G1.4
Data science and machine learning in geodesy

G1.4

EDI
Data science and machine learning in geodesy
Convener: Benedikt Soja | Co-conveners: Kyriakos BalidakisECSECS, Maria KaselimiECSECS, Ryan McGranaghanECSECS, Randa Natras
vPICO presentations
| Thu, 29 Apr, 11:00–11:45 (CEST)

vPICO presentations: Thu, 29 Apr

Chairperson: Benedikt Soja
11:00–11:05
11:05–11:07
|
EGU21-4665
|
ECS
Milad Asgarimehr, Caroline Arnold, Felix Stiehler, Tobias Weigel, Chris Ruf, and Jens Wickert

The Global Navigation Satellite System Reflectometry (GNSS-R) is a novel remote sensing technique exploiting GNSS signals after reflection off the Earth's surface. The capability of spaceborne GNSS-R to monitor ocean state and the surface wind is recently well demonstrated, which offers an unprecedented sampling rate and much robustness during rainfall. The Cyclone GNSS (CyGNSS) is the first spaceborne mission fully dedicated to GNSS-R, launched in December 2016.

Thanks to the low development costs of the GNSS-R satellite missions as well as the capability of tracking multiple reflected signals from numerous GNSS transmitters, the GNSS-R datasets are much bigger compared to those from conventional remote sensing techniques. The CyGNSS provides a high number of unique samples in the order of a few millions monthly.  Deep learning can therefore be implemented in GNSS-R even more efficiently than other remote sensing domains. With the upcoming GNSS-R CubeSats, the data volume is expected to increase in the near future and GNSS-R “Big data” can be a future challenge. Deep learning methods are additionally able to correct the potential effects, both technical and geophysical, dictated by data empirically when the mechanisms are not well described by the theoretical knowledge. This poses the question if GNSS-R should embrace deep learning and can benefit from this modern data scientific method like other Earth Observation domains.

The receivers onboard CyGNSS cross-correlate the reflected signals received at a nadir antenna to a locally generated replica. The cross-correlation power at a range of the signal delay and Doppler frequency shift is the observational output of the receivers being called delay-Doppler Maps (DDMs). The mapped power is inversely proportional to the ocean roughness and consequently surface winds.

Few recent studies innovatively show some merits of machine learning techniques for the derivations of ocean winds from the DDMs. However, the capability of machine learning techniques, especially deep learning for an operational data derivation needs to be better characterized. Normally, the operational retrieval algorithms are developed based on an existing dataset and are supposed to operate on the upcoming measurements. Therefore, machine learning-based models are supposed to generalize well on the unseen data in future periods. Herein, we aim at the characterization of deep learning capabilities for these GNSS-R operational purposes.

In this interdisciplinary study, we present a deep learning algorithm processing the CyGNSS measurements to derive wind speed data. The model is supposed to meet an acceptable level of generalization on the upcoming unseen data, and alternatively can be used as an operational processing algorithm. We propose a deep model based on convolutional and fully connected layers processing the DDMs besides ancillary input features. The model leads to the so-far best quality of global wind speed estimates using GNSS-R measurements with a general root mean square error of 1.3 m/s over unseen data in a time span different from that of the training data.

How to cite: Asgarimehr, M., Arnold, C., Stiehler, F., Weigel, T., Ruf, C., and Wickert, J.: Deep Learning for the derivation of GNSS Reflectometry global ocean wind speed, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-4665, https://doi.org/10.5194/egusphere-egu21-4665, 2021.

11:07–11:09
|
EGU21-10267
|
ECS
|
Maria Kaselimi, Nikolaos Doulamis, and Demitris Delikaraoglou

Total Electron Content (TEC) is the integral of the location-dependent electron density along the signal path and is a crucial parameter that is often used to describe ionospheric variability, as it is strongly affected by solar activity. TEC is highly depended on local time, latitude, longitude, season, solar and geomagnetic conditions. The propagation of the signals from GNSS (Global Navigation Satellite System) throughout the ionosphere is strongly influenced by short- and long-term changes and ionospheric regular or irregular variations. 
Long short-term memory network (LSTM) is a specific recurrent neural network architecture and is capable of learning time dependence in sequential problems and can successfully model ionosphere variability. As LSTM networks “memorize” long term correlations in a sequence, they can model complex sequences with various features, where solar radio flux at 10.7 cm and magnetic activity indices are taken into consideration to provide more accurate results. 
Here, we propose a deep learning architecture to create regional TEC models around a station. The proposed model allows different solar and geomagnetic parameters to be inserted into the model as features. Our model has been evaluated under different solar and geomagnetic conditions. Also, the proposed model is tested for different time periods and seasonal variations and for varying geographic latitudes. 

How to cite: Kaselimi, M., Doulamis, N., and Delikaraoglou, D.: Recurrent Neural Networks for Ionospheric Time Delays Prediction Using Global Navigation Satellite System Observables, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-10267, https://doi.org/10.5194/egusphere-egu21-10267, 2021.

11:09–11:11
|
EGU21-545
|
ECS
|
Mostafa Kiani Shahvandi and Benedikt Soja

Graph neural networks are a newly established category of machine learning algorithms dealing with relational data. They can be used for the analysis of both spatial and/or temporal data. They are capable of modeling how time series of nodes, which are located at different spatial positions, change by the exchange of information between nodes and their neighbors. As a result, time series can be predicted to future epochs.

GNSS networks consist of stations at different locations, each producing time series of geodetic parameters, such as changes in their positions. In order to successfully apply graph neural networks to predict time series from GNSS networks, the physical properties of GNSS time series should be taken into account. Thus, we suggest a new graph neural network algorithm that has both a physical and a mathematical basis. The physical part is based on the fundamental concept of information exchange between nodes and their neighbors. Here, the temporal correlation between the changes of time series of the nodes and their neighbors is considered, which is computed by geophysical loading and/or climatic data. The mathematical part comes from the time series prediction by mathematical models, after the removal of trends and periodic effects using the singular spectrum analysis algorithm. In addition, it plays a role in the computation of the impact of neighboring nodes, based on the spatial correlation computed according to the pair-wise node-neighbor distance. The final prediction is the simple weighted summation of the predicted values of the time series of the node and those of its neighbors, in which weights are the multiplication of the spatial and temporal correlations.

In order to show the efficiency of the proposed algorithm, we considered a global network of more than 18000 GNSS stations and defined the neighbors of each node as stations that are located within the range of 10 km. We performed several different analyses, including the comparison between different machine learning algorithms and statistical methods for the time series prediction part, the impact of the type of data used for the computation of temporal correlation (climatic and/or geophysical loading), and comparison with other state-of-the-art graph neural network algorithms. We demonstrate the superiority of our method to the current graph neural network algorithms when applied to time series of geodetic networks. In addition, we show that the best machine learning algorithm to use within our graph neural network architecture is the multilayer perceptron, which shows an average of 0.34 mm in prediction accuracy. Furthermore, we find that the statistical methods have lower accuracies than machine learning ones, as much as 44 percent.

How to cite: Kiani Shahvandi, M. and Soja, B.: A new spatio-temporal graph neural network method for the analysis of GNSS geodetic data, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-545, https://doi.org/10.5194/egusphere-egu21-545, 2021.

11:11–11:13
|
EGU21-2308
|
ECS
|
Junyang Gou, Mostafa Kiani Shahvandi, Roland Hohensinn, and Benedikt Soja

The Earth Orientation Parameters (EOP) are fundamentals of geodesy, connecting the terrestrial and celestial reference frames. The typical way to generate EOP of highest accuracy is combining different space geodetic techniques. Due to the time demand for processing data and combining different techniques, the combined EOP products often have latencies from several days to several weeks. However, real-time EOP are needed for multiple geodetic and geophysical applications, including precise navigation and operation of satellites. Predictions of EOP in ultra-short time can overcome the problem of latency of EOP products to a certain extent.

In 2010, the Earth Orientation Parameters Prediction Comparison Campaign (EOP PCC) collected predictions from 20 methods, which were mainly based on statistical approaches, and provided a combined solution. In recent years, more hybrid and machine learning methods have been introduced for EOP prediction.

The rapid expansion of computing power and data volume in recent years has made the application of deep learning in geodesy increasingly promising. In particular, the Long Short-Term Memory (LSTM) network, one of the most popular variations of Recurrent Neural Network (RNN), is promising for geodetic time series prediction. Thanks to the special structure of its cells, LSTM network can capture the non-linear structure between different time epochs in the time series. Therefore, it is suitable for EOP prediction problems.

In this study, we investigate the potential of using LSTM for the prediction of Length of Day (LOD). The LOD data from a combination of space geodetic techniques are first preprocessed in order to obtain residuals. For this step, we experiment with the application of Savitzky-Golay filters, Singular Spectrum Analysis and the Gauss Markov model. We then employ LSTM networks of different architectures and its variations such as bidirectional LSTM networks to predict the LOD residuals in ultra-short time. Furthermore, we study the impact of Atmospheric Angular Momentum (AAM) and its forecast data on the predictions. The performance of this method is compared with other results of EOP PCC in a hindcast experiment under the same conditions. In addition, we assess the performance of LOD predictions using longer time series than for the EOP PCC to consider improvements of EOP products over the last decade.

How to cite: Gou, J., Kiani Shahvandi, M., Hohensinn, R., and Soja, B.: Ultra-short-term prediction of LOD using LSTM neural networks, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-2308, https://doi.org/10.5194/egusphere-egu21-2308, 2021.

11:13–11:15
|
EGU21-7178
|
ECS
|
Severin Rhyner, Denis Jordan, Dante Salvini, and Rolf Dach

The Center for Orbit Determination in Europe (CODE) hosts one of the global analysis centers of the International GNSS Service (IGS). Each day, the data of about 250 GNSS stations are processed in a highly automated manner. The processing protocols contain many quality parameters related to station coordinates, satellite orbits, and satellite/receiver clock corrections.

In the context of a reprocessing campaign, 25 years of GNSS measurements are analysed within a short timeframe and therefore a huge number of processing protocols are generated. A manual inspection of all these protocols is highly time consuming. Machine learning (ML) represents a promising approach to provide a data driven and objective evaluation of these protocols. The main objective of a ML framework is to analyse a big number of independent quality parameters in order to automatically detect individual days with problems in the data analysis. Furthermore, we expect that ML could contribute to the detection of unexpected systematics in the solutions and has the potential to improve the GNSS analysis strategy.

As a first step, we have focused on one aspect of the processing protocols, namely the orbit misclosures (discontinuities at the end of the orbital arcs) at midnight. It is known that the orbit modelling of GNSS satellites is more difficult during eclipse seasons. In order to assess the capabilities of different machine learning algorithms for our purpose, we have evaluated the magnitude of the orbit misclosures and have tried to recover the information on whether the satellite was passing the earth shadow or not. State-of-the-art ML algorithms (Random Forest and Decision Tree) showed promising results of up to 80% success rate.

How to cite: Rhyner, S., Jordan, D., Salvini, D., and Dach, R.: Application of Machine Learning for Evaluation of GNSS Processing Protocols, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-7178, https://doi.org/10.5194/egusphere-egu21-7178, 2021.

11:15–11:17
|
EGU21-13011
Galileo data integrity and consistency in the IGS consolidated navigation product
(withdrawn)
Octavian Andrei
11:17–11:19
|
EGU21-3569
|
ECS
Bertrand Rouet-Leduc, Romain Jolivet, Manon Dalaison, Paul Johnson, and Claudia Hulbert

Systematically characterizing slip behaviours on active faults is key to unraveling the physics of tectonic faulting and the interplay between slow and fast earthquakes. Interferometric Synthetic Aperture Radar (InSAR), by enabling measurement of ground deformation at a global scale every few days, may hold the key to those interactions. 
However, atmospheric propagation delays often exceed ground deformation of interest despite state-of-the art processing, and thus InSAR analysis requires expert interpretation and a priori knowledge of fault systems, precluding global investigations of deformation dynamics. 
We show that a deep auto-encoder architecture tailored to untangle ground deformation from noise in InSAR time series autonomously extracts deformation signals, without prior knowledge of a fault's location or slip behaviour.
Applied to InSAR data over the North Anatolian Fault, our method reaches  2 mm detection, revealing a slow earthquake twice as extensive as previously recognized.
We further explore the generalization of our approach to inflation/deflation-induced deformation, applying the same methodology to the geothermal field of Coso, California. 

How to cite: Rouet-Leduc, B., Jolivet, R., Dalaison, M., Johnson, P., and Hulbert, C.: Deep Learning for Autonomous Extraction of Millimeter-scale Deformation in InSAR Time Series, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-3569, https://doi.org/10.5194/egusphere-egu21-3569, 2021.

11:19–11:21
|
EGU21-766
Chen Xingyu, Ran Jiangjun, Xin Linyang, and Yan Zhengwen

Variations of lake areas and shorelines can effectively reflect hydrological and climatic changes. This research focuses on the automatic and simultaneous extraction of lake areas and shorelines from optical remote sensing images and SAR images, and then analyze the area changes of lakes in Tibet Plateau, in order to provide some insights for Plateau wetland environment changes. In our research, we design a novel end-to-end lightweight multitask CNN and a modified deep CNN to automatically extract those. The experimental results over the testing image patches achieve the Accuracy of 0.9962, Precision of 0.9912, Recall of 0.9982, F1-score of 0.9941, and mIoU of 0.9879, which align with or even are better than those of mainstream semantic segmentation models (UNet, DeepLabV3+, etc.). Especially, the in-situ shoreline of the Selinco Lake located in the Central and Southern Tibetan Plateau is also collected by GPS measurements to evaluate the results of the proposed method further and the validation indicates a high accuracy in our results (DRMSE: 30.84 m, DMAE: 22.49 m, DSTD: 21.11 m), with only about one-pixel deviation for Landsat-8 images. On the basis of the preceding verification results, the sequential variations of Tibetan Plateau lakes are captured and reveal Tibetan Plateau lakes generally show an increasing trend. Such as the Selinco Lake which has an expansion trend from 1660 Square kilometers to 2410 Square kilometers, grown by 45% over half a century. It is expected that these conclusions will provide some valuable information on the variations of the Tibetan Plateau wetland environment.

Figure 1. In-situ GPS trajectory (orange line) and the predictive edge pixels (black line) in the comparison region (yellow box).

Figure 2. The variation of Selinco Lake in the Tibetan Plateau over time. The orange line is the area variation we get, other color variations are obtained by other agencies.

Figure 3. Temporal and spatial distribution of shoreline in the eastern and western regions of Selinco Lake.

How to cite: Xingyu, C., Jiangjun, R., Linyang, X., and Zhengwen, Y.: A Novel Multitask CNN for Automatically Extracting Shoreline Variations of Lakes in Qinghai-Tibet Plateau from 1970 to 2020, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-766, https://doi.org/10.5194/egusphere-egu21-766, 2021.

11:21–11:23
|
EGU21-12784
|
ECS
|
Ashutosh Tiwari, Avadh BIhari Narayan, and Onkar Dikshit

Multi-temporal interferometric synthetic aperture radar (MT-InSAR) technique has been effectively used to monitor deformation events over the last two decades. The processing steps generally involve pixel selection, phase unwrapping and displacement estimation. The pixel selection step takes most of the processing time, while a reliable method for phase unwrapping is still not available. This study demonstrates the effect of using deep learning (DL) architectures for MT-InSAR processing. The architectures are applied to reduce time computations and further to improve the quality of pixel selection. Some promising results for pixel selection have been shown earlier with the proposed architecture. In this study, we investigate the performance of the proposed architectures on newer datasets with larger temporal interval. To achieve this objective, the models are retrained with interferometric stacks covering larger temporal period and large time steps (for better estimation of interferometric phase components). Pixel selection results are compared with those obtained using open access algorithms used for MT-InSAR processing.

How to cite: Tiwari, A., Narayan, A. B., and Dikshit, O.: A deep learning approach for efficient multi-temporal interferometric synthetic aperture radar (MT-InSAR) processing, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-12784, https://doi.org/10.5194/egusphere-egu21-12784, 2021.

11:23–11:25
|
EGU21-1070
|
ECS
Roland Horvath, Sandor Toth, and Balint Magyar

The key infrastructural elements of the geodetic application of the Interferometric Synthetic Aperture Radar (InSAR) are the integrated benchmarks which are satellite technologies, besides the traditional geodetic technologies, therefore they serve as benchmarks of the Global National Satellite System (GNSS) and InSAR.

Previously, the Satellite Geodetic Observatory (SGO) has already built a network of the passive corner reflectors (SENGA) near the Hungarian GPS Geokinematic Reference Network. This infrastructure is added by an active corner reflector (called transponder) which is the first device according to our knowledge in Hungary. We have been testing the transponder in recent months. The scope of our work is the detection of the intensity of the emitted radar signal by the Sentinel-1 C-band satellite VV polarisation sensor using GAMMA Remote Sensing Software with 6 day repeat cycle availability of satellite images in ascending and descending passes. Hence, we could monitor and compare of the pixel-intensity (expressed in decibel) before and after the installation. The value of the pixel is increased around 15-20 dB and we had chance to set the Radar Cross Section (RCS=31 dBm2) against the results of existing researches. During the testing period the ECR was placed on the rooftop of the SGO, but in the short-term the design of the relocation of the device as InSAR Persistent Scatterer has also been developed.

One of the goals of our research is the incorporation of the transponders into the SENGA network which is needed to be expanded, examination and determination of the conditions of this integration.

 

How to cite: Horvath, R., Toth, S., and Magyar, B.: Testing of the usability of the C-band Electronic Corner Reflector (ECR-C) in Hungary based on Sentinel-1 SAR data, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-1070, https://doi.org/10.5194/egusphere-egu21-1070, 2021.

11:25–11:27
|
EGU21-10092
|
ECS
Alberto Pastorutti and Carla Braitenberg

Partitioning of the Earth surface in "provinces": tectonic domains, outcropping geological units, crustal types, discrete classes extracted from age or geophysical data (e.g. tomography, gravity) is often employed to perform data imputation of ill-sampled observables (e.g. the similarity-based NGHF surface heat flow map [1]) or to constrain the parameters of ill-posed inverse problems (e.g. the gravimetric global Moho model GEMMA [2]).

We define provinces as noncontiguous areas where quantities or their relationships are similar. Following the goodness metric employed for proxy observables, an adequate province model should be able to significantly improve prediction of the extrapolated quantity. Interpolation of a quantity with no reliance on external data sets a predictivity benchmark, which a province-based prediction should exceed.
In a solid Earth modelling perspective, gravity, topography, and their relationship, seem ideal candidates to constrain a province clustering model. Earth gravity and topography, at resolutions of at least 100 km, are known with an incomparable sampling uniformity and negligible error, respect to other observables.

Most of the observed topography-gravity relationship can be explained by regional isostatic compensation. The topography, representing the load exerted on the lithosphere, is compensated by the elastic, thin-shell like response of the latter. In the spectral domain, flexure results in a lowpass transfer function between topography and isostatic roots. The signal of both surfaces, superimposed, is observed in the gravity field.
However, reality shows significant shifts from the ideal case: the separation of nonisostatic effects [3], such as density inhomogeneities, glacial isostatic adjustments, dynamic mantle processes, is nontrivial. Acknowledging this superposition, we aim at identifying clusters of similar topography-gravity transfer functions.

We evaluate the transfer functions, in the form of admittance and correlation [4], in the spherical harmonics domain. Spatial localization is achieved with the method by Wieczorek and Simons [5], using SHTOOLS [6]. Admittance and correlation spectra are computed on a set of regularly spaced sample points, each point being representative of the topo-gravity relationship in its proximity. The coefficients of the localized topo-gravity admittance and correlation spectra constitute each point features.

We present a set of experiments performed on synthetic models, in which we can control the variations of elastic parameters and non-isostatic contributions. These tests allowed to define both the feature extraction segment: the spatial localization method and the range of spherical harmonics degrees which are more sensible to lateral variations in flexural rigidity; and the clustering segment: metrics of the ground-truth clusters, performance of dimensionality reduction methods and of different clustering models.

[1] Lucazeau (2019). Analysis and Mapping of an Updated Terrestrial Heat Flow Data Set. doi:10.1029/2019GC008389
[2] Reguzzoni and Sampietro (2015). GEMMA: An Earth crustal model based on GOCE satellite data. doi:10.1016/j.jag.2014.04.002
[3] Bagherbandi and Sjöberg (2013). Improving gravimetric–isostatic models of crustal depth by correcting for non-isostatic effects and using CRUST2.0. doi:10.1016/j.earscirev.2012.12.002
[4] Simons et al. (1997). Localization of gravity and topography: Constraints on the tectonics and mantle dynamics of Venus. doi:10.1111/j.1365-246X.1997.tb00593.x
[5] Wieczorek and Simons (2005). Localized spectral analysis on the sphere. doi:10.1111/j.1365-246X.2005.02687.x
[6] Wieczorek and Meschede (2018). SHTools: Tools for Working with Spherical Harmonics. doi:10.1029/2018GC007529

How to cite: Pastorutti, A. and Braitenberg, C.: Geologic provinces from unsupervised learning: synthetic experiments in clustering of localized topography/gravity admittance and correlation spectra, EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-10092, https://doi.org/10.5194/egusphere-egu21-10092, 2021.

11:27–11:29
|
EGU21-14999
|
ECS
Avadh Bihari Narayan, Ashutosh Tiwari, Govind Sharma, Balaji Devaraju, and Onkar Dikshit

The spherical approximation of the fundamental equation of geodesy defines the boundary value problems. Stokes’s integral provides the solution of boundary value problems that enables the computation of geoid from the properly reduced gravity measurements to the geoid. The stokes integral can be evaluated by brute-force numerical integration, spectral methods, and least-squares collocation. There is a trade-off between computation time and accuracy when we chose numerical integration technique or any spectral method. This research will compare time complexity and the accuracy of different spectral methods (1D-FFT, 2D-FFT, Multi-band FFT) and numerical integration technique for the region in the lower Himalaya, around Nainital, Uttarakhand, India. 

How to cite: Narayan, A. B., Tiwari, A., Sharma, G., Devaraju, B., and Dikshit, O.: Comparison of spectral methods for the evaluation of Stokes integral , EGU General Assembly 2021, online, 19–30 Apr 2021, EGU21-14999, https://doi.org/10.5194/egusphere-egu21-14999, 2021.

11:29–11:45