Monitoring soil salinity using time-lapse electromagnetic conductivity imaging

Maria Catarina Paz , Mohammad Farzamian , Ana Marta Paz , Nádia Luísa Castanheira , Maria Conceição Gonçalves , Fernando Monteiro Santos 1 1 Instituto Dom Luiz, Faculdade de Ciências da Universidade de Lisboa, Campo Grande, Edifício C1, Piso 1, 1749-016 Lisboa, 5 Portugal 2 CIQuiBio, Barreiro School of Technology, Polytechnic Institute of Setúbal, Rua Américo da Silva Marinho, 2839-001 Lavradio, Portugal 3 Instituto Nacional de Investigação Agrária e Veterinária, Avenida da República, Quinta do Marquês (edifício sede), 2780157 Oeiras, Portugal 10


Introduction 35
Lezíria Grande de Vila Franca de Xira (hereafter called Lezíria de Vila Franca) is an important agricultural system of alluvial origin located by the estuary of river Tejo, northeast of Lisbon, Portugal (Fig. 1), where soil faces risk of salinization due to the marine origin of part of the sediments, tidal influence of the estuary, irrigation practices, and projected evolution of future climate with increasing temperature and decreasing precipitation. Traditional soil salinity investigations have been conducted in the study area using the electrical conductivity of a saturated soil paste extract (ECe, dS m −1 ) as a proxy for soil salinity. 40 However, they were limited to few boreholes and involved soil sampling, which restricted the analysis to point information, often lacking representativeness at the field scale. In addition, borehole drilling is invasive and not feasible to conduct over large areas, given the large number of boreholes that needs to be made.
Electromagnetic induction (EMI) is widely used as a non-invasive and cost-effective solution to map soil properties over large areas. EMI measures the apparent electrical conductivity of the soil (ECa, dS m −1 ), which is primarily a function of soil salinity, 45 soil texture, water content, and cation exchange capacity; however, in a saline soil, soil salinity is generally the dominant factor responsible for the spatiotemporal variability of soil ECa. EMI surveys have been successfully used in conjunction with soil sampling to assess soil salinity through location-specific calibration between measured ECa and soil salinity (e.g. Triantafilis et al., 2000;2001;Corwin and Lesch, 2005;Bouksila et al., 2012). However, the ability of this method for mapping soil salinity distribution with depth is limited. This is because EMI measures ECa, a depth-weighted average conductivity measurement, 50 which does not represent the soil electrical conductivity (σ, mS m −1 ) with depth. More recently, a state-of-the-art approach called electromagnetic conductivity imaging (EMCI) has permitted to obtain σ from the inversion of multi-height and/or multi sensor ECa data (Monteiro Santos, 2004;Dafflon et al., 2013;von Hebel et al., 2014;Farzamian et al., 2015;Shanahan et al., 2015;Jadoon et al, 2015;Moghadas et al., 2017). When comparing σ with the soil properties sampled in boreholes, such as ECe, soil water content, pH, among others, a calibration process is developed through a regression between σ and the soil 55 properties. This way, EMCI can be converted to a map of the soil properties which show strong correlation with σ. This methodology has been applied in Lezíria de Vila Franca to study soil salinity risk (Farzamian et al., 2019;Paz et al., 2019b), and salinity and sodicity risk (Paz et al., 2019a) in which EMCI has been converted to ECe and sodium adsorption ratio. When repeated over a period of time, EMCI of a study area is called time-lapse EMCI, and can be used to monitor the dynamics of soil salinity and other soil properties. Time-lapse EMCI has been successfully used to monitor soil water content (Huang et 60 al., 2017;2018;Moghadas et al., 2017) although, to our knowledge, its potential for monitoring soil salinity has not been previously investigated.
This study aims to evaluate the potential of time-lapse EMCI and a previously developed regional calibration to predict the spatiotemporal variability of soil salinity, and to monitor and evaluate soil salinity dynamics in the study area. For this purpose, EMI measurements and soil sampling were carried out between May 2017 and October 2018 at four locations with different 65 salinity levels across the study area. EMI measurements were performed with a single-coil instrument (EM38), collecting ECa data in the horizontal and vertical orientations and at two heights, and then inverted to obtain EMCI, which provides a vertical distribution of σ. Finally, σ was converted to ECe through the previously developed regional calibration. Soil samples were collected along the EMI transects, and used for laboratory determination of ECe. These data were used as an independent test set to evaluate the ability of the regional calibration to predict the spatiotemporal variability of soil salinity, and to generate 70 soil salinity maps for each date of data collection.

Study area
The study was carried out in Lezíria de Vila Franca, a peninsula of alluvial origin surrounded by the rivers Tejo and Sorraia, and the Tejo estuary, located 10 km northeast of Lisbon, Portugal, as shown in Fig. 1. Soils in this region have fine to very 75 fine texture and are classified as Fluvisols in the northern part and as Solonchaks in the southern part, according to the Harmonized World Soil Database (Fischer et al., 2012). Climate is temperate with hot and dry summers, according to the Köppen classification. Daily measurements of precipitation, mean temperature and reference evapotranspiration recorded during the study period at the meteorological station represented by the blue circle in Fig. 1b, are shown in Fig. 2. Land use in this area (of about 130 km 2 ) is constituted by irrigated annual crops in the northern part and mainly by rainfed pastures in the 80 southern part. Irrigation is assured by an infrastructure that covers most of the area, collecting surface water at the confluence of the two rivers. The irrigation water has low salinity with electrical conductivity typically below 0.5 dS m -1 and sodium https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License. adsorption ratio below 1 (mmolc L −1 ) 0.5 . The area exhibits a north-south soil salinity gradient which influences the distribution of land use types and which is probably due to the regional distribution of the marine fraction of sediments and to the saline influence of the estuary on groundwater in the southern part. 85 Four locations were chosen in the study area, as presented in Fig. 1b, with numbers 1 to 4. Locations 1, 2, and 3 are cultivated with annual rotations of irrigated herbaceous crops in spring and annual ryegrass (Lolium multiflorum) in the autumn, with ploughing usually once a year. During the study years (2017 and 2018), the spring crop at location 1 was tomato drip irrigated, and at locations 2 and 3 was maize irrigated by centre pivots. Location 4 is a rainfed spontaneous pasture that hasn't been ploughed at least in the last ten years. During the study period, location 1 was irrigated from 12 April to 23 July 2017 and from 90 30 May to 23 September 2018; location 2 was irrigated from 17 June to 11 October 2017 and from 24 May to 22 September 2018; and location 3 was irrigated from 17 May to 10 September 2017 and from 06 June to 17 September 2018.
Groundwater level is shallow, as expected in an estuarine environment, and has saline characteristics. In the southern part of the study area, closer to the estuary, the depth and salinity of groundwater are influenced by tidal variation.

Electromagnetic induction data acquisition and inversion
EMI data was acquired using the EM38 instrument (Geonics Ltd, Mississauga, Canada) on five dates at locations 1 and 4, and on six dates at locations 2 and 3, during the period of May 2017 to October 2018. Measurements on the two first dates were continuously acquired along 100 m transects using a GPS (Rikaline 6010) for registration of the position. Subsequent EMI 105 measurements were acquired at positions 1 m apart along 20 m transects (Fig. 1c), overlapping the medium section of the 100 m transects. ECa was collected at two heights from the soil surface (0.15 and 0.4 m) in the horizontal and vertical dipole orientations. Inversion of ECa data to obtain σ was carried out using a 1-D laterally constrained inversion algorithm (Monteiro Santos et al., 2011). The ECa responses of the model were calculated through forward modelling based on the full solution of the Maxwell equations (Kaufman and Keller, 1983). The subsurface model used in the inversion process consisted of a set of 110 1-D models distributed according to the position of the ECa measurements. The subsurface model at each measurement position was constrained by the neighbouring models, allowing the use of the algorithm in regions characterized by high conductivity contrast. An Occam regularization (De Groot-Hedlin and Constable, 1990) based approach was used to invert the ECa data.
To run the algorithm, several parameters are selected, such as the type of inversion algorithm, the number of iterations, and the smoothing factor (λ) that controls the roughness of the model. The optimal inversion parameters for the present conditions 115 were obtained in previous studies for the study area (Farzamian et al., 2019).

Soil sampling and laboratory analysis
Soil samples were collected at the same time of EMI surveys along the transects, as shown in content and ECe. In the laboratory, water content was obtained using the gravimetric method, and then converted to volumetric water content (θm 3 m −3 ) after bulk density determination. ECe was measured with a conductivity meter (WTW 1C20-0211 inoLab) in the extract collected from the soil saturation paste. In this study, the soil is classified according to its ECe level as non-saline (ECe<2 dS m −1 ), slightly-saline (2-4 dS m −1 ), moderately-saline (4-8 dS m −1 ), highly-saline (8-16 dS m −1 ), and severely saline (>16 dS m −1 ), according to the terminology proposed by Barrett-Lennard et al. (2008). 125 https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License.

Prediction of ECe from time-lapse EMCI
A regional calibration to predict ECe from σ was previously developed for the study area resulting in the linear equation ECe = 0.03σ -1.05 (Farzamian et al., 2019). This calibration was termed "regional" because the equation was obtained using all ECe and σ data collected at four locations in the study area. Farzamian et al. (2019) tested the regional and location-specific calibrations, verifying that they have comparable prediction ability. However, the regional calibration can be used at any new 130 location in the study area, within the range of measured ECe, which makes it highly suitable for mapping and monitoring salinity in the study area. The regional calibration was based on data collected during May and June 2017 and was validated using a leave-one-out-cross-validation method with good results (RMSE = 2.54 dS m −1 ). The detailed calibration and crossvalidation procedures are described in Farzamian et al. (2019).
In the present study, the regional calibration was used to predict ECe from time-lapse EMCI. The predicted ECe and ECe 135 measured from soil samples, collected from July 2017 to October 2018, were used to validate the regional calibration as an independent test set. Its prediction ability was evaluated by calculating the root mean square error (RMSE), the coefficient of determination (R 2 ) between the measured and predicted ECe, the Lin's concordance correlation coefficient (CCC), and the mean error (ME). The RMSE is the square root of the mean of the squared differences between the measured and predicted ECe, indicating how concentrated the data is around the linear regression. In this study we used two degrees of freedom for a 140 more robust calculation of RMSE. The coefficient of determination (R 2 ) indicates how well the predicted ECe approximate the measured ECe. When this is 1, it means the predictions coincide with the measurements. Lin's CCC measures the agreement between the measured and predicted ECe evaluating how close the linear regression is to the 1:1 relationship and ranges from −1 to 1, with perfect agreement at 1 (Lin, 1989). ME is the mean of all differences between the measured and predicted ECe and evaluates whether the linear regression consistently over-and underestimates the predicted ECe. Therefore, the prediction 145 is more precise and less biased when the RMSE and the ME are closer to zero.
https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License. Figure 3 shows the variation of θ and ECe with time at the sampling site located in the middle of each transect (Fig. 1c), at locations 1 to 4. At location 1, θ increases with depth and the lower subsoil (1.2-1.5 m) is permanently saturated within the 150 study period. In the more superficial layers until 0.9 m depth, the influence of rainfall, evapotranspiration, and irrigation is noticeable. For instance, in the topsoil, θ peaks in January 2018 and lowers during the dry seasons, because drip irrigation during the dry seasons has a localized effect and there is high water uptake by the crop. At location 2, unlike the other locations, the lower subsoil is unsaturated. The influence of rainfall, evapotranspiration and irrigation is also noticeable. At locations 3 and 4, θ also increases with depth and the intermediate and lower subsoil layers are permanently saturated. 155

Temporal variation of measured θ and ECe
Regarding ECe, at location 1 the values observed are always below 1 dS m −1 , except for the topsoil in September and October 2018, which is probably due to fertigation practises during the irrigation period. At location 2, ECe generally increases with depth. All layers show a peak in June and July 2018, probably due to fertigation practises. At location 3, ECe reaches higher levels than at the previous locations, exceeding 4 dS m −1 , which is the generally accepted threshold for the classification of   9-1.2 m), and lower subsoil (1.2-1.5 m), measured at the sampling site located in the middle of each transect, at locations 1 to 4, during the study period.
https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License. Figure 4 shows ECe predicted with the regional calibration versus the measured ECe and the 1:1 line, with points identified in terms of date of measurement (Fig. 4a) and depth of measurement (Fig. 4b). Prediction of ECe with the regional calibration 170 using data collected from July 2017 to October 2018 resulted in a RMSE of 3.22 dS m −1 and R 2 of 0.89, which indicates satisfactory prediction ability, given the large range of ECe (52.35 dS m −1 ). The high global Lin's CCC of 0.93 shows accord between measured and predicted ECe. The ME is −1.30 dS m−1, indicating that the regional calibration globally overestimates ECe. Figure 4a and Fig. 4b show that the points are generally scattered around the 1:1 line and it is not possible to identify variations depending on the date or depth of the measurement. In order to analyze the prediction ability at each location,  ECe is generally overestimated. At location 2, ECe is more overestimated in deeper soil layers (Fig. 4d) which is likely due to a previously identified influence of clay content that consistently increases with depth at this location, while it is rather uniform or declines with depth at the other locations (Farzamian et al., 2019). 180

Prediction of ECe using the regional calibration
The validation procedure used in this study gives lower prediction ability for the regional calibration than the previously obtained with the leave-one-out-cross-validation (see section 2.4). This can be justified because the test set is completely independent from the dataset used to develop the calibration. Furthermore, this test set is composed of measurements collected over a wider period of time (18 months). During this period, soil properties, which are also known to influence σ, such as θ, change (as shown in Fig. 3), which introduces larger variability in the measurements. However, and given the large range of 185 ECe (52.35 dS m −1 ), a RMSE of 3.22 dS m −1 is acceptable for this type of non-invasive and indirect method. The regional calibration could be further developed by including measurements taken over a longer period of time in the calibration process, in order to include a wider range of variation of soil properties.   Figure 5 shows the soil salinity maps (ECe predicted using the regional calibration) at locations 1 to 4 for each date of the EMI surveys, categorized into 6 salinity classes, ranging from non-saline to severely-saline. The measured ECe and the groundwater level at the sampling site located in the middle of each EMI transect are also shown.  The salinity maps for location 1 show that the soil is generally non-saline, with slightly saline zones in all dates except for 205

Spatiotemporal mapping of soil salinity from time-lapse EMCI 195
October 2018. These saline zones occur in the top soil layers until 0.9 m depth (topsoil, subsurface and upper subsoil), and represent an overestimation of the soil salinity when compared to the measured ECe of the sampling point (which is invariably non-saline). This overestimation tendency is in agreement with Fig. 4d, where the very low range of spatiotemporal variations of soil salinity at this location can also be observed. In such conditions, other soil properties, such as θ, dominate the small variations of σ, and therefore the ability to predict salinity from σ at this location was reduced. Our previous studies with both 210 location-specific and regional calibrations tested at this location showed similar results (Farzamian et al., 2019).
At location 2 the salinity maps show an increase of salinity with depth from non-saline at the topsoil to highly-saline in the lower subsoil, with exception of July 2018, where the entire soil profile is moderately saline. The increase of soil salinity in upper soil layers in July 2018 can be attributed to fertigation practices for the maize cultivation that introduced salts into the soil profile. The salinity maps also show the overestimation of salinity occurring mainly at deeper soil layers, which agrees 215 with the results presented in Fig. 4d and discussed in section 4.2.
At location 3 soil salinity is well predicted in May 2017 but tends to be slightly overestimated in the remaining dates, especially in July 2018. The salinity maps show that salinity increases with depth reaching severely-saline in May 2017 and October 2017. This can be due to the influence of the saline groundwater (as seen in Fig. 3, the intermediate and lower subsoil layers are permanently saturated). The groundwater level is above 1.5 m in January 2018, although the salinity of the deeper 220 soil layers (>0.9 m) decreases compared to May and October 2017, which could be due to washing of the profile by rainfall.
At location 4 the trend of increasing salinity with depth is accurate in all dates, but it tends to be slightly underestimated. The salinity maps show that salinity increases from non-saline in topsoil to severely-saline in lower subsoil. This is probably related to the saline groundwater level above 1.5 m. During the dry period of the year, salinity of the lower subsoil reaches the highest values (June 2018). 225 Comparison of the salinity maps between locations confirms the previously known north-south soil salinity spatial gradient of the study area, that is, from location 1 to 4, soil salinity generally increases. Soil salinity dynamics at each location reveals fluctuations in time related to the input of salts and water either through irrigation, precipitation or groundwater level and https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License.
salinity. Location 1 tends to have non-saline characteristics, which can be attributed to good quality irrigation water and to the fact that this location is far from the estuary, making it less prone to the presence of saline groundwater. At locations 2 and 3, 230 the salinity maps show an increase of soil salinity in the upper layers during the dry season (when irrigation occurs), which decreases in the following months with increased rainfall (Fig. 2). At the rainfed location 4, it is also visible an increment of salinity along the entire profile during the dry season. This is likely due to the influence of the saline groundwater and capillary rise along the profile.

Conclusions 235
In this study, EMI and soil sampling data collected between May 2017 and October 2018 were used, together with a previously developed regional calibration, to predict the spatiotemporal variability of soil salinity. This procedure allowed to further validate the regional calibration with an independent test set. This validation resulted in lower prediction ability than that previously resulting from cross-validation, not only because the test set was independent, but also because it was collected over a wider period of time, during which the variation of soil properties is larger. The validation used in this study resulted in 240 a RMSE of 3.22 dS m −1 , which is acceptable given the large range of ECe (52.35 dS m −1 ). As a result, the regional calibration approach still stands as an expeditious method to predict soil salinity in the study area over time. The regional calibration could be further developed by studying new locations across the study area in order to include a wider range of variation of soil properties. Also, a longer period of observation could further improve the regional calibration. Furthermore, the influence of static soil properties (i.e., that do not vary in time), such as clay content, could be tackled with the use of maps of the variation 245 of soil salinity between two consecutive dates, which allows removing the static effect in the EMCIs.
Relatively to the inversion process, in the absence of a time-lapse inversion algorithm, ECa data was inverted independently.
This method can distort the inversion results, since the reference model and a priori information are not considered. Further research involves time-lapse inversion algorithms that are being developed to invert data collected with EMI sensors, which can generate EMCIs of higher precision. 250 The methodology used in this study allowed the creation of soil salinity maps displaying the spatiotemporal patterns of soil salinity at four locations in the study area. The salinity maps reveal fluctuations in time related to the input of salts and water https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License. either through irrigation, precipitation or groundwater level and salinity. In a regional perspective, soil salinity dynamics in the study area may be explained by a combination of spatial distribution of the marine fraction of soil, with irrigation practices in the study area and saline groundwater in the southern part. Continuous monitoring of salinity in the study area, along with 255 detailed data collection about irrigation, precipitation, evapotranspiration, leaching, groundwater flow, and tides, can be helpful to further study soil salinity dynamics.
Time-lapse EMCI has proven to be a valid methodology for evaluating risk of soil salinization, and can further support the evaluation and adoption of proper agricultural management strategies, especially in irrigated areas, where continuous monitoring of soil salinity dynamics is required. 260 https://doi.org/10.5194/soil-2019-99 Preprint. Discussion started: 5 February 2020 c Author(s) 2020. CC BY 4.0 License.