintenance and deformation early warning.
How to cite: Wang, J., Li, Z., and Jiang, W.: Deep Neural Networks for GNSS Coordinate Time Series Modeling and Prediction, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-4633, https://doi.org/10.5194/egusphere-egu25-4633, 2025.
EGU25-15605 | ECS | Posters virtual | VPS23
Deep Learning Approaches for Zenith Total Delay EstimationThu, 01 May, 14:00–15:45 (CEST) | vP1.10
Zenith Total Delay (ZTD) is a crucial parameter for understanding the effects of atmospheric conditions on satellite signals, constituting a fundamental aspect of precision positioning and atmospheric modeling applications. Traditional methods for ZTD estimation, including GNSS observations, numerical weather prediction models, and interpolation techniques, encounter critical limitations such as generalization constraints, sparse data availability, insufficient spatial coverage, high computational costs, and limited adaptability to dynamic atmospheric changes. Deep learning techniques provide substantial benefits, including processing large and complex datasets, enabling dynamic modeling, and delivering rapid and accurate estimations.
This study integrates real-time GNSS observations with high-resolution atmospheric reanalysis data from the ERA5 dataset to develop deep learning-based methods for ZTD estimation. GNSS data were sourced from 17 IGS tropospheric stations strategically selected to represent diverse geographic and climatic conditions. These stations supplied ZTD values and their temporal variations at 5-minute intervals, spanning February 2023 to January 2024. ERA5 data, offering hourly atmospheric parameters, necessitated the alignment of GNSS temporal resolution with ERA5 for spatial modeling. The spatial distribution of GNSS data was optimized using interpolation techniques to enhance the quality of inputs for deep-learning models.
The findings highlight the potential of deep learning techniques to enhance ZTD estimation processes. Future research will focus on integrating additional datasets, such as InSAR, to achieve higher spatial resolution and improved accuracy. Moreover, advanced deep learning architectures, including attention mechanisms, will be investigated to refine estimation methods and broaden their applications in atmospheric and geospatial studies.
How to cite: Tekin Ünlütürk, N. and Bak, M.: Deep Learning Approaches for Zenith Total Delay Estimation, EGU General Assembly 2025, Vienna, Austria, 27 Apr–2 May 2025, EGU25-15605, https://doi.org/10.5194/egusphere-egu25-15605, 2025.
