Estimation of soil water content using electromagnetic induction sensors under different land uses

The complex nature of podzolic soils makes investigating their subsurface challenging. Near-surface geophysical techniques, like electromagnetic induction (EMI), offer significant assistance in studying podzolic soils. Multi-coil (MC-EMI) and multi-frequency (MF-EMI) sensors were selected to maximize soil water content (SWC) prediction in this study. The objectives were to (i) compare apparent electrical conductivity (ECa) measurements from the MC and MF-EMI sensors under different land use conditions, (ii) investigate the spatial variation of ECa, SWC, texture, soil organic matter (SOM), and bulk density (BD) under different land use conditions, and (iii) use statistical and geostatistical analysis to evaluate the effectiveness of ECa measurements in characterizing SWC under different land use conditions, considering the texture, SOM, and BD contents. The study found that MC-EMI had statistically significant relations (p-value < 0.05) with SWC relative to the MF-EMI. Multiple linear regression (MLR) models were also shown to be more effective in representing SWC variations (higher coefficient of determination and lower root mean square error) than simple linear regression models. MC-EMI sensor provided better SWC predictions compared to the MF-EMI sensor, possibly due to larger sampling depths differences between time domain reflectometry measured SWC (SWCTDR) and MF-EMI sensor than those between SWCTDR and MC-EMI sensor. Lastly, cokriging of measured SWC was found to offer more accurate maps than cokriging of predicted SWC obtained from MLR across different land use conditions. The study has shown that EMI may not be highly effective for shallow depths, and ECa can be affected by various soil properties, making it difficult to extrapolate other parameters. However, EMI still shows promise as a reliable method for predicting SWC in boreal podzolic soils. Research into EMI’s usefulness for this purpose has yielded promising results, as indicated in this study. Further investigation is needed to fully harness the potential of this promising technique.


Introduction
In order to estimate soil water content (SWC) using electromagnetic induction (EMI) sensors under different land uses, it is crucial to understand how soil properties respond to different agricultural management practices in optimizing agricultural operations and inputs since soil properties vary at diverse spatial scales mainly due to factors such as soil management, and land use (Behera et al 2018). Understanding spatial variations can be achieved by applying a suite of advanced information, data analysis, and communication technologies such as remote sensing (geophysical techniques), geographic information systems (GIS), global positioning systems (GPS), and artificial intelligence (Sishodia et al 2020). Conventional methods of sampling and analyzing soil properties such as SWC, texture, soil organic matter (SOM), and bulk density (BD) to understand spatial variability tend to be destructive, time consuming and expensive, and provide only point information. Consequently, these methods can hinder timely decision-making due to the sparse representation of the desired soil property's spatial variation (Allred et al 2008). New geophysical techniques, such as EMI provide a rapid, The study site was carefully chosen for its usefulness as a research test field for research activities in the province. Research activities carried out on the test field involve measuring the levels of nitrogen fertilizer, in order to determine how to keep nitrogen in the soil for longer periods of time while minimizing gaseous and leaching losses. By successfully accomplishing this, local farmers will be able to identify which crops can be grown well under which management condition in the province. The soil at the study site was classified as a reddish-brown to brown Podzol developed on a gravelly sandy fluvial deposit with >100 cm depth to bedrock and a 2%-5% slope (Croquet 2016). The area of the agricultural land considered was 924 m 2 . This area was on fallow from 2001 to 2018. In 2017 and 2018, the land was sprayed with RoundUp® and ploughed three times to control the canary reed grass. In 2019, large rocks were cleared with an excavator and the land was treated with a combination of RoundUp® and Aragon® tank mix, before ploughing, disk harrowing, rock raking, and rock picking. Five silage corn-based rotations were initiated; corn (2019), followed by corn, faba bean, wheat, canola, or an oat/pea mix in 2020. In 2021, the whole area was seeded to silage corn. The field road is adjacent to the agricultural land and 240 m 2 in area, serving as access for yearly crop rotation of corn, with canola, faba bean, and wheat. The field road was graded to facilitate access to other parts of the study area for measurements and experienced traffic from field vehicles and personnel. In contrast, the agricultural land also witnessed higher human activitydue to various research activities conducted in the area. The field road was included in this study because this land use condition is an important factor to consider when conducting studies on SWC since it offers the opportunity to continuously monitor changes in SWC and compares results between different fields. Additionally, the field road has the potential to act as a conduit for water, allowing it to move through the landscape quickly and efficiently, potentially leading to increased SWC levels in the surrounding areas (agricultural land and natural forest). Moreover, this land use condition can provide an indication of the degree of human influence on the landscape, enabling further study of the effects of land use change on SWC levels. A 50 m 2 area was selected from a recently cleared natural forest to compare its soil property data with the other land use conditions of interest (agricultural land and field road) (figure 1). The natural forest was cleared in 2018 and the area on the cleared natural forest was selected rather than a standing forest due to the specifics of the MC-EMI sensor. Due to the thick forest with many obstacles, it was not possible to survey the woody area at constant phase which would result inconsistent data. Hence, the open space of a recently cleared natural forest was the ideal option for collecting quality data. This choice, while limiting in some ways, allowed the full capabilities of the sensor to be utilized, giving the researchers a more complete picture of the variability of the studied soil properties.
The area received a total rainfall of 409 mm and had an average a mean temperature of 8.18°C (figure 2) based on three months of reported data (1 Sep-30 Nov. 2021) from the nearby Deer Lake weather station A (http://climate.weather.gc.ca.). The daily total precipitation and temperature (minimum, maximum and mean) data obtained from Deer Lake weather station A were used to calculate the daily potential evapotranspiration PET. PET was calculated using the modified Hargreaves-Samani equation (equation (1)) for the Deer Lake weather station A (Perera 2021 Where T max is the daily maximum temperature, T min is the daily minimum temperature, T mean is the daily mean temperature, R a is the extraterrestrial radiation (MJ/m 2 /day), and λ is the latent heat of vaporization = 2.45 (MJ/kg).

Soil sampling and analysis
Before the commencement of the EC a study, a simple random sampling technique was employed to collect data from the three land use conditions. Nine composite samples were taken from each land use condition in order to obtain a representative sample of the entire population. This cost-effective and reliable sampling technique allows researchers to obtain an unbiased and accurate representation of the population. Soil samples were collected at a depth of 0-20 cm (figure 3) and analysed for a variety of properties, including texture (by hydrometer method), SOM (by loss on ignition), SWC and BD for the different land use conditions using protocols from Carter and Gregorich (2007). The SWC was measured using a time domain reflectometry (TDR), Fieldscout TDR 350, by vertically inserting 0-20 cm probes at each sampling time. At each sampling location, three TDR data readings were taken at nine separate points in order to calculate an average for each sampling location. This approach ensures that the collected data is reliable and unbiased, allowing for meaningful conclusions to be drawn from the study.

Electromagnetic induction survey
The MC-EMI sensor namely, CMD-mini-explorer (GF instruments, Brno, Czech Republic) and the MF-EMI sensor, GEM-2 (Geophex Ltd, Raleigh, USA) were used to map soil spatial and temporal variability of EC a . The CMD Mini-explorer operates at a frequency of 30 kHz. This device consists of a probe and a handheld control unit that communicate with each other via Bluetooth. It features a single transmitter coil alongside three receiver coils, with inter-coil spacing of 32 cm, 71 cm, and 118 cm. This EMI sensor operates in modes, VCP and HCP coil configuration, in which the former can sense integral depths up to 25, 50 and 90 cm, and the latter up to 50, 100, and 180 cm (Altdorff et al 2020). Additionally, the device is designed with a temperature stability of ±1 mS/m per 10°C change in temperature and is suitable for outdoor use between −10°C and +50°C temperatures (GF Instruments, Brno, Czech Republic).
The GEM-2 is a lightweight, digital MF-EMI sensor with a single transmitter and a receiver coil, as demonstrated by Tang et al (2018) and Won et al (1996). Its components include a ski that encloses all sensing elements, an electronics enclosure, an IPaq display assistant, and an external GPS connector. Its frequency range is 0.3 kHz to 90 kHz (Sadatcharam 2019). The sensor has a factory file for frequencies which can be modified to the user's desired set (Geophex Ltd, Raleigh, USA). The inter-coil spacing between the transmitter and receiver coil is 166 cm, and the bucking coil is located at approximately 1 m from the transmitter coil, cutting off the primary field from the receiver sensor (Sadatcharam 2019). Selecting too many frequencies will lower the resolution, as Sadatcharam (2019) and Altdorff et al (2020) have demonstrated. The magnitude of the selected operating frequency is inversely proportional to the depth of investigation (DOI); so, the higher the selected operating frequency, the shallower the DOI. However, a higher frequency results in a higher resolution. The GEM-2 can operate in both HCP and VCP coil configurations. In comparison, the MF sensor can operate at lower frequencies and has a higher DOI, while the MC sensor is better suited for detecting conductivity variation in the shallow soil depths (Sadatcharam 2019).
EMI surveys were conducted using both sensors across the different land use conditions on 20 Oct., and 11 Nov. 2021. Both sensors were operated in the horizontal and vertical coplanar coil orientations as done by previous researchers in the same area (Altdorff et al 2018, Badewa et al 2018, Sadatcharam 2019. All three coils were simultaneously used when employing the MC-EMI, while four different frequencies (2.8, 18.3, 38.3, and 80.0 kHz) were manually set to simultaneously measure soil EC a when using the MF-EMI sensor. The theoretical DOI for the MC-EMI sensor at three coils spacing (0.32, 0.71, and 1.18 m) was 25, 50, and 90 cm and 50, 100, and 180 cm for VCP and HCP coil orientations, respectively. The DOI for the MF-EMI sensor at coil separation of 1.66 m are 125 and 250 cm for VCP and HCP coil orientations, respectively. The selected frequencies have been reported to be suitable for shallow subsurface investigations (Won et al 1996, Wang et al 2022. The EMI surveys were carried out in a bi-directional order over the three land use conditions while maintaining a line spacing of approximately 1 m. A global positioning system (GPS) was attached to the EMI sensors to enable the collection of georeferenced data to produce georeferenced maps. Before each survey, the instrument was warmed up for at least 30 min for temperature adaptation to prevent data drift and ensure high-quality data, in accordance with the protocols developed by Robinson et al (2004).
Soil temperature was measured at a depth of 0-20 cm depth for all three land use conditions using the Bi-Metal Dial soil temperature probe (HBE International Inc.). The 'Sheets and Hendrickx temperature correction model', was used to correct soil temperature (equation (2)). where EC t is the EC a data collected at measured soil temperature (°C), EC 25 is the temperature corrected EC a at 25°C, and t is the soil temperature. Negative value observations were considered noise and subsequently eliminated.

Statistical analysis
Descriptive statistics and analysis of variance (ANOVA) of measured EC a , BD, SWC, SOM, and texture were carried out under different land use conditions and management practices. One-way ANOVA was employed to determine the distinctions in soil EC a and SWC among the various land use conditions. To evaluate the normality of residuals, the Kolmogorov-Smirnov test was utilized, and the Fisher LSD test at 10% significance was used to compare the least square means. The combined use of these two tests allowed for an accurate determination of the differences in soil EC a and SWC levels between the different land use conditions. Pearson's correlation analysis was employed in order to identify the strength of the correlation between EC a and other soil properties in the different land use scenarios. This analysis was conducted to gain a better understanding of the relationship between these soil properties and the potential effects of land use on them. The results of this analysis were integral in helping researchers to assess the current state of soil properties under different land use conditions, as well as to make predictions about the potential impact of future land use. Furthermore, this analysis allowed us to identify any potential relationships between EC a and other soil properties that could inform decisions about land management. Low clay content and the rocky nature of the study area impeded the relation between EC a and the other soil properties; therefore, a 90% confidence interval was used to assess the precision of estimated statistics in this study. A 90% confidence interval has also been used by Badewa et al (2018) to examine the strength of the EC a −SWC relationship. To determine the sole attributes which were the most influential predictors of EC a , multiple linear regression (MLR) was performed using backward elimination of less influencing variables. The MLR equations obtained from the first series of surveys (first day) were used to predict SWC for the second series of surveys (second day) under each land use condition. The measured SWC (using TDR) on the second day and the predicted SWC (by employing MLR models) were compared using a 1:1 plot and root mean square error (RMSE). The slope and the intercept of the prediction lines under each land use condition were compared statistically with those of the 1:1 line. The variance inflation factor (VIF) obtained from each regression model was used to test for multicollinearity. The VIF measures the amount of variance of the estimated regression coefficient that is magnified if the independent variables are correlated. VIF is calculated as A VIF = 1 indicates that multi-collinearity does not exist between predictors, whereas VIFs ranging between 1 and 5 indicate moderately correlated variables. On the other hand, a VIF greater than 5 indicates multicollinearity among the predictors (Shrestha 2022). All statistical analyses were performed with Minitab 17 (Minitab 17 Statistical Software ).

Geostatistical analysis
Cokriging with EC a as a covariate was performed on the measured SWC and on the predicted SWC data from the most accurate generated regression models. This was done to determine the amount of variation in accuracies between the maps obtained from the measured SWC and the predicted SWC in each land use condition. Cokriging uses multiple datasets to investigate graphs of cross-correlation and autocorrelation (equation (4) where u and v are the target and covariate variables, respectively. The two variates u and v are cross-correlated, and the covariate contributes to the estimation of the target variate; μ is the Lagrange coefficient; n is the number of sampling points used in estimation; γ(xi, xj) is the value of variogram corresponding to a vector with origin in xi and extremity in xj; λi is the weight associated with the data. Variograms describing the spatial dependence of a spatially random field were used to analyze the spatial structure of SWC. Several variogram models (linear, exponential, circular, gaussian, spherical, and power models) were considered when performing cokriging for creating SWC maps. Each variogram was characterized by three parameters: range, sill, and nugget. The nugget/sill ratios were used to characterize the spatial dependence of observations. Spatial dependence was characterized as strong (below 25%), moderate (25% and 75%), or weak (above 75%). The variogram model with the lowest RMSE based on the cross-validation (jackknifing) results were selected (Sówka et al 2020). Cross-validation is a technique used to assess the accuracy of a predictive model. It works by splitting the data into two distinct sets: a training set and a test set. Subequently, the model undergoes training using the training set and and is evaluated using the test set. The performance of the model is then evaluated based on how well it predicts the values in the test set. Cross-validation can help to identify overfitting, which occurs when a model is overly complex and fits the training data too closely. Crossvalidation can also be used to compare different models and select the best one for a given dataset. All variograms were assumed to be isotropic. Interpolated maps were then created using Surfer 24 (Golden Software Inc).

Results and discussion
Soil texture The conversion of forests into managed agricultural land is known to cause the deterioration of physical soil properties, leaving the land more susceptible to erosion. Soil erosion has a profound effect on the land, as it can reduce soil depth, altering the texture and leading to the loss of essential nutrients and SOM. This can have significant impacts on both the soil and the local ecosystem (Tellen and Yerima 2018). Therefore, it is imperative that strong steps are taken to mitigate soil erosion, in order to protect the environment and preserve the land for future generations.
The soil texture class was sandy loam for all three land use conditions (table 1). Sand and silt content did not change under the different land use conditions. Whereas clay content increased significantly from the converted lands to the natural forest. It has been suggested that vegetation-covered land increases the clay content of sandy loam soil, relative to bare lands. Root growth, litter decomposition and the formation of humus, predominant in the natural forest, are thought to influence the fixation of fine soil particles and SOM (Xia et al 2020). It is evident that the difference in clay content between natural forest and converted lands reflects the impacts of land use on the soil erosion process. This is because erosion is a selective process with respect to particle size distribution (Hao et al 2019).
ANOVA provides evidence that land use impacts resulting from soil erosion are reflected to some degree in the soil texture. Therefore, it is important to consider the implications of land use on erosion when assessing the soil texture. The soil texture attributes (sand, physical clay, and clay) of the sandy loam soil in the study area exhibited low sensitivity (CV < 15%). It is well-known that soil texture is an intrinsic property that reflects the parent material rather than the environmental conditions. It was reasonable to assume that texture would exhibit low sensitivity to land use. However, the results of ANOVA indicated that the texture factor (clay content) was significantly different between natural forest and the converted land.

Bulk density
Soil BD is an important indicator of soil compaction and health. According to Kakaire et al (2015), a higher soil BD implies less water is held by the soil at field capacity, while a lower density indicates soils which are less compacted and therefore have greater water retention. This finding is corroborated by the work of Hopmans et al (2021), who determined that soil BD was lower in a native forest relative to converted land.
BD was determined at 0-10 cm and 10-20 cm soil depth intervals. Land use conversions from natural forests to managed lands affect the compaction, porosity, and BD in soil (Kar et al 2022). BD was in the order of field road > agricultural land > natural forest (tables 1 and 2). The lower BD in the natural forest when compared to other land use conditions could be due to a favorable soil structure under forest vegetation and a stable soil environment devoid of anthropogenic activities over long periods. BD is significantly higher in the converted lands, likely due to direct influences such as compaction caused by agricultural field practices, as well as indirect influences such as the effects of land use on SOM (Smith et al 2016). Converting natural forests into agricultural land significantly increases BD (Yang et al 2020) possibly due to soil compaction and loss of soil organic carbon that occurs because of soil plowing and manipulation (Fraga et al 2020). The field road allowed access for vehicles and humans to different parts of the study area resulting in higher BD compared to ohter two land uses. Additionally, the agricultural land also experienced increased human traffic due to various research activities conducted there, contributing to its relatively higher BD. BD increased with depth under the different land use conditions.
These results could be attributed to the dynamism of organic carbon and the dissipation of organic components, which are present in greater concentrations at shallower depths and are likely to have a diluting effect, thereby reducing the BD (Bronick andLal 2005, Nwite et al 2018). The CV of BD within each land use condition was generally low (CV< 15%). Variations in BD among the different land use conditions were found to be minimal at deeper layers (10-20 cm) when compared to the surface soil layer (0-10 cm) (tables 1 and 2) thus indicating that various land use practices have a greater impact on soil BD at shallow depths when compared to deeper depths.

Soil organic matter
It is well established that cultivated soils generally have a lower SOM content when compared to native ecosystems, due to the increased aeration of soil which accelerates the decomposition of SOM (Tellen and Yerima 2018). SOM is critical for improving the physical properties of soil, increasing cation exchange capacity and water-holding capacity, and for increasing soil structural stability by binding particles into aggregates. Anthropogenic activities such as tillage (hoeing, plowing), biomass burning, residual removal, overgrazing, and drainage are thought to be responsible for the decreased SOM content observed under different land use conditions at various elevations (Steinmetz et al 2016). SOM plays an important role in soil health, as it prevents nutrient leaching, makes nutrients accessible to plants, and acts as a buffer to resist strong changes in pH (Roose and Barthes 2001). Carbon content is also known to be an essential element of an overall healthy soil (Kome et al 2019). SOM was found to be highest in the natural forest and lowest in the field road. This could be attributed to the higher clay content found in the natural forest. Soils with relatively high clay content, such as the natural forest, tend to stabilize and maintain more SOM than those with low clay content (Corwin and Scudiero 2019). The lesser amount of litter input in the field road when compared to the agricultural land and natural forest could have resulted in the low SOM in the field road (Morris 2004). Furthermore, when natural forest is converted into managed agricultural lands, SOM decomposes rapidly due to changes in temperature, aeration, and SWC (Corwin and Scudiero 2019). The decrease in SOM associated with land use conversion indicates the necessity of sustainable cropping systems, such as the addition of SOM, crop residues, crop rotation, and agroforestry using fast-growing leguminous trees to mitigate the negative effects of cultivation. Fallowing land has been found to not only improve soil fertility, but also reduce soil variability, which is beneficial for both practical and experimental agriculture.
The CV of SOM under the three land use conditions were generally low (CV < 15% ; tables 1 and 2) and varied in the order: natural forest < field road < agricultural land. The lower CV in the natural forest indicates a more stable spatial pattern (Atwell and Wuddivira 2019).

Soil water content
Research has shown that land use exerts a significant influence on SWC levels (Zhang et al 2021, Guo and Zhou (2021)). This primarily occurs due to differences in water consumption characteristics of vegetation and their impact on the root distribution of the soil. Changes in land use conditions can cause SWC levels to increase, decrease, or fluctuate. Studies have revealed that the effect of trees and shrubs on soil moisture levels is significant and can be observed in the entire soil profile (0-100 cm) after land conversions have taken place (Guo and Zhou (2021)). These changes are attributed to the differences in water uptake, as the variability in root distribution affects soil moisture levels in accordance with land use practices.
The mean SWC values were higher on the first day of the survey than on the second day (table 3), which could be attributed to the significantly higher total rainfall (84.2 mm) on the first survey day in comparison to the second survey day (19.3 mm) (figure 2). Furthermore, the mean SWC values were higher in the natural forest  than in the agricultural land and field road. This could be ascribed to the deep litter layer (high SOM) found in the natural forest, which reduces surface evaporation and improves water retention capabilities. This finding agrees with that of Morris (2004), who noted the expansion of fine soil pores that retain water against gravitational drainage in coarser textured soils such as podzolic soils. Conversely, the low SWC in agricultural land and field road might be attributed to the higher rate of evapotranspiration from crops and grasses, alson with possible decreease in infiltration caused byhigher surface runoff due to factors like surface crusting and compaction. The SWC was significantly higher on field roads relative to the agricultural land. This could be explained by the mechanical disruption of pore arrangements by practices such as tillage that lowers SWC in cultivated soils such as agricultural land (Dou et al 2020). SWC differed significantly under each land use condition, (tables 1 and 2) indicating that SWC is sensitive to land use changes. The spatial variability of SWC across the land use conditions were generally low (CV< 15%) (Warrick and Nielsen1980). Agricultural land and the field road displayed a higher CV of SWC in comparison to the natural forest (tables 1 and 2), suggesting the influence of management practices such as tillage, compaction influence the heterogeneity of SWC (Atwell and Wuddivira 2019).

Apparent electrical conductivity
In the study area, EC a values recorded at shallower depths (VCP C1, i.e., DOI of 0-0.25 m) were generally higher than at deeper depths (VCP C2, DOI = 0-0.5 m and HCP C1, DOI = 0-0.5 m) across the land use conditions, (tables 1 and 2) indicating that conductivity decreases with depth in the study area. The MC-EMI sensor's VCP C3, HCP 2, and HCP 3 provided more than 50% negative values and could not be analyzed. Aside from the 38 kHz frequency (HCP) from the MF-EMI sensor, all other frequencies registered negative values (more than 50% of the data); meaning, their measurements could not be analyzed. This is indicative of the limited usefulness of the MF-EMI in shallow soil investigations (Calamita et al 2015). However, its' ability to characterize wooded areas and acquire greater amounts data is an added value in the exchange of information between hydrology and geophysics (Calamita et al 2015).
EC a had medium sensitivity (15% < CV < 35%) to land use in the agricultural land and the field road while low sensitivity was observed in natural forests due to low anthropogenic activities (Atwell and Wuddivira 2019). Interestingly, CV for EC a generally increased with DOI which deviated from expectations since deeper depths are not typically exposed to weather or anthropogenic disturbances.
One-way ANOVA revealed low EC a readings across the different land use conditions with significant differences between natural forests and other land use conditions on both survey days (table 3); this difference could be attributed to high SOM, SWC, and clay content in natural forests (Jonard et al 2013).

Correlation analysis
Analysis of the data presented in tables 4, 5, and 6 indicates a positive correlation between EC a and silt and clay content across various land uses. Studies have suggested that the EC a of a soil is strongly determined by the clay content due to physical contact between soil particles which increases electrical conductivity (Brogi et al 2019; Grubbs et al 2019). In this study, the low clay content in the study area may explain the lack of a significant relationship between EC a and clay content across the different land uses. However, due to the low clay content, there was an increased silt-EC a interaction, which is consistent with the results of Khan et al (2016) and Sadatcharam (2019). Clay particles are smaller than silt particles and have a greater capacity to absorb and retain moisture, and thus when clay content is low, silt particles are more likely to affect EC a measurements. This interaction can provide valuable insight into the soil's water-holding capacity and other important soil characteristics. The negative significant correlation observed between sand content and EC a from the MC-EMI sensor across various land use conditions (tables 4, 5, and 6) suggests that the MC-EMI is more suited to represent variations in sand and silt contents compared to the MF-EMI sensor in this area. Such results are not unexpected, as numerous previous studies have observed that an increase in sand content leads to a decrease in EC a . This effect has been attributed to the non-conductive nature of sand particles.
The results of this study further demonstrated a strong correlation between EC a and SWC across all three land use conditions (tables 4, 5 and 6). This suggests that SWC is the major driver of EC a , an assertion that is supported by the findings of Knight and Endres (1990) and Badewa et al (2018) which argue that an increase in SWC can lead to an increase in EC a due to the additional solutes and conductive particles present in the soil solution, as well as a decrease in air thickness. The fact that EC a values are higher when SWC increases also accounts for why the correlation between EC a and SWC was stronger under natural forest than agricultural land or field road (tables 4 and 5): wet soils contain more SWC as well as more EC a than dry soils (Nocco et al 2019a(Nocco et al , 2019b. While EC a may serve as a proxy for measuring SWC (Badewa et al 2018), this correlation is subject to other factors such as compaction, which may be more pronounced in managed ecosystems like agricultural land or field road resulting in lower correlations between EC a and SWC than those found in natural forest (table 6).
The study demonstrates that the natural forest was the most suitable setting for investigating the relationship between EC a and SOM.This is due to the higher levels of SOM present in natural forests in comparison to agricultural and field road conditions. These latter conditions are characterized by the higher presence of human activities and the lack of vegetation , which can lead to reduce the correlation between EC a and SOM. The most significant influence of SOM on the MC-EMI C1 was observed at the 0 − 20 cm sampling depth, indicating that lower DOI is an important factor in this relationship. As such, this study provides evidence that the relationship between EC a and SOM can be impacted by land use, and that natural forests are the best option to examine this relation.
Surprisingly, the data gathered from both EMI sensors showed a general negative correlation with BD across different land use conditions, similar to the findings of Sadatcharam (2019). It was originally thought that as BD increased, EC a would also increase due to the decreased pore space, which would result in more solid contacts Table 4. Correlation matrix of soil properties under agricultural land for first survey day. † EC a = apparent electrical conductivity; VCP = vertical coplanar mode; HCP = horizontal coplanar mode; C1 = Coil 1; C2 = Coil 2; SWC = soil water content; SOM = soil organic matter; BD-1 = bulk density at 0-10 cm depth; BD-2 = bulk density at 10-20 cm depth; Average BD = average of BD-1 and BD-2. ‡ Significance is reported at 0.1 ( * ), 0.05 ( ** ) and 0.001 ( *** ), NS (non-significant correlations). Correlation coefficient (r) is reported in coloured boxes. Table 5. Correlation matrix of soil properties under field road for first survey day. † EC a = apparent electrical conductivity; VCP = vertical coplanar mode; HCP = horizontal coplanar mode; C1 = Coil 1; C2 = Coil 2; SWC = soil water content; SOM = soil organic matter; BD-1 = bulk density at 0-10 cm depth; BD-2 = bulk density at 10-20 cm depth; Average BD = average of BD-1 and BD-2. ‡ Significance is reported at 0.1 ( * ), 0.05 ( ** ) and 0.001 ( *** ), NS (non-significant correlations). Correlation coefficient (r) is reported in coloured boxes. and ions being available to conduct electricity (Corwin and Scudiero 2019). The results from this study, however, indicated that the EMI may not be a reliable method for representing BD variations at the study site. This is because EMI surveys measure the EC a of the soil, which can vary significantly between different soil types and locations and is not necessarily an accurate representation of the soil's BD. Moreover, EMI surveys are limited in their ability to measure the exact depth of the soil, resulting in inaccurate readings when attempting to measure BD variations.

Regression analysis
The use of MLR is a valuable tool when attempting to evaluate the accuracy of estimating SWC from EC a in the presence of multiple predictors such as texture, SOM, and BD contents of the soil being investigated. MLR can identify how each predictor influences the SWC, as well as whether the combination of predictors is sufficient to accurately estimate the SWC. Such an analysis provides valuable insights into the hydrological properties of the soil, and can help optimize agricultural practices, such as irrigation and fertilization, to promote the growth of crops. MLR predictions were compared to simple linear regression predictions during data processing, and it was found that MLR predictions were more accurate relative to simple linear regression predictions based on the higher R 2 and lower RMSE values obtained (table 7).
The VIF obtained from the MLR ranged from 1.00 to 1.51 across different land use conditions, indicating the absence of multi-collinearity in the developed regression models (Shrestha 2022). The accuracy of prediction improved significantly with increasing SWC, which is in line with the findings of other researchers, demonstrating high sensitivity of both sensors to measure SWC with increasing SWC level (Hassan-Esfahani et al 2015). The generated MLR models generally over-predicted SWC across different land use conditions (table 7). Generally, EC a was able to explain more variations in soil properties in the natural forest compared to the field road, where it was only seen to be able to explain variations related to SWC. The best model for agricultural land was obtained using coil 1 of the MC-EMI sensor in the VCP mode (EC a VCP C1). This was possibly due to the similar DOI between EC a (DOI = 0-25 cm) and the soil samples taken at 0-20 cm depth. EC a from this coil orientation significantly explained approximately 92.11% of the variations in SWC, SOM, and silt collectively at approximately 0-25 cm depth range. EC a did not explain variations in sand and clay in this land use condition regardless of the EMI sensor used and were subsequently removed from all generated models for this land use condition. EC a readings from the MC-EMI sensor in coil orientations VCP C2 and HCP C1 could only explain significant variations in SWC (68.06% and 82.95%, respectively). SWC and BD were the dominant significant factors being explained by the variation in EC a recorded from the MF-EMI sensor. The high DOI coupled with the highly sensitive nature of the MF-EMI sensor (Won et al 1996) could have resulted in low R 2 values relative to the MC-EMI sensor due to the significant influence of external features and other shallow soil properties (Farooque et al 2012). Table 6. Correlation matrix of soil properties under natural forest for first survey day. † EC a = apparent electrical conductivity; VCP = vertical coplanar mode; HCP = horizontal coplanar mode; C1 = Coil 1; C2 = Coil 2; SWC = soil water content; SOM = soil organic matter; BD-1 = bulk density at 0-10 cm depth; BD-2 = bulk density at 10-20 cm depth; Average BD = average of BD-1 and BD-2. ‡ Significance is reported at 0.1 ( * ), 0.05 ( ** ) and 0.001 ( *** ), NS (non-significant correlations). Correlation coefficient (r) is reported in coloured boxes.
It is likely that external factors such as soil-to-sensor distance variation, plant roots, and residues have an effect on the accuracy of the EMI sensor . Additionally, moist soils are more favorable for EC a surveys than dry soils, as seen by Brevik et al (2006), which can also explain the higher predictive accuracy found in the natural forest relative to the agricultural land and the field road (Sadatcharam 2019). Measurement errors associated with measuring BD via core sampling, transportation and drying may account for some limitations when predicting soil properties such as BD as well as errors associated with the core drying method when measuring SWC (Mouazen and Al-Asadi 2018).
Hence, both MC-EMI and MF-EMI sensors have been demonstrated to have the potential to predict subsurface soil properties based on EC a readings. However, results suggest that the MC-EMI sensor has a greater advantage in accurately predicting these properties in comparison with the MF-EMI sensor, particularly in managedlands such as agriculture and field roads, where measurement errors associated with the core drying method can be eliminated by keeping samples sealed until the drying test is conducted (Mouazen and Al-Asadi 2018).
Statistical comparison of TDR-measured soil moisture against predicted soil using selected models under the different land use condition The results of a comparison between the measured SWC (obtained from TDR) and the predicted SWC (obtained from VCP C1 using MLR) using the 1:1 line at α = 0.1 showed no significant difference (slope = 1 and intercept = 0) in the agricultural land (table 8). This result verifies that the MLR model was able to accurately predict SWC using EC a VCP C1 ( figure 4(a)). Significant differences were observed between the SWC prediction lines obtained from EC a VCP C2, EC a HCP C1, and EC a HCP 38 kHz and their respective 1:1 line at α = 0.1 in the agricultural land (slope ≠ 1 and intercept = 0) (table 8). Therefore, the results of the SWC predictions were found to be generally unreliable when using the EC a VCP C2, EC a HCP C1, and EC a HCP 38 kHz coil orientations. A comparison of the prediction values to the 1:1 line demonstrated that the predictions were consistently underestimated (figures 4(b)-(d), respectively). The prediction error in SWC when using EC a VCP C2, EC a HCP C1, and EC a HCP 38 kHz coil orientations could be due to the disparity in sampling depths between the TDR and these EMI coil orientations since the sampling depths of these EMI coil orientations are deeper (> 20 cm) than those of the TDR sampling depths (measured SWC) taken 0-20 cm depth (Calamita et al 2015;Altdorff et al 2017).
Furthermore, the analysis revealed that there were no significant differences between the prediction line obtained from EC a HCP 38 kHz and the 1:1 line at α = 0.1 in the field road (slope = 1 and intercept = 0) (table 8). T-test revealed that significant differences existed between the 1:1 line and prediction lines obtained using EC a VCP C1 (slope ≠ 1 and intercept = 0), EC a VCP C2 (slope ≠ 1 and intercept ≠ 0), and EC a HCP C1 (slope ≠ 1 and intercept ≠ 0) at α = 0.1 in the field road (table 8; figures 5(a)-(c), respectively). The insignificant difference between the prediction line and the 1:1 line when using the MF-EMI sensor (EC a HCP 38 kHz) could be attributed to the sensor's ability to produce thermally stable measurements in dry soils such as those found in the field road (Won et al 1996).
Lastly, the results of the comparison between the predicted SWCs (obtained from EC a VCP C1, EC a HCP C1, and EC a HCP 38 kHz) using their respective 1:1 lines revealed that there were no significant differences in the natural forest (table 8; figures 6(a)-(d), respectively). The slope of the regression line between the measured and predicted SWCs was equal to 1 and the intercept was 0, indicating that the predicted SWCs closely correlated with the measured values. These findings suggest that more EMI coil orientations were able to estimate SWC reliably in the natural forest relative to the other land use conditions. There were significant differences between the prediction line obtained from EC a VCP C2 and its 1:1 line in the natural forest (slope ≠ 1 and intercept = 0) ( figure 6(b)). While the MLR techniques may be able to accurately predict SWC in a natural forest, more research is needed to better understand the disparities that may arise when using EC a VCP C2.

Variography and kriging interpolated surface of soil water content
Cokriging is a powerful interpolation technique that provides useful estimates of data in areas where the data is sparse or of low quality. However, there are a few important limitations and assumptions to consider when using cokriging. Firstly, cokriging requires the availability of both the primary variable (SWC) and the covariate (EC a ) to be measured at the same locations, making it unable to interpolate data from one location to another. Secondly, the linear relationship between the primary variable and the covariate must be established for cokriging to be effective, and finally, the covariate must be an appropriate predictor of the primary variable. The last point is especially important when using EC a as a covariate, as it is not always a reliable predictor of SWC. It is thus important to carefully consider these limitations and assumptions of cokriging before using it. EC a = apparent electrical conductivity; VCP = vertical coplanar mode; HCP = horizontal coplanar mode; C1 = Coil 1; C2 = Coil 2; SWC = soil water content; SOM = soil organic matter; BD = bulk density at 0-20 cm depth; RMSE = root mean square error; R 2 = coefficient of determination Interpolation methods are increasingly being used as a tool to improve the estimation of SWC spatial distribution due to limited access to accurate observation data (Xie et al 2020). Since the MC-EMI VCP C1 gave the most accurate prediction from MLR, this coil orientation was used as a covariate for the interpolation of SWC using cokriging. A total of 4002, 1142, and 1037 EC a samples were collected in the agricultural land, field road, and natural forest, respectively using the MC-EMI VCP C1, while only 9 SWC samples were collected in    each land use condition. A visual comparison of the maps of measured SWC obtained from cokriging with EC a as a covariate, and the maps of predicted SWC from cokriging with EC a as a covariate in each land use condition, revealed notable discrepancies between both maps. These differences could likely be attributed to the degree of accuracy of the cokriging predictions, and the unique environmental conditions found in each land use class. However, the trends in the variability of both the measured and predicted SWC were similar. The various variograms are shown in appendix for further reference.

Agricultural land
The interpolated spatial maps for measured SWC for the agricultural land are displayed in figure 7(a), for which a spherical variogram model was the best fit for the spatial representation in the cokriging interpolation. The nugget effect of 5 indicated that additional sampling of SWC at smaller distances may be needed to detect spatial dependence and create a more accurate map. The range value of 8 m indicated that SWC values influenced neighboring SWC values over higher distances compared to other land use conditions, and the spatial dependence was categorized as medium (nugget:sill = 0.3), based on the classification of Cambardella et al (1994). Since the most accurate MLR prediction of SWC in this land use condition was obtained from the MC-EMI VCP C1, cokriging with measured EC a as a covariate was used to generate the spatial maps ( figure 8(a)). The SWC maps obtained from the MLR prediction (figure 8(a)) depicted low SWC content similar to the maps obtained from the TDR-measured SWC (figure 7(a)). A spherical variogram of the SWC maps obtained from the MLR prediction, with a range of 8 m, nugget effect of 0, and a partial sill of 20 depicted a strong spatial dependence (nugget:sill = 0) ( figure 8(a)). The accuracy of the SWC map obtained from the TDR-measured SWC was approximately 50% higher (RMSE = 0.09%) relative to the predicted SWC map (RMSE = 0.18%) possibly due to the influence of other soil properties such as SOM and silt, as seen in the MLR model.

Field road
The spherical variogram model was seen as the best fit for the spatial representation of the TDR-measured SWC in the cokriging interpolation ( figure 7(b)). The measured SWC maps in the field road generally depicted higher SWC content (35%-40%) relative to the agricultural land (25%-35%), as indicated by the results from the ANOVA (table 3). The observed range of SWC values (4 m) in this land use condition was substantially smaller than that of the agricultural land, indicating that SWC values in this land use condition were more heavily influenced by neighbouring SWC values at shorter distances. A partial sill of 10 and the high nugget value (10) also suggested that additional sampling of SWC at smaller distances may be needed to detect spatial dependence and create a more accurate map. Furthermore, the spatial dependence was classified as medium (nugget: sill = 0.5). The interpolated spatial map of the MLR predicted SWC using EC a as a covariate is shown in figure  8(b). Based on cross-validation, the accuracy of the TDR-measured SWC map was higher (RMSE = 0.12%) than the MLR predicted SWC map (RMSE = 0.20%).

Natural forest
The interpolated spatial maps of the TDR-measured SWC in the natural forest, displayed in figure 7(c), were generated using cokriging with a spherical variogram model as the best-fit representation of its spatial variation. The measured SWC maps generally showed a high SWC content of 40%-50%, with range values of 2 m, nugget values of 1, and partial sill values of 5.5. The small nugget value indicated that further sampling of SWC at shorter distances may not be necessary to demonstrate spatial dependence. Compared to other land use conditions, the range values were lower for wet soils (natural forest) and increased with decreasing SWC (field road and agricultural land) (Zhang et al 2019). The land use condition also exhibited a high spatial dependence (nugget: sill = 0.15). Furthermore, the natural forest revealed the strongest spatial dependence. The variogram of the MLR predicted SWC map in the natural forest (figure 8(c)) had a range of 3.4 m, nugget of 0, and partial sill of 20, and showed a strong spatial dependence (nugget:sill = 0). The accuracy of the TDR-measured SWC map obtained from cross-validation was higher (RMSE = 0.05%) when compared to the MLR predicted SWC map (RMSE = 0.15%). This is likely due to the significant influence of other soil properties such as SOM, BD, and silt as seen in the MLR model. The differences between measured and predicted SWC maps can have significant implications for the study of soil water dynamics in various environments. These discrepancies can provide insight into the limitations of current measurement techniques, predictive models, and the complex nature of soil water processes. Potential reasons for these differences include differences in spatial resolution, depth discrepancies, parameterization, land cover changes, vegetation heterogeneity, and water management strategies. For example, ground-based measurements may not capture fine-scale variability, which can lead to differences between measured and predicted maps. Additionally, land use and land cover changes can significantly affect soil water dynamics, as vegetation alters the expected SWC patterns. Understanding these discrepancies has implications for water management, agriculture, climate studies, and future research directions. For instance, stakeholders can use this information to improve data collection, refine models, and optimize water use and irrigation practices. Moreover, recognizing the limitations of current measurement and modeling techniques can lead to the development of advanced measurement techniques and improved modeling approaches. Additionally, largescale mapping of the spatial and temporal variability of soil properties and states is critical in delineating management zones to improve the efficiency of agricultural inputs through the advancement of precision agriculture. The implications of the differences between measured and predicted SWC maps are multifaceted and require continued research and data collection to enhance our understanding of soil water dynamics.

Practical implications of this research
The findings of this research have far-reaching practical implications for various stakeholders, including farmers, land managers, and policymakers. For example: (i) the research provides valuable insights for making informed land-use decisions, helping to identify areas of soil best suited for specific crops or land use activities; (ii) it can facilitate the development of accurate soil water prediction models, aiding in effective irrigation management and water conservation; (iii) the research can inform soil management strategies, particularly precision agriculture with targeted and site-specific resource application; (iv) it enables soil quality and fertility monitoring, allowing farmers to implement appropriate soil management practices to maintain soil health and productivity; (v) the findings can be crucial in informing policy decisions related to land use and soil management; and (vi) they also facilitate precision agriculture, leading to optimized resource utilization, cost reduction, and minimized environmental impacts. As a whole, the research makes a significant contribution to sustainable land management, increased agricultural productivity, and the reduction of negative environmental impacts.

Conclusion
EMI covers a range of depths, which is controlled by inter-coil separation and operation frequency, particularly in relatively shallow soils. However, due to its susceptibility to various soil properties, accurately estimating soil properties solely from EC a can be challenging. This study aimed to evaluate the effectiveness of EMI for measuring and predicting SWC in different land use conditions. The study revealed that measuring EC a using MC-EMI and MF-EMI sensors provided important information for maximizing the accuracy of SWC prediction under different land use conditions. This was obtained by establishing a workable relationship between EC a and the other investigated soil properties sampled at 20 cm across the different land use conditions. Generally, EC a was found to be significantly correlated to SWC and SOM, with SWC exhibiting the strongest relation with EC a across the different land use conditions in this study site. The correlation results between EC a and SWC were the strongest in the moistest soil (natural forest).
Measuring EC a using the MC-EMI (C1-VCP) provided the best predictions across the tested land use conditions compared to the other EMI coils employed in this study. Cokriging of TDR-measured SWC with EC a as a covariate revealed more accurate maps than cokriging of MLR predicted SWC with EC a as a covariate. The comparison of both sensors revealed that although MF-EMI used 8 integral depths from 4 different frequencies in 2 coil orientations, only the EC a −38 kHz in HCP coil orientation showed significant interactions with the other soil properties. On the other hand, the MC-EMI showed 6 integral depths that displayed significant interactions with the other soil properties. This result indicates that the MC-EMI sensor is more suitable for SWC characterization in this study area than the MF-EMI sensor. Additionally, the MC-EMI sensor provides a higher level of accuracy and precision for SWC characterization than the MF-EMI sensor. This is particularly beneficial in land use applications where SWC needs to be measured precisely and reliably. The findings from this study also demonstrate that georeferenced mobile soil EC a measurements carried out by using an MF-EMI or MC-EMI can aid in characterizing soil spatial variability rapidly. Furthermore, EC a can serve as a soil quality indicator for soil productivity and aid in site-specific agronomic management essential for land use conversion. The research conducted has the potential to have a considerable influence on land-use choices, by aiding in the recognition of the most suitable soil types for particular crops or activities. Additionally, it can be utilized to create precise soil water prediction models to facilitate effective irrigation and water conservation. Furthermore, it can enable soil quality and fertility monitoring, allowing farmers to preserve soil health and productivity. Additionally, it can provide insight into policy decisions in relation to land use and soil management and promote precision agriculture with improved resource utilization. Ultimately, the research contributes to sustainable land management, increased agricultural productivity, and reduced negative environmental impacts.
Further studies are needed to investigate inverse modeling to obtain 3D maps of SWC, as well as considering other environmental factors such as temperature and vegetation cover, in order to create a more comprehensive understanding of soil spatial variability and formulate more accurate prediction models. Additionally, more extensive field measurements should be conducted to enhance the robustness and reliability of the prediction models by including additional covariates such as temperature, which could help improve accuracy in predicting SWC across different land use conditions. Variogram models of predicted soil water content from MLR models obtained from cokriging for agricultural land (a), field road (b) and natural forest (c).
Simple linear regression between apparent electrical conductivity (EC a ) and soil moisture content (SWC) at 90% confidence