 Full paper
 Open access
 Published:
Timelapse inversion of onedimensional magnetotelluric data
Earth, Planets and Space volumeÂ 70, ArticleÂ number:Â 27 (2018)
Abstract
We present a new tool for modelling timelapse magnetotelluric (MT) data, an emerging technique for monitoring changes in subsurface electrical resistivity. Timelapse MT data have been acquired in various settings, including sites of hydraulic fracturing, dewatering and sequestration. It has been shown in other geophysical techniques that the most effective way to model timelapse data is with simultaneous inversion, which uses information from all timesteps to produce models with higher accuracy and fewer artefacts. We introduce this method to model timelapse 1D MT data. As with a standard MT inversion, our routine penalises spatial roughness at each timestep, however we also introduce temporal regularisation. The inversion is simple to apply, requiring only the ratio between regularisation parameters and the desired level of misfit from the user. The algorithm is tested on both synthetic data, and a case study. We find that in the synthetic example our inversion successfully retrieves the main characteristics of the test model and introduces minimal artefacts, even in the presence of significant noise. We also test the effect of changing the ratio of regularisation parameters. In the case study, we produce an easily interpretable model that compares favourably with previous inversions of the synthetic data. We conclude that timelapse modelling of 1D MT data can be a valuable tool for imaging subsurface change.
Introduction
In recent years, the magnetotelluric method (MT) has been increasingly used as a costeffective technique for subsurface resistivity monitoring (see Rees etÂ al. 2016, for a general introduction). As this is a relatively new application for MT, there are few modelling tools available. Peacock etÂ al. (2012) have used differences in MT phase tensors to qualitatively interpret resistivity changes due to fluid injection in an enhanced geothermal system (EGS) at Paralana, South Australia. This technique has also be used by Didana etÂ al. (2017) interpreting change at the Habernero EGS. Recent work includes parameterising resistivity changes as a threedimensional (3D) plume structure and inverting using Markov Chain Monte Carlo (RosasCarbajal etÂ al. 2015). Another approach has been to use 1D layerstripping, where the effect of overlying structures is removed to model the timevarying magnetotelluric responses at depth (Ogaya etÂ al. 2016). Standard MT modelling tools such as inversion can also be adapted, for example Rees etÂ al. (2016) use cascading twodimensional (2D) inversions to model a coalseam gas depressurisation, where the results from the preinjection inversion are used as a priormodel for the postinjection. An improvement on cascaded inversion has been implemented by RosasCarbajal etÂ al. (2012), where differenced 2D MT inversions have achieved high accuracy by subtracting the priormodel response from all data and reducing the data error. Highaccuracy simultaneous timelapse 1D MT inversions have also been approximated using 2D MT codes (e.g. Rees etÂ al. 2016; Didana etÂ al. 2017); however, this approximation leads to small errors in the calculation of the forwardmodel and hence inversion result.
The problem of inverting timelapse data is fairly well researched in other geophysical methods. Hayley etÂ al. (2011) have compared the various techniques used for inverting electrical resistivity tomography (ERT) timelapse data. They showed that a simultaneous inversion of timelapse ERT data gives a superior result compared to independent inversions, cascading inversions or differenced inversions. Their preferred technique was a special case of the 4D algorithm of Kim etÂ al. (2009), which has also successfully been applied to gravity (Karaoulis etÂ al. 2013a), induced polarisation (Karaoulis etÂ al. 2013b), and seismic tomography (Karaoulis etÂ al. 2015). This technique has not yet been presented for magnetotelluric inversion.
We present an implementation of a simple timelapse algorithm to simultaneously invert 1D MT monitoring data. Firstly, we present the results from synthetic inversions of 1D magnetotelluric data to test the validity of the algorithm. We then present inversions from a case study of data from a coalseam gas production survey (Rees etÂ al. 2016).
Inversion method
The algorithm is an Occam style inversion (Constable etÂ al. 1987), seeking to find the model which fits the data to a desired level whilst introducing the least amount of structureâ€”in this case in both temporal and spatial domains. We seek to find the model (\(\mathbf {m}\)) which minimises the model roughness (\(R(\mathbf {m})\)) subject to the constraint that the model misfit (\(\chi ^2(\mathbf {m})\)) is less than a desired target misfit (\(\chi ^2_\mathrm{target}\)). That is,
\(\chi ^2\) misfit is given by the weighted differences of n 1D MT forward responses (\(F(\mathbf {m})\)) with the data (d) and data errors (\(\sigma\)):
It is more convenient to express the misfit in terms of rootmeansquare error (RMS), given by
The model roughness is the sum of regularisation in both spatial (S) and temporal (T) directions, weighted by a factor \(\beta\):
The temporal regularisation is given by the squared sum of the model changes in time:
and the spatial regularisation by the sum of second derivatives at each time slice:
The optimisation is performed using the Sequential Least SQuares Programming (SLSQP) algorithm available in scipy (Jones etÂ al. 2001), which is based on the algorithm by Kraft and Schnepper (1989). The algorithm does not require any tradeoff parameter between data misfit and smoothing, as this is encapsulated in the target RMS. Data are read into the program using the mtpy module (Krieger and Peacock 2014).
Due to the efficiency of the 1D MT algorithm, which is calculated analytically, the model readily reaches an acceptable misfit within a few iterations. Subsequent iterations reduce the roughness of the model.
Synthetic study
Numerical data are first used to study the results of the timelapse MT algorithm. The data are produced from a model of an idealised resistivity change at a single site for a coalseam gas pump into a shallow aquifer. The initial resistivity is a 10Â \(\Omega\)m halfspace. At survey day 4 a sharp change occurs at a depth of 665â€“958Â m, with the resistivity dropping to 2Â \(\Omega\)m. The input model for the synthetic study is shown in Fig.Â 1.
Magnetotelluric impedance data are produced at 17 frequencies ranging from 0.14 to 6.26Â Hz. The data were computed using the standard analytic 1D forward algorithm. The 1D assumption is valid if the resistivity changes are sufficiently laterally continuous, which we would expect for changes in a lateral coalbed. The assumption is not perfect however, as 3D effects would be present during the development of the resistivity front. The data were then contaminated with 5% Gaussian random noise on the impedance tensor, which is a reasonably high estimate of the errors expected in field operations. The data errors were fixed at 5%.
The inversion was conducted with a \(RMS_\mathrm{target}\) of 1.0. Smaller target RMS values result in overfitted models which mapped noise in the data as temporal change, whilst larger target RMS values underfit the change in the data. The \(\beta\) value, which weights the relative importance of temporal and spatial change, was varied between in three separate inversions with values of \(\beta = 100\), \(\beta = 1000\) and \(\beta = 10,000\) in order to show the effect of this parameter.
For each inversion the target RMS was achieved within 5 iterations, however additional smoothing iterations were run, with a total of 100 iterations for each inversion.
Results and discussion
FigureÂ 2 shows the resistivity for three synthetic inversions with varying ratios of temporal smoothing to spatial smoothing. On the left panels the absolute resistivity is shown, and on the right panels the change in resistivity relative to the initial resistivity is displayed.
The main feature in common to all three figures is the strong decrease in conductivity from day 2â€“3 onwards. In each inversion, the change is diffusively spread out across multiple layers, with the amount of spread dependent on the value for \(\beta\). The diffusive spreading is most evident with the smallest \(\beta\) value in Fig.Â 2a, and the largest value of \(\beta\) in Fig.Â 2c has the changes confined to almost a single layer.
Compared to the input model, each inversion model underestimates the change in the data. The relative change in the input model is an 80% drop over two layers, whereas the three inversion models have drops in the order of 50â€“60%. The extent of the conductive change is overestimated by the \(\beta = 100\) model and underestimated by the \(\beta = 10,000\) model, with the middle \(\beta = 1000\) model providing a fairly accurate image. With increasing \(\beta\) the models also become rougher spatially. In the \(\beta = 10,000\) scenario, for example, the model changes from 1 to 100Â \(\Omega\)m between three layers.
Example model fit curves from the first and final days are overlain in Fig.Â 3 for the \(\beta = 1000\) inversion, plotted alongside the synthetic data used in the tests. The curves show that a good fit to the data is achieved at convergence. They also highlight the substantial change in the synthetic data due to the resistivity changes.
The inversions were reasonably accurate in terms of the absolute resistivities and the relative changes in resistivity. The smaller \(\beta\) value, however, spread the changes out over a large spatial area. The algorithm will preference changes over as many layers as possible for two reasons. Firstly, the temporal roughness is calculated on the square of the change in any one layer, which means that it is preferable to spread changes over multiple layers if possible. Secondly, and more importantly, the entire model is smoothed in the spatial dimension. The lowest \(\beta\) parameter resulted in a less concentrated area of change, with smaller changes spread over a greater area. The highest \(\beta\) weighting gave a stronger concentration of change and a smaller overall change in resistivity. The model, however, became rougher in the spatial dimension leading to improbable absolute resistivities.
It is worth noting that the inverted models accurately pick the time when the resistivity change occurs. There is a slight change in resistivity before the inversion; howeverÂ this change is unavoidable as the small data misfit which it introduces is offset by the improved temporal smoothing. If a priori information exists regarding the time \(t_c\) when a resistivity change occurs, it can be included by restricting all temporal change for \(t < t_c\).
Case study
An example from a coalseam gas field is used to further examine the algorithm. A coalseam gas monitoring experiment is ideal to test the 1D timelapse MT algorithm, as changes occur at shallow depths where the MT response is better approximated by a 1D algorithm, and we expect any changes to be laterally continuous through the coal seam.
We use timelapse MT data from a 2013/2014 survey of a coalseam gas field in the Surat Basin, Queensland, Australia (Rees etÂ al. 2016). Data for this survey were collected along two lines near several production wells. The wells extracted both water and gas from depths of 400â€“700Â m, and a resulting resistivity change was observed by Rees etÂ al. (2016) in differenced MT inversions.
A schematic of the survey design is shown in Fig.Â 4, which also features production well locations. Site 105 was used for our inversion, which is the same site used for the 1D inversions in Rees etÂ al. (2016). Data were used from 10 days between January 23, 2014 and February 19, 2014. The site is nearby the most active production well and had reasonable data quality during these days. MT data are taken from the YX mode as these are of higher quality than the XY mode data.
Identical to the synthetic study, the data have 17 frequencies ranging from 0.14 to 6.26Â Hz. We invert these data using the presented timelapse algorithm. We use parameters \(RMS_\mathrm{target}=1.0\), \(\beta =1000\) and an error floor of 2%. The inversion is run for 100 iterations and completed within 20Â min on a 4 core machine.
Results and discussion
The inversion resistivity model is shown in Fig.Â 5, with the resistivity differences shown in the same figure. A slight reduction in resistivity between 500 and 700Â m begins on day 9 and slowly increases with time, culminating with a drop of roughly 15% in resistivity by the final day. There is also a slight (\(\approx\) 5%) increase in resistivity in the area between 100 and 500Â m. Finally, there is also a reduction in resistivity at depth; however, this is smaller than the resistivity drop in the higher zone. Notably this would be near the penetration depth of the data, and would be less constrained than the other areas of the model. The entire model resistivity is between 1.25 and 16Â \(\Omega\)m. When considering the commencement date of pumping, shown as a dashed line in Fig.Â 5, we see that the resistivity changes occur slightly before the commencement of dewatering. In the synthetic model, temporal changes were smoothedÂ by the algorithm toÂ include dates before the change in the target model, and we would suggest that the same phenomenon is occurring here.
Looking at the fit of the response curves in Fig.Â 6, we see that the model slightly underfits some of the changes in the MT data. This is expected as the algorithm introduces the minimum change required to fit the data. This would place the estimations of 15% changeÂ in resistivity as conservative and shows the importance of obtaining highquality data, as tighter error bars would lead to stronger changes in resistivity. Notably there is a significant dip in the model fit curves between the first (in blue) and final (in green) time slices.
The most important area to consider in interpretation is the area of the resistivity decrease between 500 and 700Â m, as the other resistivity changes in the model are much smaller and should not be overinterpreted. There seems to be a strong link between the start of the resistivity change and the dewatering event at W4. There is no gas extraction during this time. Hence, our new model would agree with the interpretation in Rees etÂ al. (2016) that the changes are due to the increased permeability of dewatered coalseams, potentially due to the reduction in the coal matrixÂ as gas is released from the matrix.
Unlike the synthetic example, the changes in the model were slow, with a build up to the maximum change in the final day included in the inversion. This would lead us to believe that resistivity changesÂ resulting from coalseam gas pumping are related not just to the rate of pumping, but also the amount of material that has been removed from the pores.
Conclusion
We presented a simple timelapse inversion of 1D MT data, which we tested on synthetic data and casestudy data. Both inversions resulted in defined areas of resistivity change with few artefacts in the models. The synthetic inversion obtained resistivity changes which slightly underestimated the changes in the input model; however, it accurately retrieved spatial and temporal locations, as well as absolute resistivities. We investigated the effect of the weighting parameter \(\beta\) between spatial and temporal smoothing and showed that there is a tradeoff between models which are overly sharp spatially and models which overestimate the extent of the resistivity change. We also showed results from applying the algorithm to a case study with coalseam gas MT data. The inversion resulted in small changes in the area of the coalseams.
Compared to other modelling techniques, our algorithm has several advantages. Full 3D modelling is highly computationally intensive. A 1D approximation is simple to compute, which leads to aÂ rapidly converging model. This allows the modeller to trial several parameters and select the best model. It also allows a large amount of data to be inverted at onceâ€”it is feasible to invert data from each day for several months, which would result in an extremely large model space in 3D. Compared to independent inversions, or timelapse inversions using 2D MT codes as approximations, it has previously been shown that simultaneous inversion results in a model with fewer artefacts and stronger constraints on areas of change.
One of the disadvantages of the technique is that it is unable to fit any 2D or 3D aspects of the data. Further work could expand the technique into higher spatial dimensions, or to deal with anisotropy. It would be useful to extend the inversion into 3D space and incorporate all data into the inversion; however, the long processing times of higherdimensional MT forward codes would make the resulting code extremely computationally intensive.
Abbreviations
 MT:

magnetotellurics
 1D:

onedimensional
 RMS:

rootmeansquared error
References
Constable S, Parker R, Constable C (1987) Occamâ€™s inversion: a practical algorithm for generating smooth models from electromagnetic sounding data. Geophysics 52(3):289â€“300. https://doi.org/10.1190/1.1442303
Didana YL, Heinson G, Thiel S, Krieger L (2017) Magnetotelluric monitoring of permeability enhancement at enhanced geothermal system project. Geothermics 66:23â€“38. https://doi.org/10.1016/j.geothermics.2016.11.005
Hayley K, Pidlisecky A, Bentley LR (2011) Simultaneous timelapse electrical resistivity inversion. J Appl Geophys 75(2):401â€“411. https://doi.org/10.1016/j.jappgeo.2011.06.035
Jones E, Oliphant T, Peterson P et al (2001) SciPy: open source scientific tools for python. http://www.scipy.org/.Â Â Accessed 10 Mar 2017
Karaoulis M, Revil A, Minsley B, Todesco M, Zhang J, Werkema DD (2013a) Timelapse gravity inversion with an active time constraint. Geophys J Int 196(2):748â€“759
Karaoulis M, Revil A, Tsourlos P, Werkema DD, Minsley BJ (2013b) IP4DI: a software for timelapse 2D/3D DCresistivity and induced polarization tomography. Comput Geosci 54:164â€“170. https://doi.org/10.1016/j.cageo.2013.01.008
Karaoulis M, Werkema DD, Revil A (2015) 2D timelapse seismic tomography using an active timeconstraint (ATC) approach. Lead Edge 34(2):206â€“212. https://doi.org/10.1190/tle34020206.1
Kim JH, Yi MJ, Park SG, Kim JG (2009) 4D inversion of DC resistivity monitoring data acquired over a dynamically changing earth model. J Appl Geophys 68(4):522â€“532. https://doi.org/10.1016/j.jappgeo.2009.03.002
Kraft D, Schnepper K (1989) SLSQPâ€”a nonlinear programming method with quadratic programming subproblems. DLR, Oberpfaffenhofen
Krieger L, Peacock JR (2014) MTpy: a python toolbox for magnetotellurics. Comput Geosci 72:167â€“175
Ogaya X, Ledo J, Queralt P, Jones AG, Marcuello l (2016) A layer stripping approach for monitoring resistivity variations using surface magnetotelluric responses. J Appl Geophys 132:100â€“115. https://doi.org/10.1016/j.jappgeo.2016.06.014
Peacock JR, Thiel S, Reid P, Heinson G (2012) Magnetotelluric monitoring of a fluid injection: example from an enhanced geothermal system. Geophys Res Lett 39(17):3â€“7
Rees N, Carter S, Heinson G, Krieger L, Conway D, Boren G, Matthews C (2016) Magnetotelluric monitoring of coalseam gas and shalegas resource development in Australia. Lead Edge 35(1):64â€“70. https://doi.org/10.1190/tle35010064.1
RosasCarbajal M, Linde N, Kalscheuer T (2012) Focused timelapse inversion of radio and audio magnetotelluric data. J Appl Geophys 84:29â€“38. https://doi.org/10.1016/j.jappgeo.2012.05.012
RosasCarbajal M, Linde N, Peacock J, Zyserman FI, Kalscheuer T, Thiel S (2015) Probabilistic 3D timelapse inversion of magnetotelluric data: application to an enhanced geothermal system. Geophys J Int 203(3):1946â€“1960
Authorsâ€™ contributions
DC performed algorithm development and testing with help from GH and JR. NR helped in the casestudy test by assisting with the data selection and modelling. All authors read and approved the final manuscript.
Acknowledgements
We would like to thank QGC Pty Limited for their funding of the data collection, as well The Australian Geophysics Observing System and AuScope for providing the instrumentation. We thank S. Rana, G. Hicks, P. McKelvey and A. Smart for their help with field logistics, and S. Schnaidt, G. Boren, A. Tullis, B. Warren, S. MacMillan and C. Matthews for their assistance in field work. Finally, S. Carter and O. Putland and thanked for their help with data processing.
Competing interests
The authors declare that they have no competing interests.
Availability of data and materials
Included in supplementary material.
Consent for publication
Not applicable.
Ethics approval and consent to participate
Not applicable.
Funding
Not applicable.
Publisherâ€™s Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.
About this article
Cite this article
Conway, D., Heinson, G., Rees, N. et al. Timelapse inversion of onedimensional magnetotelluric data. Earth Planets Space 70, 27 (2018). https://doi.org/10.1186/s406230180791y
Received:
Accepted:
Published:
DOI: https://doi.org/10.1186/s406230180791y