Difference between revisions of "JSG T.29"

From Icctwiki
Jump to: navigation, search
m (Pendl moved page JSG0.16 to JSG T.29)
Line 1: Line 1:
<big>'''JSG T.29: Earth’s inner structure from combined geodetic and geophysical sources'''</big>
+
<big>'''JSG T.29: Machine learning in geodesy'''</big>
  
Chairs: ''Robert Tenzer (China)''<br>
+
Chairs: ''Benedikt Soja (USA), chair''<br>
Affiliation: ''Comm. 2 and 3''
+
Affiliation: ''Commissions 2, 3 and 4''
  
 
__TOC__
 
__TOC__
Line 8: Line 8:
 
===Introduction===
 
===Introduction===
  
The satellite gravimetry missions, CHAllenging Mini-satellite Payload (CHAMP), the GRavity field and Climate Experiment (GRACE) and the Gravity field and steady-state Ocean Circulation Explorer (GOCE), significantly improved our knowledge on the external gravitational field of the Earth at the long-to-medium wavelengths (approximately up to a spherical harmonic degree of 250). Such improved information in terms of the accuracy and resolution has been utilized in studies of the Earth’s interior for a better understanding of the Earth’s inner structure and processes occurring within the lithosphere and sub-lithospheric mantle. Whereas the long-wavelength spectrum of the Earth’s gravitational field comprises mainly the signature of deep mantle density heterogeneities attributed to mantle convection, the medium wavelengths reflect the density structure of more shallow sources within the lithosphere. This allows studying and interpreting in more detail the gravitational features which are related to the global tectonism (including the oceanic subduction, orogenic formations, earthquakes, global lithospheric plate configuration, etc.), sub-lithospheric stresses, isostatic mechanisms, glacial isostatic adjustment, and other related geodynamic phenomena. Moreover, the Global Gravitational Models (GGMs) have been extensively used in studies of the lithospheric density structure and density interfaces such as for the gravimetric recovery of the Moho depth, lithospheric thickness as well as structure of sedimentary basins.  
+
Due to the exponential increase in computing power over the last decades, machine learning has grown in importance for several applications. In particular, deep learning, i.e., machine learning based on deep neural networks, typically performed on extensive data sets (“big data”), has become very successful in tackling various challenges, for example, image interpretation, language recognition, autonomous decision making or stock market predictions. Several scientific disciplines have embraced the capability of modern machine learning algorithms, including astronomy and many fields of geosciences.  
  
Since the gravity observations could not be used alone to interpret the Earth’s inner density structure due to a non-uniqueness of inverse solutions (i.e. infinity many 3-D density structures could be attributed to the Earth’s gravity field), additional information is required to constrain the gravimetric methods for interpreting the Earth’s interior. These constraining data comprise primarily results of seismic surveys as well as additional geophysical, geothermal and geochemical parameters of the Earth. Moreover, numerous recent gravimetric studies of the Earth’s interior focus on the global and regional Moho recovery. The classical isostatic models (according to Airy and Pratt theories) are typically not able to model realistically the actual Moho geometry, due to the fact that the isostatic mass balance depends on loading and effective elastic thickness, rigidity, rheology of the lithosphere and viscosity of the asthenosphere. Moreover, geodynamic processes such as the glacial isostatic adjustment, present-day glacial melting, plate motion and mantle convection contribute to the time-dependent isostatic balance. To overcome these issues, processing strategies of combining gravity and seismic data (and possibly also additional constraining information) have to be applied to determine the actual Moho geometry.  
+
The field of geodesy has seen a significant increase in observational data in recent years, in particular from Global Navigation Satellite Systems (GNSS) with tens of thousands of high-quality permanent stations, multiple constellations, and increasing data rates. With the upcoming NISAR mission, the InSAR community needs to prepare for handling daily products exceeding 50 GB. In the future, the next-generation Very Long Baseline Interferometry (VLBI) Global Observing System (VGOS) will deliver unprecedented amounts of data compared to legacy VLBI operations. Traditional data processing and analysis techniques that rely largely on human input are not well suited to harvest such rich data sets to their full potential. Still, machine learning techniques are not yet adopted in geodesy.  
  
The gravimetric methods applied in studies of the Earth’s inner density structure comprise - in principle - two categories. The methods for the gravimetric forward modeling are applied to model (and remove) the gravitational signature of known density structures in order to enhance the gravitational contribution of unknown (and sought) density structures and interfaces. The gravimetric inverse methods are then used to interpret these unknown density structures from the refined gravity data. It is obvious that the combination of gravity and seismic data (and other constraining information) is essential especially in solving the gravimetric inverse problems.  
+
Machine learning in geodesy has the potential to facilitate the automation of data processing, detection of anomalies in time series and image data, their classification into different categories and prediction of parameters into the future. Machine learning and, in recent years, deep learning methods can successfully model complex spatio-temporal data through the creation of powerful representations at hierarchical levels of abstraction. Furthermore, machine learning techniques provide promising results in addressing the challenges that arise when handling multi-resolution, multi-temporal, multi-sensor, multi-modal data. The information contained in GNSS station position time series is essential as it can help derive important conclusions related to hydrology, earthquakes, or volcanism using machine learning. Other important applications are tropospheric and ionospheric parameters derived from GNSS where automated detection and prediction could be beneficial for improved severe weather forecasting and space weather monitoring, respectively. InSAR data will benefit in particular from efficient image processing algorithms based on machine learning, facilitating the detection of regions of interest. In several of these cases, the development of scalable deep learning schemes can contribute to more effectively handling and processing of large-scale spatio-temporal data.  
  
This gives us the platform and opportunities towards improving the theoretical and numerical methods applied in studies of Earth’s interior from multiple data sources, primarily focusing but not restricting only to combining gravimetric and seismic data. It is expected that the gravity data could improve our knowledge of the Earth’s interior over significant proportion of the world where seismic data are sparse or completely absent (such large parts of oceanic areas, Antarctica, Greenland and Africa). The gravity data could also provide additional information on the lithospheric structure and mechanisms, such as global tectonic configuration, geometry of subducted slabs, crustal thickening of orogenic formations and other phenomena.
+
Traditional machine learning techniques for geodetic tasks include convolutional neural networks for image data and recurrent neural networks for time series data. Typically, these networks are trained by supervised learning approaches, but certain applications related to autonomous processing will benefit from reinforcement learning.
 +
 +
The field of machine learning has expanded rapidly in recent years and algorithms are constantly evolving. It is the aim of this JSG to identify best practices, methods, and algorithms when applying machine learning to geodetic tasks. In particular, due to the “black box” nature of many machine learning techniques, it is very important to focus on appropriate ways to assess the accuracy and precision of the results, as well as to correctly interpret them.
  
 
===Objectives===
 
===Objectives===
  
* Development of the theoretical and numerical algorithms for combined processing of gravity, seismic and other types of geophysical data for a recovery of the Earth’s density structures and interfaces.
+
* Identify geodetic applications that could benefit from machine learning techniques, both in terms of which data sets to use and which issues to investigate.  
* Development of fast numerical algorithms for combined data inversions.
+
* Create an inventory of suitable machine learning algorithms to address these problems, highlighting their strengths and weaknesses.
* Development of stochastic models for combined inversion including optimal weighting, regularization and spectral filtering.
+
* Perform comparisons between machine learning methods and traditional data analysis approaches, e.g., for time series analysis and prediction.
* Better understanding of uncertainties of interpreted results based on the error analysis of input data and applied numerical models. Geophysical and geodynamic clarification of results and their uncertainties.
+
* Focus on error assessment of results produced by machine learning algorithms.
* Recommendations for optimal data combinations, better understanding of possibilities and limiting factors associated with individual data types used for geophysical and geodynamic interpretations.
+
* Identify open problems that come with the automation of data processing and generation of geodetic products, including issues of reliability.
 +
* Develop best practices when applying machine learning methods in geodesy and establishing standardized terminology.
  
 
===Program of activities===
 
===Program of activities===
  
* Launching of a web page with emphasis on exchange of ideas and recent progress, providing and updating bibliographic list of references of research results and relevant publications from different disciplines.
+
* Create a web page about machine learning in geodesy to provide information and raise awareness about this topic. The page will include:
* Work progress meetings at the international symposia and presentation of research results at the appropriate sessions.
+
** inventory of algorithms, see above,
* Possible collaboration between various geoscience study groups dealing with the modeling of the Earth’s interior and related scientific topics.  
+
** benchmark datasets to test the performance of these algorithms,
 +
** comprehensive record of previous activities/publications related to machine learning in geodesy,
 +
** description of activities by the JSG members.  
 +
* Work toward a state-of-the-art review paper about machine learning in geodesy co-authored by the JSG members.
 +
* Promote sessions and presentation of the research results at international scientific assemblies (IAG/IUGG, EGU, AGU) and technique-specific meetings (IGS, IVS, ...).  
  
 
===Members===
 
===Members===
  
'' '''Robert Tenzer (China), chair''' <br /> Lars Sjöberg (Sweden) <br /> Mohammad Bagherbandi (Sweden) <br /> Carla Braitenberg (Italy) <br /> Mehdi Eshagh (Sweden) <br /> Mirko Reguzzoni (Italy) <br /> Xiaodong Song (USA) <br />''
+
'' '''Benedikt Soja (USA), chair ''' <br /> Kyriakos Balidakis (Germany) <br /> Clayton Brengman (USA) <br /> Jingyi Chen (USA) <br /> Maria Kaselimi (Greece) <br />
 +
Ryan McGranaghan (USA) <br /> Randa Natras (Germany) <br /> Simone Scardapane (Italy) <br />''

Revision as of 10:54, 10 June 2020

JSG T.29: Machine learning in geodesy

Chairs: Benedikt Soja (USA), chair
Affiliation: Commissions 2, 3 and 4

Introduction

Due to the exponential increase in computing power over the last decades, machine learning has grown in importance for several applications. In particular, deep learning, i.e., machine learning based on deep neural networks, typically performed on extensive data sets (“big data”), has become very successful in tackling various challenges, for example, image interpretation, language recognition, autonomous decision making or stock market predictions. Several scientific disciplines have embraced the capability of modern machine learning algorithms, including astronomy and many fields of geosciences.

The field of geodesy has seen a significant increase in observational data in recent years, in particular from Global Navigation Satellite Systems (GNSS) with tens of thousands of high-quality permanent stations, multiple constellations, and increasing data rates. With the upcoming NISAR mission, the InSAR community needs to prepare for handling daily products exceeding 50 GB. In the future, the next-generation Very Long Baseline Interferometry (VLBI) Global Observing System (VGOS) will deliver unprecedented amounts of data compared to legacy VLBI operations. Traditional data processing and analysis techniques that rely largely on human input are not well suited to harvest such rich data sets to their full potential. Still, machine learning techniques are not yet adopted in geodesy.

Machine learning in geodesy has the potential to facilitate the automation of data processing, detection of anomalies in time series and image data, their classification into different categories and prediction of parameters into the future. Machine learning and, in recent years, deep learning methods can successfully model complex spatio-temporal data through the creation of powerful representations at hierarchical levels of abstraction. Furthermore, machine learning techniques provide promising results in addressing the challenges that arise when handling multi-resolution, multi-temporal, multi-sensor, multi-modal data. The information contained in GNSS station position time series is essential as it can help derive important conclusions related to hydrology, earthquakes, or volcanism using machine learning. Other important applications are tropospheric and ionospheric parameters derived from GNSS where automated detection and prediction could be beneficial for improved severe weather forecasting and space weather monitoring, respectively. InSAR data will benefit in particular from efficient image processing algorithms based on machine learning, facilitating the detection of regions of interest. In several of these cases, the development of scalable deep learning schemes can contribute to more effectively handling and processing of large-scale spatio-temporal data.

Traditional machine learning techniques for geodetic tasks include convolutional neural networks for image data and recurrent neural networks for time series data. Typically, these networks are trained by supervised learning approaches, but certain applications related to autonomous processing will benefit from reinforcement learning.

The field of machine learning has expanded rapidly in recent years and algorithms are constantly evolving. It is the aim of this JSG to identify best practices, methods, and algorithms when applying machine learning to geodetic tasks. In particular, due to the “black box” nature of many machine learning techniques, it is very important to focus on appropriate ways to assess the accuracy and precision of the results, as well as to correctly interpret them.

Objectives

  • Identify geodetic applications that could benefit from machine learning techniques, both in terms of which data sets to use and which issues to investigate.
  • Create an inventory of suitable machine learning algorithms to address these problems, highlighting their strengths and weaknesses.
  • Perform comparisons between machine learning methods and traditional data analysis approaches, e.g., for time series analysis and prediction.
  • Focus on error assessment of results produced by machine learning algorithms.
  • Identify open problems that come with the automation of data processing and generation of geodetic products, including issues of reliability.
  • Develop best practices when applying machine learning methods in geodesy and establishing standardized terminology.

Program of activities

  • Create a web page about machine learning in geodesy to provide information and raise awareness about this topic. The page will include:
    • inventory of algorithms, see above,
    • benchmark datasets to test the performance of these algorithms,
    • comprehensive record of previous activities/publications related to machine learning in geodesy,
    • description of activities by the JSG members.
  • Work toward a state-of-the-art review paper about machine learning in geodesy co-authored by the JSG members.
  • Promote sessions and presentation of the research results at international scientific assemblies (IAG/IUGG, EGU, AGU) and technique-specific meetings (IGS, IVS, ...).

Members

Benedikt Soja (USA), chair
Kyriakos Balidakis (Germany)
Clayton Brengman (USA)
Jingyi Chen (USA)
Maria Kaselimi (Greece)
Ryan McGranaghan (USA)
Randa Natras (Germany)
Simone Scardapane (Italy)