Geological hazards caused by both natural forces and human-induced disturbances, such as land subsidence, earthquakes, tectonic motion, mining activities, coastal erosion, volcanic activities, and permafrost alterations, cause great adverse effects to earth’s surface. The preservation of a comprehensive record detailing past, present, and future surface movements is imperative for effective disaster risk mitigation and property protection. Interferometric Synthetic Aperture Radar (InSAR) is widely recognized as a highly effective and extensively employed geodetic technique for comprehending the spatiotemporal evolution of historical ground surface deformation. However, it only reveals the past deformation evolution process and the deformation update is slowly considering the long revisit cycle of satellites. Deformation evolution in the future is also crucial for preventing and mitigating geological hazards. Unlike traditional mathematical-statistical models and physical models, machine learning methods provide a new perspective and possibility to efficiently and automatically mine the time series information over a large-scale area. In the context of InSAR time series prediction over large areas, the previous researches do not consider the spatiotemporal heterogeneity caused by various factors over a large-scale area and mainly focus on one typical deformation point.

Therefore, in this study, we present a framework designed to predict large-scale spatiotemporal InSAR time series by integrating independent component analysis (ICA) and a Long Short-Term Memory (LSTM) machine learning model. This framework is developed with a specific focus on addressing spatiotemporal heterogeneity within the dataset. The utilization of the ICA method is employed to identify and capture the displacement signals of interest within the InSAR data, enabling the characterization of independent time series signals associated with various natural or anthropogenic processes. Additionally, a K-means clustering approach is incorporated to partition the study area into spatiotemporal homogeneity subregions across a large-scale region, aiming to mitigate potential decreases in model accuracy caused by data heterogeneity. Subsequently, LSTM models are constructed for each cluster, and optimal parameters are determined. The proposed framework is rigorously tested using simulated datasets and validated against two real-world cases—land subsidence in the Willcox Basin and post-seismic deformation following the Sarpol-e Zahab earthquake. Comparative analysis demonstrates that the proposed model surpasses the original LSTM, resulting in a 34% and 17% improvement in average prediction accuracy, respectively. The spatial prediction results in 60 days over the two cases are mapped with high accuracy.

This study introduces an integrated framework that seamlessly integrates InSAR data processing with machine learning techniques such as LSTM to enhance our ability to predict deformation over large-scale geographical areas. The adaptability of the proposed model has made it an alternative to numerical or empirical models, especially when detailed on-site data is scarce or challenging to obtain. While our immediate applications have focused on scenarios on land subsidence and post-seismic deformation, the broader implications of our methodology are evident. We anticipate the proposed framework will be expanded to various application domains, including mining, infrastructure stability, and other situations involving sustained motions. The proposed framework will ultimately contribute to more informed decision-making and risk assessment in complex dynamic systems.

How to cite: Peng, M., Motagh, M., Lu, Z., Xia, Z., Guo, Z., Zhao, C., and Quan, Y.: Improving ground deformation prediction in satellite InSAR using ICA-assisted RNN model , EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10467, https://doi.org/10.5194/egusphere-egu24-10467, 2024.

Global Navigation Satellite Systems (GNSS) are best known for their accurate positioning, navigation and timing capabilities. In total, over 20.000 permanent high-grade GNSS stations are available worldwide, the positions of which are monitored with millimeter accuracy. Thanks to the high accuracy and the fact that these stations are mounted on the ground, subtle movements due to hydrological changes and crustal deformation can be observed. Thus, the GNSS observations contain valuable geophysical information. Although many geodetic applications require these movements to be properly understood and potentially corrected, this is not trivial due to the complexity of the interactions within the Earth’s system. Therefore, there is a severe lack of available models explaining residual GNSS station movements beyond conventionally modeled effects. On the opposite, if these movements are properly understood, GNSS observations might contribute to the correct interpretation of emerging environmental changes.

This study exploits the wealth of satellite-derived Earth observation (EO) data to derive suitable models to explain GNSS station movements. We combine GNSS station coordinate time series and EO variables with the help of machine learning techniques to benefit from various types of information. While the target vector consists of concatenated GNSS station coordinate time series over Europe, EO variables such as precipitation, soil water, snow water equivalent, and land cover data are used as input features. Different machine learning models, including Random Forest, XGBoost, and Multilayer Perceptron, are investigated and compared. Additionally, a sensitivity analysis is performed to determine the individual impact of EO variables to quantify what drives GNSS movements, which in turn, might allow monitoring the corresponding Earth system processes based on GNSS observations.

How to cite: Crocetti, L., Schneider, R., and Soja, B.: Explaining GNSS station movements based on Earth observation data, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10706, https://doi.org/10.5194/egusphere-egu24-10706, 2024.

Displacement time series from Global Navigation Satellite System (GNSS) at daily rates are used commonly to investigate and understand the processes controlling Earth's surface deformation originating from tectonic processes such as postseismic slip, slow slip events and viscoelastic relaxation, but also non-tectonic processes such as reflectometry, atmospheric sensing and remote sensing. For each individual research field, different parts of the total recorded GNSS displacement time series are of intrest. A major difficulty is the modeling and isolation of non-tectonic seasonal signals that are established to be related with non-tidal surface loading.

In the past, many methods were developed with some success based on Kalman filters, matrix factorization and various approaches using curve fitting to separate the tectonic and non-tectonic contributions. However, these methods still have some difficulties in  isolating the seasonal loading signals especially in the presence of interannual variations in the seasonal loading pattern and steps in the time series.

We present here a deep learning model trained on a globally distributed, continuous 8-10 years long dataset of ~8000 stations PPP-GNSS displacement time series from NGL to estimate the seasonal loading signals using a global non-tidal surface loading model developed at ESM-GFZ. We compare our model to other statistical methods for isolation of the seasonal with the established method of subtraction of the non-tidal surface loading signals (hydrological loading, and non-tidal atmospheric and oceanic loading) as our baseline. We also present the evaluation of our model and its capabilities in reducing the seasonal loading signal as well as parts of the high-frequency scattering in the original GNSS time series.

How to cite: Çökerim, K., Bedford, J., and Dobslaw, H.: A Globally Trained Deep Learning Model for Estimation of Seasonal Residual Signals in GNSS displacement time series, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10290, https://doi.org/10.5194/egusphere-egu24-10290, 2024.

Accurate ionospheric models are essential for single-frequency high-precision Global Navigation Satellite Systems (GNSS) applications. Global ionospheric maps (GIMs), which depicts the global distribution of vertical total electron content (VTEC), are a widely used ionospheric product provided by the International GNSS Service (IGS). To meet the increasing need for real-time applications, the IGS real-time service (RTS) has been established and offers real-time (RT) GIMs that can be used for real or near-real time applications. However, the accuracy of present RT GIMs is still significantly lower compared with the final GIMs. IGS RT GIMs show an RMSE of 3.5-5.5 TECU compared to the IGS final GIMs. In this study, we focus on enhancing the accuracy of RT GIMs through machine learning (ML) approaches, specifically a classical Convolutional Neural Network (CNN) and a Generative Adversarial Network (GAN). The objective is to bridge the gap between the RT GIMs and the final IGS GIMs. This is achieved by using RT GIMs as input and final GIMs as target. The ML approach is applied to the IGS combined RT GIMs and Universitat Politècnica de Catalunya (UPC) RT GIMs. The performance of the improved RT GIMs is evaluated in comparison to the combined IGS final GIM.

We consider over 11'000 pairs of RT GIMs and final GIMs. Over a comprehensive test period spanning 3.5 months, the proposed approach shows promising results with an enhancement of more than 30% in accuracy of RT GIMs. Especially for regions with high VTEC values, which are most critical, the results show a significant improvement. The results demonstrate the model’s great potential in generating more accurate and refined real-time maps.

The integration of ML techniques proves to be a promising avenue for refining and augmenting the precision of real-time ionospheric maps, thereby addressing critical needs in the realm of space weather monitoring and single-frequency applications.

How to cite: Iten, M., Mao, S., and Soja, B.: Enhanced Real-time Global Ionospheric Maps using Machine Learning, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-3117, https://doi.org/10.5194/egusphere-egu24-3117, 2024.

It is known that natural hazards such as volcanic eruptions, earthquakes, and tsunamis can trigger acoustic and gravity waves (AGWs) that could reach the ionosphere and generate electron density disturbances known as Travelling Ionospheric Disturbances (TIDs). These disturbances can be investigated in terms of variations in the ionospheric total electron content (TEC) measurements, collected by continuously operating ground-based Global Navigation Satellite Systems (GNSS) receivers. The VARION (Variometric Approach for Real-Time Ionosphere Observation) algorithm is a well-known real-time tool for estimating TEC variations. It is based on single-time differences of geometry-free combinations of GNSS carrier-phase measurements.

Artificial Intelligence (AI), particularly in machine learning, offers computational efficiency and data handling, leading to its exploration in ionospheric studies. In this context, the abundance of data allows the exploration of a VARION-based machine learning classification approach to detect TEC perturbation. For this purpose, we used the VARION-TEC variations from the 2015 Illapel earthquake and tsunami, leveraging the distinct ionospheric response triggered by the event.

We employed machine learning algorithms, specifically Random Forest (RF) and XGBoost (XGB), using the VARION-core observations (i.e., dsTEC/dt) as input features. We formulated a binary classification problem using supervised machine learning algorithms and manually labelled the time frames of TEC perturbations as the target variable. We considered two elevation cut-off time series, namely 15° and 25°, to which we applied the classifier. XGBoost with a 15° elevation cut-off dsTEC/dt time series reached the best performance, achieving an F1 score of 0.77, recall of 0.74, and precision of 0.80 on the test data. More in detail, regarding the testing samples, the model accurately classified 183 out of 247 (74.09%) samples of sTEC variations related to the earthquake and tsunami (True Positives, TP). Moreover, 2975 out of 3021 (98.49%) testing samples were correctly classified as containing no sTEC variations caused by an earthquake (True Negatives, TN). However, 64 out of 247 samples (25.91%) were erroneously classified as not containing sTEC variations related to the event (False Negatives, FN), while 46 out of 3021 (1.51%) were wrongly classified as containing sTEC variations related to the earthquake and tsunami (False Positives, FP).

This model showed a 75-second average deviation in predicting perturbation time frames for testing links, equivalent to 5 steps in the 15-second time series intervals. This highlights the algorithm's potential for early detection of ionospheric perturbations from earthquakes and tsunamis, aiding in early warning purposes.

Finally, the model efficiently detects TIDs within 2-3 minutes, showing an impressive computational efficiency, crucial for effective early warning systems. It relies only on the VARION-generated real-time TEC time series (dsTEC/dt), enabling its application in an operational real-time setting using real-time GNSS data.

In conclusion, this work demonstrates high-probability TEC signature detection by machine learning for earthquakes and tsunamis, which can be used to enhance tsunami early warning systems.

How to cite: Fuso, F., Ravanelli, M., Crocetti, L., and Soja, B.: Using Machine Learning for identifying TEC signatures related to earthquakes and tsunamis: the 2015 Illapel event case study, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-1853, https://doi.org/10.5194/egusphere-egu24-1853, 2024.

Spaceborne GNSS Reflectometry (GNSS-R) is a novel remote sensing technique providing accumulating data volume with global coverage and enhanced temporal resolution. The reflected pre-existing L-Band signal of opportunity transmitted by the Global Navigation Satellite System contains information about the reflection surface properties which can be quantified and converted into data products for further studies. To retrieve such information, Artificial intelligence (AI) models are implemented to estimate geophysical parameters based on the GNSS-R observations. With more and more complex deep learning models being proposed and more and more input features being considered, understanding the decision-making process of the models and the contributions of the input features becomes as important as enhancing the model output accuracy.

This study explores the potential of the Explainable AI (XAI) to decode complex deep learning models for ocean surface wind speed estimation trained by the Cyclone GNSS (CYGNSS) observations. The input feature importance is evaluated by applying the SHAP (SHapley Additive exPlanations) Gradient Explainer to the model on an unseen dataset. By analyzing the SHAP value of each input feature, we find that in addition to the two known parameters that are used in the operational wind speed retrieval algorithm, other scientific and technical ancillary parameters, such as the orientation of the satellite and the signal power information are also useful for the model.

We seek to offer a better understanding of the deep learning models for estimating ocean wind speed using GNSS-R data and explore the potential features for more accurate retrieval. In addition to building an efficient model with effective inputs, XAI also helps us to discover the important factors found by models which can enhance the physical understanding of the GNSS-R mechanism.

How to cite: Xiao, T., Asgarimehr, M., Arnold, C., Zhao, D., Mou, L., and Wickert, J.: Explainable AI for GNSS Reflectometry: Investigating Feature Importance for Ocean Wind Speed Estimation, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12715, https://doi.org/10.5194/egusphere-egu24-12715, 2024.

Accurate prediction of Earth orientation parameters (EOPs) is critical for astro-geodynamics, high-precision space navigation, and positioning, and deep space exploration. However, the current models' prediction accuracy for EOPs is significantly lower than that of geodetic technical solutions, which can adversely affect certain high-precision real-time users. In this study, we introduce a simultaneous prediction approach for Polar Motion (PM) and Celestial Pole Offsets (CPO) employing deep neural networks, aiming to deliver precise predictions for both parameters.
The methodology comprises three components, with the first being feature interaction and selection. The process of feature selection within the context of deep learning differs from traditional methods for machine learning, and may not be directly applicable to theme since they are designed to automatically learn relevant features. Consequently, we propose in this step a solution based on feature engineering to select the best set of variables that can keep the model as simple as possible but with enough precision and accuracy using recursive feature elimination and the SHAP value algorithm, aiming to investigate the influence of FCN (Free Core Nutation) with its amplitude and phase on the CPO forecasting. This investigation is crucial since FCN is the main source of variance of the CPO series. Considering the role represented by the effective Angular Momentum functions (EAM), and their direct influence on the Earth's rotation, it is logical to assess numerically the impact of EAM on the Polar motion and FCN excitations. SHAP value aids in comprehending how each feature contributes to final predictions, highlighting the significance of each feature relative to others,  and revealing the model's dependency on feature interactions.
During the second phase, we formulate two deep-learning methods for each parameter. The first Neural Network incorporates all features, while the second focuses on the subset of features identified in the initial step. This stage primarily involves exploring feature and hyperparameter tuning to enhance model performance. The SHAP value algorithm is also used in this stage for interpretation. 
In the final phase, we construct a multi-task deep learning model designed to simultaneously predict ( CPO ) and ( PM ).  This model is built using the optimal set of features and hyperparameters identified in the preceding steps. To validate the methodology, we employ the most recent version of the time series from the International Earth Rotation and Reference Systems Service (IERS), namely IERS 20 C04 and EAM provided by the German Research Center for Geosciences (GFZ). We focus on a forecasting horizon of 90 days, the practical forecasting horizon needed in space-geodetic applications.
In the end, we conclude that the developed model is proficient in simultaneously predicting ( CPO ) and ( PM ). The incorporation of ( EAM ), sheds light on its role in CPO excitations and Polar Motion predictions.

How to cite: Guessoum, S., Belda, S., Ferrándiz, J. M., Begga, A., Karbon, M., Schuh, H., Modiri, S., and Heinkelmann, R.:  Feature Selection and Deep Learning for Simultaneous Forecasting of Celestial Pole Offset (CPO) and Polar Motion (PM), EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-10575, https://doi.org/10.5194/egusphere-egu24-10575, 2024.

One big challenge in the analysis and interpretation of geodetic data is the separation of the individual signal and noise components contained in the data. Specifically, the global, temporal gravity data obtained by the GRACE and GRACE Follow-On satellite missions contain spatial-temporal gravity signals caused by all kinds of mass variations in the Earth’s system. While only the sum of all signals can be measured, for geophysical interpretation, an extraction of individual signal contributions is necessary.

Therefore, our aim is to develop an algorithm solving the signal separation task in global, temporal gravity data. Since the individual signal components are characterized by specific patterns in space and time, the algorithm to be found needs to be able to extract patterns in the 3-dimensional latitude-longitude-time space.

We propose to exploit the pattern recognition abilities of deep neural networks for solving the signal separation task. Our method uses a multi-channel U-Net architecture which is able to translate the sum of various signals as single-channel input to the individual signal components as multi-channel output. The loss function is a weighted sum of the L2 losses of the individual signals.

We perform a supervised training using synthetic data derived from the updated Earth System Model of ESA. The latter consists of separate datasets for temporal gravity variations caused by mass redistribution processes in the atmosphere, the oceans, the continental hydrosphere, the cryosphere and the solid Earth.

In our study, we use different parts of this dataset to form training and test datasets. In this fully-synthetic framework, the ground truth of the individual signal components is also known in the testing stage, allowing a direct computation of the separation errors of the trained separation model.

In our contribution, we present results on optimizing our algorithm by tuning various hyperparameters of the neural network. Moreover, we demonstrate the impact of the number of superimposed signals and the definition of the loss function on the signal separation performance of our algorithm.

How to cite: Heller-Kaikov, B., Pail, R., and Werner, M.: Signal separation in global, temporal gravity data using a multi-channel U-Net, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-12487, https://doi.org/10.5194/egusphere-egu24-12487, 2024.

Snow is a key variable of the global climate system and the hydrological cycle, as well as one of the most critical sources of freshwater. Therefore, measurements of snow-related parameters such as seasonal snow height (SSH) or snow-water-equivalent (SWE) are of great importance for science, economy and society. Traditionally, these parameters are either measured manually or with automated ground-based sensors, which are accurate, but expensive and suffer from low temporal and spatial resolution.

A new alternative for such systems is the use of GNSS observations, by application of the GNSS interferometric reflectometry (GNSS-IR) method. The technique enables users to infer information about soil moisture, snow depth, or vegetation water content. Signal-to-Noise Ratio (SNR) observations collected by GNSS receivers are sensitive to the interference between the direct signal and the reflected signal (often referred to as “multipath”). The interference pattern changes with the elevation angle of the satellite, the signal wavelength, and the height of the GNSS antenna above the reflecting surface. By comparing this reflector heights estimated for snow surfaces with those from bare soil conditions, snow height can be determined.

The estimation of reflector heights, and respectively SSH, is typically carried out using Lomb-Scargle Periodogram (LSP) spectrum analysis. This study investigates the potential of machine learning methods for this task, using similar input parameters as the standard GNSS-IR retrieval. Results from different supervised algorithms such as Random Forest (RF) or Gradient Boosting (GB) are shown for different GNSS sites and experimental setups. First investigations indicate that snow heights can be successfully obtained with machine learning, with results less noisy than with classical approaches.

How to cite: Aichinger-Rosenberger, M. and Soja, B.: Exploring the performance of machine learning models for the GNSS-IR retrieval of seasonal snow height, EGU General Assembly 2024, Vienna, Austria, 14–19 Apr 2024, EGU24-14724, https://doi.org/10.5194/egusphere-egu24-14724, 2024.