tTEM20AAR: a benchmark geophysical dataset for unconsolidated ﬂuvio-glacial sediments

. Quaternary deposits are complex and heterogeneous. They contain some of the most abundant and extensively used aquifers. In order to improve the knowledge of the spatial heterogeneity of such deposits, we acquired a large (more than 1400 hectares) and dense (20 m spacing) Time Domain ElectroMagnetic (TDEM) dataset in the upper Aare Valley, Switzerland. TDEM is a fast and reliable method to measure the magnetic ﬁeld directly related to the resistivity of the underground. In this paper, we present the inverted resistivity models derived from this acquisition, and all the necessary data in order to perform 5 different inversions on the processed data (https://doi.org/10.5281/ZENODO.4269887 (Neven et al., 2020)). The depth of investigation ranges between 40 to 120 m depth, with an average data residual contained in the standard deviation of the data. These data can be used for many different purposes: from sedimentological interpretation of quaternary environments in alpine environments, geological and hydrogeological modeling, to benchmarking geophysical inversion techniques.

sharp variations due to the nearest neighbors interpolation method used during the workflow. Therefore, there is a need for a 25 better understanding of Quaternary sedimentary heterogeneity, in order to better constrain geological models, knowledge that could be applied in the Aare valley or for any fluvio-glacial filling area.
Near-surface geophysics such as DC resistivity, electromagnetic or seismic methods can bring important information in terms of the spatial distribution of facies. However, they are usually carried out in restricted areas to answer specific local questions, and do not help to understand the variations of geology at the valley scale. Consequently with such datasets, it is difficult to 30 develop general modeling methods that can be applied to fluvio-glacial deposits, or even to come up with simple conceptual models for large scale filling. In order to fill this gap of information, and provide a valley scale fluvio-glacial resistivity map, we conducted in January 2020 a large geophysical survey using tTEM (towed Transient Electromagnetic) system (Auken et al., 2019) in the upper Aare Valley, Switzerland. The tTEM-system provides a very detailed (both vertically and horizontally) resistivity model. The tTEM20AAR dataset covers a section of the valley of approximately 1-2 km width and 16 km long. The 35 fields were mapped with a line spacing of 20 meters, resulting in about 1500 hectares of covered land (see Fig. 1). The raw tTEM data were processed to suppress and removed noisy data parts, and then inverted to a resistivity model using specially constrained inversion algorithm (Viezzoli et al., 2008). The resulting resistivity model consists of 57'000 1D models of 30 layers. The depth of investigation varies, from 40 to 120 meters depth, primary driven by lithological/resistivity variations. The resulting resistivity model explains (fits) the recorded data well within the estimated data uncertainty. The resistivity model 40 reveals new and very interesting geophysical/geological structures of the subsurface at a fine resolution.
The tTEM20AAR data set can be used for several purposes. It can be used as a benchmark to test and compare geophysical inversion procedures for tTEM systems. For example, in this work the inversion was carried out in 1D for every vertical sounding but using constraints to account for lateral continuity. The full 3D inversion is much more demanding, but the data set could be used to compare it with the results published in this paper. Stochastic inversion (Mosegaard and Tarantola, 1995;45 Linde et al., 2017) using different methods or types of prior knowledge could also be applied on this data set and be compared with the published results. More generally, Quaternary formations are highly heterogeneous and constitute a challenge for geostatistical and uncertainty modeling (De Marsily et al., 2005). Sharing this data set will allow to test and compare various methods to interpolate the properties of the underground and construct models that can be used for various purposes. The integration of geophysical methods to constrain hydrogeological models is also a very important field of research (Binley et al., 50 2015). The tTEM20AAR data set may help testing the development of innovative methods for the construction of groundwater models. It is important to note in this perspective that the Upper Aare Valley has been extensively studied and a consequent amount of additional data are distributed by the Swiss authorities : a few thousand of logged boreholes, 46 stations of present or historic groundwater level measurements, a complete bedrock elevation model and Quaternary deposits thickness, as well as a complete hydrological and geological map. All these data cover the same site, and in combination with the geophysical 55 data presented in this paper, it forms an ideal test site for multiple data integration. Especially, we consider this site ideal for benchmarking new methods that integrates hydrological and geophysical data, or more generally that may link porosity, permeability or lithology with resistivity. Improvement in such methods may strongly improve hydrogeological modeling in such environments, subject to high local facies variations. Finally, from a more geological perspective, the tTEM20AAR data https://doi.org/10.5194/essd-2020-390   (Faraday's law of induction) for mapping the electrical conductivity (conductivity=1/resistivity) of the subsurface. A detailed description of the TEM principle can be found in (Christiansen et al.). The layout of the tTEM system is shown in Fig. 2. configuration: a high moment (HM) with a high inductive current of 30 A and a low moment with a lower inductive current of 5 A. Such configuration has the advantage of being able to resolve shallow targets with the low moment and its associated fast turn off time, and to reach higher penetration depth with the high moment. Both moments are stacked few hundreds times.
Detailed parameters are summarized in the table 1. The gate is the time interval in which the received amplitudes are averaged.

80
Due to the signal attenuation, further we get in the listening time, lower is the signal to noise ratio. In order to partially counterbalance this effect, we used a logarithmic increasing gate size related to listening time.
To ensure the data quality, the tTEM-instrumentation were calibrated prior to the survey at the Danish national TEM test site following the calibration procedure described by Foged et al. (2013). The two calibrated parameters are a time shift and an amplitude factor. The calibration was done with the ATV connected to the equipment in order to account for any shift caused 85 by it. Figure 3 shows the match between the test site reference response and the measured tTEM-response after calibration, which results in a fully acceptable match.

Field Site
The field site is the Upper Aare Valley, in central Switzerland (see Fig. 1  driving speed was between 10 to 20 km/h, depending of the terrain. Since the acquisition rate is time dependant, and not distance triggered, we also lowered the speed in noisier or less responsive areas in order to acquire a denser dataset. The spacing between the lines was approximately 20 meters. The average covered surface par day was 112 hectares, for a total of 1425 hectares. The voltage data from the receiver is measured continuously, and need to be cleaned of man-made noise and coupling. Data processing and inversion were carried out with the tTEM processing module in the Aarhus Workbench software. The objective of the processing of the tTEM-data is to remove any interference in the data from man-made installation (coupled data), suppress random noise by stacking, and finally discard the noisy late time data entering the background noise. Thus, we ensure that the resulting resistivity model represents geological structures of the subsurface without artifact from man-made 100 installation. Processing of the dB/dt data comprises of the following steps, which is a standard EM processing workflow : -Detection of coupled structures in the data and filtering.
-Averaging of raw data to suppress random noise. Raw data are averaged using a moving average filter with narrow time windows in early times and wider in the late times.
-Creation of vertical soundings every 2.5 s which corresponds approximately to a spacing of 10 m. The exact distance 105 can vary depending on driving speed.
-Automatic filtering of the averaged data for removal of late-time data points entering the background noise.
-Visual assessment of all dB/dt data and manual removal of coupled data not detected by the automatic filtering.
-Evaluation and adjustment of the data processing based on preliminary inversion results.
Furthermore GPS data are lag-corrected to geographical positioned data/models at center between transmitter and receiver 110 coils. The data uncertainty consists of a minimum of 3% as uniform data standard deviation (STD) plus the STD calculated from the data stacking. Averaged data resulting with STD over 30% are discarded from inversion.

Inversion
The electrical resistivities of the underground are then estimated using a series of 3D constrained 1D-inversions. The 1D inversion is based on the AarhusInv code (Auken et al., 2015;Kirkegaard et al., 2015). This code is an implementation of a 1D 115 non-linear damped least-squares solution, with a modeled transfer function for the TEM instrumentation. This function takes into account the transmitter waveform, the instrument low pass filters, the receiver bandwidth, the system geometry, the gate widths and the instrument front gate. However, in such an standalone 1D inversion, each model is totally independent of the neighboring ones. To account for the lateral continuity expected in geological environments, the spatially constrained inversion (SCI) (Viezzoli et al., 2008) method was used. It applies 3D constraints to 1D inversion models both along and across the 120 mapping lines, with a weight that is decreasing with distance. All the inversions were carried out with the Aarhus Workbench software.
The SCI inversion can be used with two different schemes of regularization: smooth or sharp. The smooth scheme tends to minimize abrupt changes in resistivity, in the vertical and horizontal directions. On the other hand, the sharp regularization  scheme tends to minimize the number of resistivity changes, but will consequently result in more abrupt resistivity transitions 125 and a potential more blocky model appearance. Both regularizations were used, and are included in the output data.
For each resistivity model, we estimate the depth of investigation (DOI) using a method based on the Jacobian Sensitivity matrix (Christiansen and Auken, 2012). This method has the advantage of taking into account the full transfer function, including system geometry, data uncertainty and the resistivity model. Two DOI thresholds values in the sensitivity matrix were used to provide the reported DOI-standard, and the DOI-conservative values. As a guideline, the resistivity structures above the DOI 130 conservative value are strongly data driven, while resistivity structures below the DOI standard value are weekly represented in the data. Normally one would blank the resistivity models below DOI standard value.
Inversion setup for the smooth and sharp inversions are summarized in the table 2.

Data Validation
where d obs is the observed data, d f wr is the forward data, σ d is the uncertainty of the observed data and N is the total number of data point. A data residual below 1 indicates that our final model response is within one standard deviation of the data, when  Figure 5b shows, a single data curve (error bars) and the forward response (line) from the resistivity model in Fig. 5a. The associated data-misfit for this model is 0.27. The data misfit for the all smooth inversion models is plotted in Fig. 6. As seen in Fig. 6, the data misfit is in general well below one and fully acceptable. 95% of the data is within 1 standard deviation, with a global misfit average of 0.65 and 0.52 respectively for the sharp and smooth models. A manual inspection of the high data misfit models revealed that they are all associated to highly resistive models, and/or are close to man made electromagnetic noise such as roads, fences, or train tracks. A good example is the extreme south of the acquisition, that is one of the most resistive areas. This situation logically leads to a lower signal to noise ratio, and due to the spatial constraints of the inversion, it will consequently leads to an higher data misfit. However, they are usually restricted to only a few local data-points, and the models are similar to neighbouring ones that has acceptable misfit. We therefore decided 150 to keep them in the dataset.
Finally, users of the data should be aware that the footprint of the equipment is at least 9m at the surface (size of the equipment) and is increasing with depth and wave diffusion. Consequently, a sharp vertical transition in the geology for example, will tend to appear oblique in the resistivity data due to this effect. The resistivity models proposed here are only the one that fits the best our data. However, the processed but uninverted data are provided, in order for others to perform inversions  After the data processing and the inversion, the uninverted processed data, the resistivity models and the associated forward responses from the smooth and sharp inversions had been produced. These data (Neven et al., 2020) are provided in column based ASCII files. Each file structure is outlined in the following subsections. The can be freely downloaded 160 on the Zenodo platform (https://doi.org/10.5281/ZENODO.4269887). The forward code used for the tTEM data presented here is the AarhusInv developped by the Aarhus University Hydrogeophysics group (Auken et al., 2015;Kirkegaard et al., 2015). The AarhusInv code is free to use for research purpose, and can then be combined with any other inversion technique (https://hgg.au.dk/software/aarhusinv/).

Processed data file
with unc down and unc up being the absolute lower and upper uncertainties, DATA the processed z-component dB/dt data value and DATASTD the relative uncertainty. The structure is outlined in the following

Synthetic response file
The Forward_Data_Sharp.dat and Forward_Data_Smooth.dat files contains the forward responses of the sharp and smooth resistivity models. The structure of the forward data files is the same as the Processed_Data.dat file except that the forward responses does not have associated data uncertainties. Detailed file structure is provided in table 5. The forward 185 response was produced with the TEM 1D AarhusInv code.

Conclusions
In this work, we presented the acquisition of more than 1400 hectares of tTEM measurement in the Upper Aare Valley, Switzerland. The dataset presented here is unique in terms of coverage, and combines it with a high spatial resolution. It reveals the presence of various so far unexplained geological structures, in the shallow and in the deeper depths. This data set could 190 be used for a wide range of different applications, including methodological developments as well as a better understanding of fluvio-glacial aquifers that represent a major source of drinking water supply in many places in the world.