Accurate soil maps are critical for efficient and sustainable nitrogen and water management in specialty crops like citrus. Variations in soil properties, such as texture and moisture content, remarkably influence the availability and uptake of nutrients and water. Soil texture, determined by the proportion of sand, silt, and clay particles, affects key soil processes such as water retention, drainage, and nutrient-holding capacity. These processes, in turn, influence the movement and availability of water-soluble nitrogen and other macro and micro elements applied as fertilizers and diffused within the soil profile. Nitrogen-deficient citrus plants are stunted, whereas excessive nitrogen promotes vegetative growth and increases susceptibility to diseases that damage fruit, kill spurs, and may reduce yields in following years (Muhammad et al., 2018). Timing of nitrogen fertilization is crucial in citrus. Peak nitrogen uptake in citrus trees happens during blooming and early fruit growth (Muhammad et al., 2018). In areas with coarse-textured soils (e.g., sandy soils), the high permeability and low water-retention capacity may lead to nitrogen leaching, particularly in the form of nitrate (NO₃ ^-), beyond the root zone and into groundwater. Nutrient leaching reduces the fertilizer use efficiency by crops and increases the risk of groundwater contamination, contributing to issues such as eutrophication of water bodies and pollution of drinking water sources. Conversely, in fine-textured soils (e.g., clay-rich soils), higher water retention and slower drainage may lead to waterlogged conditions, especially in low-lying areas. Under such anaerobic conditions, denitrification processes may dominate, converting nitrate into gaseous forms of nitrogen such as nitrous oxide (N₂O), a potent greenhouse gas. Fields characterized by wide variability of soil texture and terrain attributes may have contrasting propensities for nitrogen loss through leaching or denitrification (Luo et al., 2023). Without accurate mapping and management of these variations, uniform application prescriptions of nitrogen and other nutrients can result in over-application in some areas and under-application in others. Over-application can intensify nitrogen losses and environmental impacts, especially in coarser soil (Syvertsen and Smith, 1996; Muhammad et al., 2018; Luo et al., 2023).

In the United States of America, soil survey maps developed by the government (e.g., USDA-NRCS soil survey maps) provide valuable broad-scale information on soil types and properties (Beaudette and O’geen, 2009; Chaney et al., 2016). Still, they are generally inadequate for guiding detailed field-scale water and nutrient management (Scudiero et al., 2024b). These maps, typically created at scales of 1:12,000 to 1:24,000, offer an overview of soil variability across landscapes but often fail to capture the fine-scale heterogeneity within individual fields, which is critical for precision agriculture (Reyes et al., 2018).

On-the-go soil sensing technologies provide an efficient means to generate high-resolution soil maps across agricultural fields. Among these technologies, apparent soil electrical conductivity (EC_a), gamma-ray spectrometry (GRS), and other on-the-go near-ground sensing technologies have been widely studied for their potential to map soil texture (Sudduth et al., 2005; Pätzold et al., 2020).

Apparent soil electrical conductivity (EC_a) is a measure of how easily electrical current passes through the soil, influenced by factors such as soil texture, moisture content, salinity, and temperature (Rhoades et al., 1976; Corwin and Lesch, 2003). EC_a is commonly used in precision agriculture to map soil variability, as it provides indirect information about soil properties that affect crop growth, such as clay content and water-holding capacity (Corwin and Lesch, 2003). This non-invasive, on-the-go sensing method is valuable for identifying zones within a field that require different management practices (Córdoba et al., 2016). Geospatial EC_a is arguably the most used sensor measurement for field-scale soil mapping by practitioners and scientists due to its relative ease of use, cost-effectiveness, and ability to capture spatial variability in soil properties influenced by factors such as moisture content, salinity, and clay content (Corwin and Lesch, 2003; Doolittle and Brevik, 2014). However, in environments where soil salinity is expected, such as in California and other mediterranean, arid, and semi-arid irrigated farmland worldwide, EC_a may not be the ideal tool for soil texture mapping because of salinity becoming a primary factor influencing the EC_a measurement (Corwin and Scudiero, 2016). Notably, EC_a measurements should be carried out over moist soils (i.e., around field capacity or slightly drier) to ensure reliable correlations with target soil properties (Corwin and Lesch, 2005b). To this regard, micro-irrigated orchards in water-scarce environments present unique challenges for EC_a sensing due to the very-short scale spatial heterogeneity of wetting soil conditions, which influences the sensor measurements. Soil moisture levels are generally ideal for reliable EC_a surveys in the hardly accessible areas under dense canopies, where micro irrigation is applied. Conversely, easily accessible alleyways have generally much drier soils (Pedrera-Parrilla et al., 2016; Corwin et al., 2022; Scudiero et al., 2024a). Soil compaction may also be remarkably different between below-canopy areas and the alleyways due to field equipment passages and other traffic. Even after precipitation events, EC_a surveys in the alleyways of micro-irrigated orchards may potentially lead to biased representation of field-scale soil spatial variability. Corwin and Lesch (2013) recommend surveying EC_a both in alleyways and along the tree lines distinctively in orchards and vineyards.

Gamma-ray spectrometry (GRS) measures, non-invasively, the natural gamma radiation emitted by isotopes of potassium (K), uranium (U), and thorium (Th) present in the soil. The entire energy spectrum of the gamma radiation, typically around the 0.1 to 3 MeV range, is also measured as Total Counts (TC) of gamma emission. Gamma-ray emitting nuclides are naturally present in soils and rocks. At the field scale, these gamma emissions correlate with specific soil properties, such as texture and mineral composition. Some factors influence gamma-ray emissions, e.g., increasing water content and bulk density decrease the measured gamma-ray volume (Grasty, 1979; Cook et al., 1996; Reinhardt and Herrmann, 2019). Mahmood et al. (2013) and Reinhardt and Herrmann (2019) provided detailed descriptions of GRS and its use for soil mapping and precision agriculture. Unlike EC_a, which primarily reflects soil moisture and salinity, GRS provides information on the mineralogical composition of the soil, offering a different perspective on soil heterogeneity. Moreover, it is recommended that EC_a measurements are carried out in moist soils (Corwin and Lesch, 2005b), whereas GRS is attenuated by high soil moisture (Reinhardt and Herrmann, 2019). In water-scarce and dry micro-irrigated orchards, where spatial variability in soil moisture is extremely short-scaled due to localized irrigation application under the tree canopies, GRS can complement EC_a taken along the driplines by providing additional data on particle size distribution and mineral content with measurements done in the drier alleyways (Scudiero et al., 2024a). This dual-sensor approach, developed and described by Scudiero et al. (2024a), can enhance the accuracy of soil maps, enabling more precise water and nutrient management in these complex environments.

Developing field-scale models from sensor data typically requires extensive ground-truthing (Reyes et al., 2018), involving the collection of numerous soil samples to calibrate and validate the sensor measurements (Corwin and Lesch, 2005b), as otherwise sensor maps only serve as a qualitative indication of soil spatial variability (Corwin and Scudiero, 2016). For regional-scale models, two primary approaches can be utilized: universal models, which apply broadly across regions (Lobell et al., 2010; Pätzold et al., 2020), and locally adjusted models, such as analysis of covariance (ANOCOVA) regression models (Corwin and Lesch, 2014), which tailor the sensor data to specific local conditions. The choice between these approaches depends on the degree of variability within the region, the desired accuracy of the soil maps, and the available resources that can be used for the ground truthing campaign. The reliability of soil property models, whether universal or locally adjusted, is highly dependent on the rigor and consistency of sensor and soil data collection protocols (Corwin and Lesch, 2005b). Standardized calibration, maintenance, and handling of sensors are essential to ensure accurate and comparable measurements across different locations and times. Proper spatial sampling design and timing are critical to capturing the full range of soil variability (Lesch, 2005), while standardized soil sampling and laboratory procedures ensure the accuracy of ground-truthing data (Corwin and Yemoto, 2017). Consistent adherence to these protocols minimizes measurement-induced variability, leading to more reliable and generalizable soil property models (Corwin and Scudiero, 2016). Protocols for field to regional scale soil mapping with EC_a and for EC_a-directed soil sampling have been developed and updated by Dennis Corwin and colleagues at the USDA-ARS US Salinity Lab (Corwin and Lesch, 2005b; Corwin and Scudiero, 2016). However, no equivalent protocols or recommendations for GRS are available (Reinhardt and Herrmann, 2019).

Universal models, also referred as “site-independent models” (Pätzold et al., 2020) should predict soil properties at novel agricultural fields without the need for additional ground truthing. For these models to be reliable in predicting soil properties across a broad region, several conditions must be met. There should be a quantifiable mechanistic relationship between the sensor measurements and the target soil properties across the entire region. This requires that the physical processes being measured, such as electrical conductivity or gamma-ray emissions, correlate strongly and predictably with soil attributes like texture, moisture content, or mineral composition, regardless of local variations in soil type or environmental conditions. The data used to develop the model should be representative of the full range of conditions within the region, ensuring that the model is not biased toward specific soil types or microclimates. Modeling approaches like support vector machines were shown as good candidates for predicting soil properties from on-the-go sensor measurements over datasets with diverse pedogenesis (Heggemann et al., 2017). Additionally, the secondary factors influencing the sensor measurements, e.g., soil-forming processes, land use history, tillage, and other factors (Corwin and Lesch, 2005b; Reinhardt and Herrmann, 2019), are similar enough across the region that a single model can adequately describe the soil property-sensor relationships everywhere. Finally, to confirm its accuracy and generalizability, a universal model should be validated against independent data sets from different locations within the region (Ramcharan et al., 2018) and with robust cross-validation techniques (Roberts et al., 2017).

Locally adjusted models, such as the ANOCOVA method operate under the assumptions that sensors consistently measure physical processes related to the target soil property across a given geographical regions and that any secondary influencing factors can be accounted for by adding local ground-truth data any time a new site is surveyed. If this assumption holds true, in ANOCOVA models, a constant slope can be applied to the model, allowing for the estimation of a local random effect (i.e., a field-specific intercept coefficient) with limited soil sampling (Harvey and Morgan, 2009; Corwin and Lesch, 2014, 2017; Scudiero et al., 2017). This approach potentially reduces the need for extensive ground-truthing while still providing accurate soil property estimates at the regional scale (Scudiero et al., 2017). In the context of calibrating on-the-go sensor measurements to map soil texture, ANOCOVA regression relies on several assumptions to ensure accurate results. First, ANOCOVA assumes a linear relationship between the sensor measurements (covariate) and the soil texture properties (dependent variable) across the region of interest, with the slope of this relationship being consistent across different conditions. Any interaction between sensor data and geographical or management differences should not significantly alter the slope of the relationship for the ANOCOVA model to be reliable. Second, ANOCOVA assumes that the residuals, or errors, of the regression model are normally distributed and exhibit homoscedasticity, meaning that the variability of these errors is consistent across all levels of the covariate. Moreover, the model assumes that the covariate is measured without error, which is critical for the reliability of the calibration process. For this, it is recommended that the sensor data collection procedure is methodologically consistent across the entire dataset (Corwin and Lesch, 2014).

The objective of this study was to evaluate the use of GRS for mapping particle size fraction in micro-irrigated citrus orchards in California and to compare its performance to that of EC_a, which served as the benchmark. Additionally, the study aimed to test whether the ANOCOVA approach could be effectively applied to GRS data to develop accurate soil maps with reduced ground-truthing requirements.

Arguably, on-the-go soil sensors should reliably correlate with the target soil properties that they developed to justify their broad and consistent use in high-resolution soil mapping. On-the-go sensors measuring EC_a capture a complex process which is influenced by many soil properties such as water content, salinity, and soil texture. These complex interactions contributing to EC_a have been studied in agricultural soils for many decades (Rhoades et al., 1976, 1989; Corwin and Lesch, 2003; 2005a), and are fairly well understood: e.g., higher clay content, water content, and salinity contribute to higher EC_a. Moreover, it is often the case that the spatial patterns of the soil properties influencing EC_a are locally correlated at the field scale, e.g., the soil physical properties and terrain drive the spatial variability of soil moisture and salinity, making statistical EC_a-to-soil-property calibrations at the field-scale fairly straightforward (Corwin and Scudiero, 2016). Because of the complex interactions between soil properties influencing EC_a, unexpected soil relationships are sometimes reported. For example, Scudiero et al. (2016) reported a negative relationship between EC_a and clay content from fields in shallow soils on hilly landscape in Colorado, USA. The GRS as reviewed by Mahmood et al. (2013) measures a more straightforward process: the emissions of gamma radiation from radionucleotides naturally occurring in soils and rocks (mainly from ⁴⁰K, ²³⁸U, and ²³²Th) and from anthropogenic ¹³⁷Cs. The overall TC are expected to be lower in sandy soils than in finer soils (Mahmood et al., 2013; Reinhardt and Herrmann, 2019); and others, and as supported by the global Pearson correlation coefficients observed in this study ( Table 4 ). Unfortunately, at the field-scale, TC relationships with soil texture can be erratic as shown by Pätzold et al. (2020) in Germany, by Maxton and Lund (2020) in the US Midwest, and as observed in this California citrus study ( Table 4 ).

Despite the potential of observing unexpected correlations between EC_a and GRS with texture-related field properties, we observed significant relationships between most target soil properties and sites using either of the sensors ( Table 4 ). Moreover, as discussed by Scudiero et al. (2024a), data fusion (e.g., principal component analysis, multiple regression) between the two sensors can be a means to obtain very accurate soil texture maps. Unfortunately, even when leveraging the spatial information from on-the-go sensing in spatial models (Reyes et al., 2018) and/or directing the soil sampling to calibrate ordinary least square models (Lesch, 2005), the amount of ground-truth data needed to develop accurate field-scale maps is too high (e.g., dozens of soil samples per field). The burden and the cost associated with the collection of ground-truth data from numerous locations within each field is a major bottleneck for the widespread application of on-the-go sensing for quantitative and accurate soil texture mapping.

Beyond the field-scale, the potential for developing regional site-independent models that would allow predicting soil properties at new sites without the need for collecting local ground-truth data has been investigated by many scientist in the past decades. Sudduth et al. (2005) investigated the use of EC_a from different sensors to map texture and cation exchange capacity across six US Midwest states. Their study sites from six states included soils of differing parent material, weathering, levels of organic matter content, and agricultural management. They observed R² values ≥ 0.55 for the two soil properties across their entire dataset. Although encouraging, their models were not sufficiently accurate to predict field-scale texture for precision agriculture. For GRS, Heggemann et al. (2017) developed a site-independent model to predict soil texture using data collected from ten agricultural fields in Germany across sites with diverse mineralogy and parent material. They developed site-independent support vector machines models that yielded R² of 0.96 (sand), 0.93 (silt), and 0.78 (clay), and with MAE values < 0.04. However, Pätzold et al. (2020) evaluated the model by Heggemann et al. (2017) at independent fields in Germany. They observed that the “model was not generally capable of predicting soil texture at sites that were not adequately represented in the calibration set”.

ANOCOVA regression as an alternative approach to universal models has been widely discussed for EC_a-to-salinity calibrations (Corwin and Lesch, 2014, 2017; Scudiero et al., 2017). Harvey and Morgan (2009) discussed the potential of ANOCOVA for texture mapping on three fields in Texas, USA, and observed prediction errors below 4% for clay. In this study, we report calibration MAE values < 0.05 for ANOCOVA EC_a-texture predictions ( Table 5 ). Previous research concluded that ANOCOVA may be a means to reducing field-scale soil samples (Corwin and Lesch, 2014; Scudiero et al., 2017). In particular, the regression slope common for all fields is calibrated over the entire available dataset, whereas only three or more samples per field may be needed to estimate the field-specific intercept. For the first time, this research shows that minimal data (n=5 per field) can be used to calibrate accurate ANOCOVA EC_a-texture regression with very low calibration and independent evaluation errors ( Table 5 ). Generating high-resolution, accurate soil maps using as few as five ground-truth sites per agricultural field may benefit practitioners seeking to decrease the costs related to soil sampling and laboratory analyses.

ANOCOVA regression was shown in this research to be a very powerful tool to map texture with limited data, but it relies on postulating the slope between the sensor and the target soil property to be stable over multiple fields. This was not the case for GRS-TC making ANOCOVA regression modeling not feasible. Over the three fields, the TC showed expected (e.g., positive r with clay) and unexpected (e.g., negative r with clay) with the target soil texture properties. These inconsistencies make GRS unsuitable for texture mapping in California citrus orchards using ANOCOVA. Field-specific modeling may be needed instead. Mahmood et al. (2013) and Pätzold et al. (2020) indicated that differences in parent material and clay mineralogy may be responsible for contrasting GRS and texture relationships over multiple fields. If fields with unexpected GRS-texture relationships could be identified from raw GRS data or using available landscape-scale soil maps, then soil scientists may decide whether to include such fields in ANOCOVA models or to calibrate field-specific models (which require larger ground-truthing). Moreover, reliable classification of the expected nature of the GRS-texture relationship may enable the calibration of reliable site-independent models (Pätzold et al., 2020). For this reason, soils from the three citrus sites were analyzed using DRIFTS, extraction of active mineral phases, and available USDA soil maps (Beaudette and O’geen, 2009) in relationship to the raw GRS TC, K, Th, and U observed at the three sites.

Mineralogical analysis through DRIFTS and extraction of active Fe, Al, and Mn oxides highlighted notable differences in soil clay fraction mineralogy between sites. The aluminosilicate and active oxide phases probed here both represent prevalent, reactive and high surface area minerals in soil clay size fractions that are known to retetain several elements in soils, including K, Th, and U (Mcbride, 1994; Duff et al., 2002; Bachmaf and Merkel, 2011; Hongxia et al., 2016; Dublet et al., 2017; Wang et al., 2021). As such, variation in this mineralogy between sites may translate to different affinities of these isotopes for the clay fraction of these soils. In the case of the soil minerals probed by DRIFTS, there were apparent differences in clay substitution, clay composition and/or oxide chemistry between the Strathmore site and Lemon Cove and Riverside soils. These factors could have impacted surface hydroxyl group availability (e.g., 1:1 vs. 2:1 clay), cation exchange capacity and exchange selectivity, which may greatly impact mineral retention of ions, including Th and U (Duff et al., 2002; Bachmaf and Merkel, 2011; Hongxia et al., 2016; Wang et al., 2017). In addition to the distinct aluminosilicate mineralogy at the Strathmore site, differences in extractable Fe, Al, Mn, and Si provide further insight into the deviation in GRS-texture relationship at the Strathmore site. An observation apparent in our data is the role of poorly-crystalline and amorphous Fe species in U and Th retention across sites, in alignment with previous studies showing the strong association of these Fe species for U and Th through adsorption and co-precipitation (Duff et al., 2002; Dublet et al., 2017; Li et al., 2019; Wang et al., 2021). Interestingly, poorly-crystalline Al, Mn, and Si forms showed a weak inverse relationship with extracted U and Th, indicating they did not associate to the same extent as with Fe. Provided the relatively lower proportion of Fe in total extracted poorly-crystalline mineral phases at the Strathmore site compared to Lemon Cove and Riverside, it is conceivable that this reduced the affinity of Th and U for fine-grained active oxides in Strathmore. This could, in turn, weaken or alter the relationship between soil clay content and GRS through lower radionuclide retention. While these results overall indicate that mineralogy may have played a role in observed inconsistencies in assessment of soil texture across sites with GRS, more detailed mineralogical analyses (e.g., X-ray diffraction “XRD”), extended X-ray absorption fine structure “EXAFS” spectroscopy) and isotope retention studies of isolated textural fractions across sites would further explore this relationship.

The mineralogical results align with the previous suggestion that differences in clay fraction mineralogy may result in contrasting trends of GRS and clay determination between field sites (Mahmood et al., 2013; Pätzold et al., 2020). This suggests that fundamental knowledge of site mineralogy may inform the applicability of GRS for soil textural determination between sites. This could relate further to variable environmental factors that may impact mineralogy of soil clay fractions, such as parent material, drainage and climate (Mcbride, 1994). For example, formation and dissolution of poorly crystalline Fe species related to Th and U retention across the Lemon Cove, Riverside and Strathmore sites are notably sensitive to soil redox fluctuations that may in turn impact radionuclide retention (Duff et al., 2002; Winkler et al., 2018). Knowing in advance the pedological properties of a soil may therefore provide some indication of what type of GRS defined textural relationships could be expected for a given site. Defining these relationships with respect to widely available soil survey data may also represent a future area for exploration with substantial implications for GRS application at regional scales.

As smart fertilizer and water management practices become increasingly important for California’s specialty crops, driven by factors such as resource scarcity, rising costs, and regulations, the need for accurate, high-resolution soil maps will grow. Geospatial sensors, such as soil apparent electrical conductivity (EC_a) and gamma-ray spectrometry (GRS), were confirmed in this study as reliable tools for field-scale soil mapping of particle size fraction, based on data from three citrus orchards in California. Model-based sampling schemes, such as the response surface sampling design used here, enable the creation of accurate soil maps using a relatively small set of ground-truth soil samples. Multi-field modeling using universal, site-independent models may not be feasible due to unknown secondary influences on sensor measurements at the individual field level. For EC_a, locally adjusted analysis of covariance (ANOCOVA) regressions modeled particle size fractions with high accuracy. Notably, the ANOCOVA regressions can be calibrated using limited (n=5 per field) data. This novel insight marks a step forward in making high-resolution mapping affordable for practitioners and their clientele. More research is needed to understand how ANOCOVA models that use minimal soil data can be developed, calibrated, evaluated, and improved over time (e.g., when any new field is added to a preexisting dataset).

In one of the orchards (the Strathmore orchard), GRS exhibited an unexpected negative correlation with clay content, making the use of ANOCOVA for GRS-texture regressions not possible. Although such relationships have been reported previously, the causes behind them remain poorly understood. To investigate this further, GRS ratios, diffuse reflectance infrared Fourier-transform spectroscopy (DRIFTS), and acid ammonium oxalate extractable elements were analyzed at all three sites. The Strathmore orchard displayed unique GRS ratios, DRIFTS, and acid ammonium oxalate extractable element profiles compared to the other sites. These novel insights may guide future research and help predict whether positive (expected) or unexpected GRS-clay content relationships are likely to occur at a given field, based either on raw GRS data or geochemical information. Such a-priori knowledge could inform soil scientists’ decisions about ground-truthing efforts (e.g., allocating additional resources to increase the number of soil samples for an accurate map) or be used as an additional predictor in regional models, such as support vector machines or random forest regression models. Here, the odd GRS-texture relationships were observed in one site only, which had fewer sampling locations than the other two sites. Further research is needed to identify other sites showing unexpected GSR-texture relationships to understand commonalities amongst these and contrasting features compared to sites where GSR performs as expected.