Multi-Source Data Integration for Sustainable Management Zone Delineation in Precision Agriculture

Abstract

Accurate delineation of within-field management zones (MZs) is essential for implementing precision agriculture, particularly in spatially heterogeneous environments. This study evaluates the spatiotemporal consistency and practical value of MZs derived from three complementary data sources: electromagnetic conductivity (EM38-MK2), basic soil chemical properties (pH, humus, P₂O₅, K₂O, nitrogen), and vegetation/surface indices (NDVI, SAVI, LCI, BSI) derived from Sentinel-2 imagery. Using kriging, fuzzy k-means clustering, percentile-based classification, and Weighted Overlay Analysis (WOA), MZs were generated for a five-year period (2018–2022), with 2–8 zone classes. Stability and agreement were assessed using the Cohen Kappa, Jaccard, and Dice coefficients on systematic grid samples. Results showed that EM38-MK2 and humus-weighted BSP data produced the most consistent zones (Kappa > 0.90). Sentinel-2 indices demonstrated strong alignment with subsurface data (r > 0.85), offering a low-cost alternative in data-scarce settings. Optimal zoning was achieved with 3–4 classes, balancing spatial coherence and interpretability. These findings underscore the importance of multi-source data integration for robust and scalable MZ delineation and offer actionable guidelines for both data-rich and resource-limited farming systems. This approach promotes sustainable agriculture by improving input efficiency and allowing for targeted, site-specific field management.

1. Introduction

Precision agriculture (PA) in Europe and Serbia faces several agronomic, economic, and demographic challenges. Small to medium-sized farms dominate Central and Eastern Europe, limiting economies of scale and raising production costs per hectare [1]. Serbia reflects this trend, with an average farm size of 5.4 ha and pronounced land fragmentation [2]. Concurrently, rural depopulation, population aging, and youth migration contribute to a shortage of agricultural labor [3]. As a result, PA offers a promising strategy to boost productivity and cut input costs, even for small and fragmented farms. Technologies like automated soil sampling, satellite- and UAV-based remote sensing, proximal sensing, and advanced analytics are becoming essential for modern, sustainable agriculture [4]. Moreover, the concept of digital twins and the fusion of real-time field and satellite data have gained prominence for zone delineation and input optimization [5].

While several pilot programs have introduced PA technologies to larger farms—particularly in parts of Western Serbia—broader adoption remains limited among smallholders. Reported yield increases of 12%, and even up to 100% for specific high-value crops, suggest the agronomic potential of these technologies; however, uptake is often constrained by high initial costs, insufficient advisory support, and limited technical capacity [6]. In regions like Vojvodina, where digital infrastructure is somewhat more developed, empirical studies still highlight that PA tools such as drones, online platforms, and remote sensing are underutilized due to financial and educational barriers [7]. While the EU promotes PA through the Common Agricultural Policy (CAP), Serbia is gradually embracing it through pilot initiatives and national rural development strategies [8]. Even so, further investment in digital infrastructure, broadband connectivity, and targeted education is needed to enable wider adoption, particularly on fragmented and resource-limited farms [9]. The integration of digital and geospatial technologies is thus becoming a critical driver of sustainable agricultural production, offering a pathway to address structural limitations and reduce the rural–urban digital divide.

PA is a modern farming approach that combines various technologies to improve yields while conserving natural resources [10]. Recent studies emphasize that the integration of site-specific data sources, even on smallholder farms, is increasingly supported by open access remote sensing and low-cost proximal sensors, enabling broader PA adoption across Europe and transitional regions [5,11,12]. It is grounded in the acquisition and analysis of spatial data on soil and crop variability, facilitating the delineation of site-specific management zones (MZs). This zonation supports targeted application of fertilizers, water, and other inputs based on localized needs [13]. The effectiveness of this approach depends on the combined use of complementary data sources, including proximal sensors, laboratory soil analyses, and satellite imagery, which together provide a comprehensive understanding of field heterogeneity [14,15].

Methods relying on apparent soil electrical conductivity (EC), particularly those using the EM38-MK2 sensor, have become increasingly important for detecting spatial variability in key soil physico-chemical properties in a fast and cost-effective manner [16]. The EM38-MK2 performs continuous, non-invasive measurements at depths of 0–0.5 m and 0–1.5 m, using horizontal and vertical dipoles to generate four data channels per point [17,18]. These readings indirectly reflect soil texture, compaction, moisture, salinity, and, in some cases, organic matter content [17,19], thereby supporting efficient management zone (MZ) delineation and potentially lowering input costs [20]. The approach offers high sampling density, broad coverage, and rapid data acquisition without direct soil contact. However, accuracy may be reduced in excessively moist conditions, where EC is dominated by water content [21], and interference from nearby metallic structures or rocks can require additional data correction.

Satellite-based data, particularly from the Sentinel-2 platform, are increasingly utilized in precision agriculture due to their ability to deliver frequent and wide-area observations of vegetation and soil conditions [22,23]. Vegetation indices (VIs) such as NDVI, SAVI, and LCI provide estimates of biomass and photosynthetic activity, while the Bare Soil Index (BSI) helps detect exposed soil and monitor land degradation [24,25]. These indicators reflect phenological development and support informed decisions throughout the growing season—from sowing to harvest [26,27]. Key advantages include large spatial coverage, high revisit frequency, and access to multi-year time series at up to 10 m resolution. However, satellite data are affected by atmospheric conditions and mixed pixels, and VIs may be influenced by factors unrelated to soil status, such as disease or mechanical damage [25,28]. When combined with field and laboratory data, these indices enhance the robustness of multi-source models for delineating stable and agronomically relevant management zones [28].

Basic soil property (BSP) analysis is a critical element of precision agriculture, offering direct insight into soil fertility through measurements of nitrogen, phosphorus (P₂O₅), potassium (K₂O), pH, and humus content. These parameters are essential for planning fertilization, selecting suitable crops, and assessing productive potential. BSP is valued for its analytical precision and agronomic relevance, yet it remains constrained by the localized nature of sampling, higher per-sample costs, and laboratory processing time. Consequently, BSP data are often integrated with other sources to support spatial interpolation and produce continuous soil property maps that better reflect within-field variability.

The past few years have witnessed a methodological shift from single-source to multi-layer zonation strategies, where EM38-derived soil structure, Sentinel-2-based vegetation dynamics, and laboratory-measured fertility indicators are combined to form robust models [12,29]. These approaches are increasingly favored for their capacity to capture the full spectrum of soil–plant interactions and improve zone stability across different seasons and crop types.

Although previous research highlights the importance of evaluating management zone (MZ) delineation methods—particularly regarding temporal stability and the potential to replace costly field analyses with remote sensing—few studies have jointly examined EM38-MK2, basic soil properties (BSPs), and satellite-derived indices [30]. Moreover, limited attention has been given to how classification techniques such as kriging, fuzzy k-means, or percentile-based methods influence spatial reliability. Integrating these three data sources offers a more comprehensive representation of field variability: EM38-MK2 captures physical soil traits, BSP reflects chemical composition, and satellite indices characterize vegetation dynamics and biological activity.

This study employed kriging as the principal interpolation method for mapping the spatial distribution of soil properties. As a geostatistical approach, kriging leverages spatial autocorrelation to produce continuous prediction surfaces and is commonly used in delineating management zones in precision agriculture. However, its accuracy is sensitive to sampling density and spatial configuration [31], with errors increasing when sampling falls below 0.5 points per hectare or when field conditions are heterogeneous [14]. The method also assumes normally distributed data with stable variance—assumptions often unmet in practice—and relies on robust variogram estimation, which becomes challenging with limited data [32]. In such cases, kriging may yield overly smoothed outputs that obscure within-field variability. Integrating finer-resolution datasets, such as vegetation indices or apparent electrical conductivity (EC) maps, has been shown to enhance the spatial precision of interpolated zones [15].

Recent research highlights the value of multi-year satellite data—particularly NDVI during peak vegetation periods—for assessing the temporal consistency of field performance. Excluding outlier years that deviate from typical agroecological patterns improves management zone reliability [33]. Simplified zoning methods based on a few seasonal observations are gaining traction, especially for small or resource-limited farms [11]. Integrating proximal data (e.g., EM38-MK2) with Sentinel-2 imagery and applying kriging can reduce sampling demands without sacrificing spatial resolution [5]. Clustering algorithms like fuzzy c-means and k-means, using vegetation (VI) and bare soil indices (BSIs), have proven effective with high-resolution time series. NDVI correlates well with soil properties such as organic matter, texture, and pH [12], and its combination with EC data in unsupervised classification has shown promise, although further agronomic validation is needed [30]. Multivariate geostatistical models that integrate dense and sparse data—such as Vis–NIR, EC, and NDVI—are increasingly employed to delineate zones capturing the full spectrum of soil variability [34].

This study aims to evaluate various methods for delineating agricultural management zones using data from EM38-MK2 measurements, basic soil properties (BSPs), and satellite-derived indices (VI/BSI), with a focus on assessing their spatiotemporal stability over a five-year period (2018–2022). It further seeks to determine the optimal number of zones and the minimal yet effective set of input parameters for practical application in precision agriculture.

The paper is structured as follows: Section 2 details the data sources and methodology; Section 3 presents the comparative results and stability assessments; Section 4 discusses key findings and their practical implications; and Section 5 concludes with recommendations for future research and operational deployment.

2. Materials and Methods

This study investigates soil spatial variability in the context of precision agriculture by integrating proximal sensing data, satellite-derived indices, and laboratory-based soil analyses to delineate stable and agronomically relevant management zones. The methodology combines several classification approaches—such as quantile-based partitioning and weighted overlay—with spatial validation metrics including the Cohen Kappa, Jaccard, and Dice coefficients.

This section outlines the applied methodology (Figure 1), focusing on multi-source data integration, selection of informative input layers, and determination of the optimal number of management zones. Fuzzy k-means clustering was the primary classification method. Its quality was evaluated using the Fuzzy Performance Index (FPI) and Normalized Classification Entropy (NCE). These tools supported informed decision-making in defining functionally relevant management zones.

2.1. Study Area

The study was conducted in the Komareva Humka cadastral municipality, located in southeastern Vojvodina, Serbia. As part of the Belgrade agricultural basin—formerly managed by PKB and now operated by Al Dahra Serbia [35,36]—the area is characterized by flat, fertile terrain dominated by alluvial and chernozem soils. This landscape supports intensive field crop production, including cereals, industrial, and forage crops. Long-standing agricultural practices and well-developed irrigation and drainage infrastructure make the region one of Serbia’s key arable zones.

The experimental study covered 10 agricultural parcels in Vojvodina, Serbia, ranging from 16 to 92 hectares in size, with a total area of 378 ha (44°56′ N, 20°29′ E) and elevations between 64 and 80 m a.s.l. (Figure 2). Field data were collected in July 2022—on the 11th, 15th, and 19th—shortly after wheat and barley harvests. All parcels were under intensive crop production, primarily focused on field crops.

The parcels were chosen based on the availability of synchronized multi-source data, which is rarely accessible under typical field conditions. Specifically, both soil chemical properties (BSPs) and EM38-MK2 electromagnetic measurements were collected on the same day, at nearly the same time, across all fields (Figure 3). This level of spatial and temporal alignment ensured consistency in data integration and allowed for reliable comparative analysis. Although the parcels included in the study range from 16 to 92 ha, the applied methods—including BSP, EM38-MK2 scanning, interpolation, fuzzy clustering, and weighted overlay—are scalable and applicable to much smaller agricultural plots. Based on practical experience with this workflow, the approach can be reliably applied to plots as small as 0.5 ha. However, due to the 10-m spatial resolution of Sentinel-2 imagery, we recommend using this workflow for parcels above that threshold to avoid degradation of classification accuracy.

2.2. Data Collection

2.2.1. EM38-MK2 Data

Soil apparent electrical conductivity (EC) was measured using the EM38-MK2 sensor (Geonics Ltd., Mississauga, ON, Canada) at four configurations (C1, C0.5, L1, L0.5), corresponding to two depths and dipole modes. The sensor records both conductivity and magnetic susceptibility components, enabling assessment of soil texture, moisture, salinity, and, under certain conditions, organic matter and nutrient distribution [17]. Thanks to its dual-coil design and enhanced sensitivity, EM38-MK2 offers improved accuracy over earlier models [37], and it has proven effective in capturing spatial variability and delineating management zones in agricultural fields [38].

The EM38-MK2 sensor was mounted on a custom sled towed by a vehicle moving at 20 km/h, with transect spacing ranging from 30 to 40 m, depending on parcel shape. This configuration yielded an average point spacing of 8.5 m. On the smallest plot (16 ha), approximately 8000 measurements were recorded, while the largest (90 ha) generated around 27,000 readings (Figure 3). Although not fully optimized [38], the sampling ensured consistent and reliable data acquisition across all parcels.

2.2.2. Sampling and Analysis of Basic Soil Properties

Soil sampling followed standardized protocols defined by ISO 18400-102:2017, which recommends systematic or stratified approaches based on field heterogeneity, typically at one composite sample per 1–5 ha [39]. International practices such as the LUCAS survey [40], FAO guidelines [41], and national protocols in Germany [42] and the U.S. [43] emphasize spatial representativeness. In Serbia, similar recommendations apply, with 3–5 ha per sample being standard in Vojvodina.

We collected soil samples at a depth of 0–30 cm using automated probes (Duoprob 60-UP and Multiprob 120, Soil Probetechnik Peters GmbH, Quakenbrück, Germany). Each ~5 ha section contained 15–25 evenly distributed points, forming one composite sample. For example, a 16 ha parcel yielded 3 samples, while a 90 ha parcel resulted in 18 composite samples (Figure 3). Laboratory analysis included pH (KCl), humus (%), total nitrogen (%), available phosphorus (P₂O₅ mg/kg), and potassium (K₂O mg/kg), in line with accredited procedures. Sampling was performed post-harvest, ensuring timely results for fertilizer planning. The methodology aligns with both national recommendations and regional agronomic practice [44,45].

2.2.3. Sentinel-2 Satellite Imagery and Indices

In addition to soil sampling and measurements of apparent electrical conductivity, satellite data from the Sentinel-2 platform were collected for all parcels included in this study. To complement field measurements, Sentinel-2 imagery from 2018 to 2022 was used to derive vegetation indices—NDVI, SAVI, LCI—and the Bare Soil Index (BSI) for all studied parcels [46,47,48]. For each season, NDVI time series were interpolated to identify phenologically relevant dates, resulting in five temporal configurations: one (NDVI peak), two (before + peak), three (before, peak, after), four (two before, peak, after), and five dates per year (Figure 4).

BSI was calculated using three configurations: pre-sowing (minimum NDVI), post-harvest (minimum NDVI), and a combination of both (Figure 5). These temporally stratified layers supported further zonation and stability analysis.

To address the well-documented issue of NDVI saturation under dense vegetation, we included two additional indices—SAVI and LCI—that mitigate this limitation. SAVI incorporates a soil brightness correction factor to reduce background noise in medium-density vegetation, while LCI leverages Sentinel-2 red-edge bands, which are less prone to saturation and more sensitive to chlorophyll variation. This complementary index selection was intended to enhance the robustness of zonation across different canopy densities.

2.3. Spatial Analysis

Following data acquisition from EM38-MK2 and laboratory soil analysis, spatial interpolation was performed using Ordinary Kriging or Support Vector Machines (SVMs). We applied kriging using five isotropic semivariogram models—linear, linear with threshold, exponential, spherical, and Gaussian—with the optimal model selected via cross-validation [49]. All raster outputs were normalized to a 0–1 scale and resampled to 10 m spatial resolution to ensure compatibility with Sentinel-2-based VI and BSI layers.

Figure 6 includes a total of twenty-five interpolated and normalized raster layers, grouped into five categories: four EM38-MK2 layers (C0.5, C1, L0.5, L1), five basic soil property layers (humus, pH, P₂O₅, K₂O, nitrogen), one BSI index layer, and five layers each for NDVI, SAVI, and LCI derived from different seasonal dates. All layers were resampled to 10 m spatial resolution and normalized to a 0–1 scale to enable consistent comparison.

Following cloud filtering and date selection, satellite data processing included the generation of temporally integrated vegetation and bare soil index (BSI) layers. For each of the five selected dates, NDVI, SAVI, and LCI were computed and averaged into a composite raster:

$${Mean Composite}_{date}={\displaystyle \frac{NDVI+SAVI+LCI}{3}}$$

(1)

Each raster was assigned a temporal weight, depending on the number of available dates—for example, [0.3, 0.4, 0.3] for three dates (VI) or [0.6, 0.4] for BSI (pre- and post-season). Final composites were calculated using the following:

$$Final Composite=\sum _{i=1}^n{w}_i\times R_i$$

(2)

where $w_i$ is the weight and $R_i$ the date-specific raster.

Post-processing included a majority filter to reduce noise and normalization to the [0, 1] scale. The resulting temporally balanced rasters served as inputs for subsequent classification and management zone delineation.

2.4. Cluster Analysis

Beyond interpolation, we classified all datasets into management zones (MZs) using fuzzy k-means clustering. To determine the optimal number of classes, Fuzzy Performance Index (FPI) and Normalized Classification Entropy (NCE) were calculated, following established recommendations [50,51,52]. Clustering was performed with a maximum of 100 iterations and a fuzzy exponent of 1.25. Across parcels and data types (BSP and EC), the optimal number of classes typically ranged from 2 to 8, with occasional suggestions up to 10. Consequently, classifications were conducted using 2–8 classes per dataset to analyze spatial variability and identify viable management strategies. Figure 7 illustrates an example of how the same dataset is classified using an increasing number of management zones. The figure includes 14 sub-images, which display clustering results for the EM38-MK2 C0.5 layer and a BSI layer corresponding to a selected date. Each example shows zonation maps derived using 2 to 8 classes, allowing for a visual comparison of how classification complexity affects the spatial delineation of management zones.

Three classification approaches were applied across all 10 parcels: fuzzy k-means clustering, Weighted Overlay Analysis (WOA), and percentile-based classification. For each data combination, raster layers were generated using 2 to 8 classes, resulting in eight classified rasters per configuration.

For BSP data, two sets of fuzzy classifications were produced: (a) using all chemical properties and (b) humus only. For EM38-MK2, fuzzy classification included all individual depths (C1, C0.5, L1, L0.5), selected combinations (e.g., C1+C0.5), and each parameter separately.

The weighting coefficients used in WOA were based on agronomic relevance and adapted to the specific combination of parameters. For BSP-only configurations, two weighting schemes were used: the first assigned humus, phosphorus (P₂O₅), and potassium (K₂O) a ratio of 0.65:0.23:0.12; the second incorporated pH with adjusted weights (humus 0.63, P₂O₅ 0.21, K₂O 0.11, pH 0.05).

In configurations combining BSP with EM38-MK2 ECa layers (C1 and C0.5), a portion of the total weight was allocated to ECa inputs (typically 0.2 each), while the remainder was proportionally distributed across BSP parameters. These included humus, P₂O₅, K₂O, pH, and in some schemes, total nitrogen, depending on the configuration.

This flexible approach ensured that agronomically important variables remained well represented, while incorporating physical soil structure data into the zonation models.

Percentile-based classification was applied to VI and BSI layers across five growing seasons (2018–2022), using one to five selected dates per year. For BSI in 2018, three configurations were analyzed: pre-sowing, post-harvest, and their combination. Each configuration generated 8 classified rasters (2–8 classes) per parcel (Figure 7).

2.5. Zonation Stability Analysis

To assess the reliability of spatial classifications and delineated management zones (MZ), we conducted a multi-metric stability analysis. Given the importance of spatial consistency in precision agriculture, three complementary methods were applied: (1) systematic grid sampling to examine the effect of sampling resolution [53,54,55,56]; (2) Cohen’s Kappa coefficient to measure agreement between classified layers while correcting for chance [55]; and (3) the Jaccard and Dice coefficients to quantify spatial overlap in binary zonations [57,58]. This framework enables a comprehensive evaluation of zone robustness, capturing both internal consistency across methods and sensitivity to input configurations and class number. A concise overview of the underlying geostatistical and clustering methods used in this study (e.g., fuzzy k-means, kriging, and agreement indices) can be found in [50,51,55], which offer accessible introductions for readers less familiar with these techniques.

2.5.1. Systematic Grid Sampling

Zone stability was evaluated using systematic grid sampling at intervals of 100 m, 50 m, 20 m, and 10 m across all fields. On the largest parcel (90 ha), this corresponded to 88, 354, 2216, and 8857 sampling points, respectively. Values were extracted from all classified raster layers—including EC, BSP, VI, BSI, and their composites—at each grid point. This approach allowed for consistent comparison of classification outputs across methods and resolutions, minimizing random sampling effects. The extracted data served as input for computing agreement metrics such as the Cohen Kappa, Jaccard, and Dice coefficients [55,57,58].

2.5.2. Kappa Coefficient

We employed Cohen’s Kappa coefficient (κ) as the primary metric to assess agreement between classified raster layers, accounting for chance-level concordance. The Kappa coefficient ($\mathsf{\kappa }$) is defined as follows:

$$\mathsf{\kappa }={\displaystyle \frac{p_o-p_e}{1-p_e}}$$

(3)

where $p_o$ is the observed agreement and $p_e$ the expected agreement under random classification. Values range from −1 (total disagreement) to 1 (perfect agreement), with values near 0 indicating randomness. The Kappa value was calculated across all raster group combinations (EC, BSP, VI/BSI, and their weighted variants), with stratification by class number (2–8) to evaluate the impact of classification granularity on consistency.

2.5.3. Jaccard and Dice

To complement the Kappa analysis and assess spatial similarity of binarized management zones, the Jaccard and Dice coefficients were applied. The Jaccard index (J) quantifies overlap as the ratio of intersection to union:

$$\mathrm{J}={\displaystyle \frac{\left|A\bigcap B\right|}{\left|A\bigcup B\right|}}$$

(4)

The Dice coefficient (D) gives more weight to shared areas and is defined as follows:

$$\mathrm{D}={\displaystyle \frac{2\times \left|A\bigcap B\right|}{\left|A\right|+\left|B\right|}}$$

(5)

These metrics were used to evaluate spatial agreement between zones produced by different methods and data combinations (e.g., EC vs. BSP, BSP & WOA vs. VI & BSI), offering additional insight into zone consistency and robustness.

3. Results

3.1. Evaluation of EM38-MK2 Data

To assess the internal consistency, classification agreement, and zonation stability of soil electrical conductivity (EC) data from the EM38-MK2 sensor, we conducted a detailed analysis. Pearson’s correlation, Cohen’s Kappa, and spatial similarity indices (the Jaccard and Dice coefficients) were used on original, normalized, and classified rasters interpolated at four spatial resolutions (10, 20, 50, and 100 m).

3.1.1. Correlation Analysis of Raster Layers for Zonation Stability

Correlation analysis confirmed high consistency across EC raster layers at different spatial resolutions. Normalization preserved the internal ranking of values, with Pearson’s correlation coefficients equal to 1 between original and normalized rasters at all grid sizes (10–100 m). Interpolated and normalized layers exhibited strong mutual correlations, while classified rasters showed the highest agreement when the number of classes matched, regardless of EM38-MK2 channel or combination (Figure 8). Coarser grids (e.g., 100 m) resulted in higher mean correlations (≈0.615), indicating reduced local variability and enhanced zonation clarity. These findings support the use of normalized and interpolated layers for stable management zone delineation.

3.1.2. Zonation Consistency Assessment Based on the Kappa Coefficient

Cohen’s Kappa coefficient was applied to evaluate zonation consistency derived from EM38-MK2 data, accounting for variations in class number and channel configuration. The highest average agreement was observed for two-class schemes (mean Kappa = 0.6436, max = 0.9541), though with substantial variability (SD = 0.2848), indicating sensitivity to localized variation. Three-class zonation offered a more balanced outcome (mean = 0.4910, max = 0.8812, SD = 0.1560), while four- and five-class configurations maintained moderate stability (mean ≈ 0.46, SD ≈ 0.1). Beyond five classes, both the mean and maximum Kappa values declined, though variability remained low (SD = 0.0505–0.0772), reflecting weaker but stable agreement.

Analysis of channel combinations (Figure 9) showed the best average agreement between C1 and the full composite (C1C05L1L05) with the mean Kappa = 0.4935. Combinations involving C05 and C1+C05 performed similarly well (mean ≈ 0.474–0.478), while inclusion of lateral channels (L1, L05) reduced consistency, likely due to added spatial noise. The highest pairwise Kappa (0.9541) was recorded for C1 vs. C1+C05 with two classes, yet this pairing also showed the lowest value in some cases (Kappa = 0.2227), illustrating strong spatial heterogeneity. The lowest average agreement was found for C05 vs. C1C05L1L05 (mean = 0.4533; range = 0.2805–0.8754), while C1+C05 vs. full composite yielded a slightly better mean of 0.4689. Conversely, C05 vs. C1+C05 produced 0.4788, and C1 vs. full set maintained the highest average (0.4935), indicating that C1 alone can be considered a reliable indicator.

In summary, two-class zonations achieve the strongest agreement but with high variability, whereas three-class configurations provide a good compromise. Class ranges from 3 to 5 and depth-oriented configurations—especially those based on C1—offer the most stable and reliable delineation of soil management zones.

3.1.3. Multi-Class Zonation Similarity Analysis Using the Jaccard and Dice Metrics

The Jaccard and Dice coefficients were used to assess spatial consistency of EM38-MK2-based zonation maps classified into 2–8 classes, across four configurations: C1, C05, C1C05, and C1C05L1L05. As shown in

Table 1. Group-wise quantitative similarity analysis based on the Jaccard and Dice coefficients.

Group	Jaccard (Mean ± SD)	Jaccard Max	Dice (Mean ± SD)	Dice Max
C1	0.0857 ± 0.1657	0.8922	0.1244 ± 0.2214	0.943
C05	0.0771 ± 0.1517	0.9428	0.1147 ± 0.2045	0.9706
C1C05	0.0710 ± 0.1396	0.8179	0.1072 ± 0.1967	0.8998
C1C05L1L05	0.0750 ± 0.1853	0.8459	0.1010 ± 0.2277	0.9165

, single-depth configurations achieved the highest mean similarity: C1 (Jaccard = 0.0857, Dice = 0.1244) and C05 (Jaccard = 0.0771, Dice = 0.1147). Multi-channel setups showed slightly lower values. Despite low means, all groups included raster pairs with the Dice value > 0.9, indicating strong localized overlap. While overall similarity values were low, configurations based on single-depth inputs (C1 and C05) showed more consistent spatial overlap, with several raster pairs achieving a Dice coefficient above 0.9. This indicates that simpler EM38-MK2 setups can capture stable spatial structures, especially when additional channels introduce noise.

Class number significantly influenced similarity (

Table 2. Similarity trends by number of classes based on the Jaccard and Dice coefficients.

Number of Classes	Jaccard Mean	Jaccard Max	Dice Mean	Dice Max
2	0.1427	0.8179	0.21	0.8998
3	0.0966	0.8194	0.1402	0.9008
4	0.0642	0.8318	0.0979	0.9082
5	0.0318	0.6204	0.05	0.7657
6	0.031	0.9428	0.0438	0.9706
7	0.0098	0.5719	0.0143	0.7276
8	0.0041	0.4039	0.0059	0.5754

). Two-class schemes yielded the highest mean values (Jaccard = 0.1427, Dice = 0.2100), followed by three-class zonations. From four classes onward, both metrics declined, with the lowest overlap at eight classes (Jaccard = 0.0041, Dice = 0.0059), reflecting increased spatial fragmentation. A clear trend was observed with class number: zonations with two or three classes showed stronger spatial agreement, while higher class counts resulted in fragmentation and lower overlap. This confirms the trade-off between detail and spatial coherence.

Overall, zonations with 2–4 classes and single-depth inputs—especially C1 and C05—achieve the best trade-off between spatial coherence and interpretability. These configurations are suitable for standalone use or integration with additional data sources, such as vegetation indices, in precision agriculture workflows.

3.2. Evaluation of BSP Data

The effectiveness of zonation based on basic soil properties (BSPs) was evaluated using the correlation, Kappa, Jaccard, and Dice metrics.

3.2.1. Correlation Analysis of Raster Layers for Zonation Stability

Raster layers derived from humus, P₂O₅, nitrogen, pH, and K₂O showed consistently high mutual correlations across all grid resolutions (10–100 m), with mean absolute Pearson’s coefficients ranging from 0.622 to 0.623. Strong (|r| > 0.6) and very strong (|r| > 0.8) relationships were stable across the dataset. Normalized and classified layers in particular demonstrated reliable internal structure, suitable even without interpolation. Highest correlations were observed between classified layers of the same variable and among composite overlays including humus, P₂O₅, and K₂O. These results confirm the robustness of BSP-based zonation across spatial scales, especially in data-constrained scenarios.

3.2.2. Zonation Consistency Assessment Based on the Kappa Coefficient

Agreement between BSP- and humus-based zonation maps and Weighted Overlay Analysis (WOA) outputs was evaluated using Cohen’s Kappa. Values ranged from −0.4592 to 0.3822, with a low overall mean of 0.013, indicating generally weak consistency. Out of 294 comparisons, 159 yielded positive values and 135 yielded negative values, highlighting WOA’s sensitivity to input composition and weight allocation. The highest agreement (Kappa = 0.3822) was found between a two-class BSP zonation and a WOA layer weighted by humus (65%), P₂O₅ (23%), and K₂O (12%). In general, humus-only maps aligned more closely with WOA outputs than full BSP combinations, suggesting humus as the dominant factor in spatial differentiation.

3.2.3. Multi-Class Zonation Similarity Analysis Using the Jaccard and Dice Metrics

Similarity between WOA and BSP/humus-based zonation was overall low (mean Jaccard = 0.096; Dice = 0.139). However, individual cases reached high values (Jaccard = 0.8735; Dice = 0.9325), consistently associated with humus-dominant overlays. These findings mirror the Kappa results and underscore the role of humus weighting in achieving spatial coherence.

3.3. Evaluation of VI and BSI Data

3.3.1. Correlation Analysis of Raster Layers for Zonation Stability

Classified raster layers derived from vegetation indices (VIs) and bare soil index (BSI) showed stable spatial patterns across all grid sizes, with the mean Pearson correlation values ranging from 0.32 to 0.34. Resolution had minimal impact, though 100 m grids recorded the highest number of strong correlations (|r| > 0.8), indicating enhanced zonation stability at coarser scales (

Table 3. Summary of the Pearson correlations for VI and BSI classified rasters by resolution.

Grid Resolution	Mean Mean ± Std. Dev.	Min	Max	|r| > 0.6	|r| > 0.8
10 m	0.3375 ± 0.3231	−0.2140	0.9994	360	248
20 m	0.3393 ± 0.3230	−0.2167	0.9994	360	250
50 m	0.3194 ± 0.3366	−0.2631	0.9995	356	244
100 m	0.3373 ± 0.3511	−0.3534	0.9996	382	272

; Figure 10). This suggests that coarser grids may enhance zonation clarity by reducing local noise, while preserving key spatial trends in VI and BSI patterns. Full correlation matrices with raster layer names for all four grid resolutions (10 m, 20 m, 50 m, and 100 m) are provided in Supplementary Table S1.

Multi-year zonations integrating VI and BSI layers exhibited strong mutual correlations, suggesting a shared spatial structure. BSI layers based on just two dates (typically pre-sowing and post-harvest) aligned closely with multi-date VI zonations, making them effective in data-constrained scenarios. The highest inter-annual correlation (r = 0.995) was observed between a two-date BSI zonation from 16 April 2020 and 5 December 2020 and a four-date VI zonation from 2021. Other high correlations (r > 0.94) were noted between the following:

• two-date BSI and multi-date VI composites;
• annual and multi-annual zonations (e.g., 2021 vs. 2018–2022);
• spring–autumn combinations across different years.

These findings underscore that both limited-date BSI and multi-date VI layers can reliably capture temporally stable spatial variability, offering practical value for zonation in varying data environments.

3.3.2. Zonation Consistency Assessment Based on the Kappa Coefficient

Cohen’s Kappa analysis showed a declining trend in mean agreement with increasing class count (

Table 4. Summary of the Kappa metrics across class counts (2–8).

Number of Classes	Mean Kappa	Max	Min	Std. Dev.
2	0.2794	0.9761	−0.244	0.2841
3	0.2155	0.9614	−0.1423	0.2422
4	0.1798	0.955	−0.1112	0.2179
5	0.1554	0.9435	−0.082	0.2001
6	0.1389	0.9303	−0.0695	0.1865
7	0.1255	0.9306	−0.0603	0.1758
8	0.1153	0.9162	−0.0566	0.1665

). Two-class zonations yielded the highest mean Kappa (0.2794), but also the highest variability (SD = 0.2841). Conversely, seven- and eight-class schemes had lower average values (0.1255 and 0.1153), but greater internal consistency (SD = 0.1758 and 0.1665, respectively). These results indicate that simpler (2–3 class) configurations offer higher agreement but greater variability, while higher class counts provide more stable yet divergent outcomes.

The highest observed Kappa (0.9761) was between four- and five-date VI zonations from 2019, underscoring the stabilizing impact of intra-annual data richness. The lowest value (−0.2440) occurred between two-date BSI zonations from 2019 and 2020, reflecting greater instability in sparse BSI inputs.

Zonations using 3–5 seasonal dates within the same year and based on a single index produced the most reliable results. While two-class configurations offered the strongest mean agreement, they were also the most variable. Three-class schemes provided a better balance (mean = 0.2155; SD = 0.2422), maintaining coherence with moderate spatial granularity. Increasing class count beyond four led to further decline in agreement, but with reduced variability—indicating more consistent yet less similar outcomes.

These findings confirm that both class number and temporal depth influence zonation consistency. Two- and three-class schemes with multi-date imagery offer an optimal trade-off between spatial clarity and stability.

3.3.3. Multi-Class Zonation Similarity Analysis Using the Jaccard and Dice Coefficients

Spatial similarity among VI- and BSI-based zonations was evaluated using the Jaccard and Dice coefficients across index types and acquisition date combinations (

Table 5. Group-based similarity metrics (mean ± standard deviation) and maximum values for VI and BSI zonation groups.

Group	Jaccard (Mean ± SD)	Jaccard Max	Dice (Mean ± SD)	Dice Max
BSI prior to sowing	0.2412 ± 0.2343	0.8571	0.3641 ± 0.2932	0.9231
BSI after harvest	0.1902 ± 0.2179	0.8493	0.2991 ± 0.2780	0.9189
BSI both dates	0.1293 ± 0.1803	0.7023	0.1962 ± 0.2261	0.8252
VI 1 date	0.1036 ± 0.1600	0.7514	0.1561 ± 0.2064	0.8571
VI 2 dates	0.1247 ± 0.1770	0.7775	0.1861 ± 0.2236	0.8824
VI 3 dates	0.1424 ± 0.1896	0.7811	0.2144 ± 0.2369	0.892
VI 4 dates	0.1486 ± 0.1794	0.8066	0.2227 ± 0.2295	0.8922
VI 5 dates	0.1807 ± 0.2067	0.8264	0.2651 ± 0.2583	0.9052

). The highest similarity was recorded for BSI prior to sowing (Jaccard = 0.2412 ± 0.2343; Dice = 0.3641 ± 0.2932), followed by post-harvest BSI and five-date VI zonations (Dice = 0.2651 ± 0.2583). All BSI configurations outperformed VI-based layers in internal coherence. Within VI groups, similarity increased with the number of acquisition dates, while single-date layers showed the lowest agreement (Dice < 0.20). This highlights the advantage of BSI layers—particularly pre-sowing—when bare soil is exposed, offering greater spatial consistency than vegetation-based metrics.

Overall, BSI-based zonations demonstrated superior consistency, especially where bare soil is visible, whereas multi-date VI layers serve as a viable alternative when soil exposure is limited.

Analysis across class counts (

Table 6. Similarity metrics across class counts (2–8).

Number of Classes	Jaccard (Mean ± SD)	Jaccard Max	Dice (Mean ± SD)	Dice Max
2	0.359 ± 0.206	0.977	0.496 ± 0.214	0.988
3	0.241 ± 0.182	0.977	0.360 ± 0.201	0.988
4	0.184 ± 0.166	0.976	0.285 ± 0.190	0.988
5	0.150 ± 0.153	0.983	0.237 ± 0.180	0.992
6	0.128 ± 0.144	0.976	0.205 ± 0.171	0.988
7	0.111 ± 0.136	0.979	0.181 ± 0.164	0.989
8	0.099 ± 0.129	0.976	0.162 ± 0.157	0.988

) confirmed that two-class zonations yielded the highest similarity (Jaccard = 0.359; Dice = 0.496), with a gradual decline in both metrics beyond four classes. Three- and four-class schemes provided a favorable trade-off between spatial resolution and internal agreement, while higher class counts led to reduced coherence and increased variability. These findings confirm that moderate complexity in class structure ensures better overlap without excessive fragmentation, especially in high-resolution satellite data.

These findings suggest that zonation schemes with 2–4 classes, based on temporally aggregated inputs—particularly BSI prior to sowing and multi-date VI layers—offer optimal stability and interpretability for defining management zones in precision agriculture (Figure 11).

3.4. Evaluation of EM38-MK2 and VI-BSI Data

3.4.1. Correlation Analysis of Raster Layers for Zonation Stability

This analysis explores spatial correspondence between EM38-MK2-derived zonation maps and those generated from satellite-based vegetation (VI) and bare soil indices (BSIs), aiming to evaluate alignment between subsurface conductivity and surface reflectance patterns.

Strong correlations were observed across all resolutions (10–100 m), with the highest values at 50 m and 100 m where smoothing reduced noise. At these scales, Pearson’s r often exceeded 0.80, especially for C05 (4-class), C1 (3-class), and L05 (5-class) configurations, with R² values up to 0.70. Even at 10 m resolution, correlations remained substantial (r = 0.55–0.70), particularly when using multi-date or multi-year composites.

While correlation strength decreased slightly at finer resolutions (10 m), values remained meaningful (r = 0.55–0.70), especially when VI/BSI layers were based on multi-date or multi-year composites. Zonations using 4–6 classes demonstrated greater stability and spatial alignment with VI/BSI, while 2-class models lacked sufficient granularity, and schemes with over 6 classes often introduced fragmentation.

Among all comparisons, the five-date VI layer from 2019 yielded the highest overall agreement with EM38 zones (mean r = 0.84; max r = 0.96). High correlations were also found for C1 and C05 zones at 100 m with VI 2021 and BSI 2020 (r > 0.90). Notably, even the 20th-best correlated layer exceeded r = 0.75. Figure 12 presents the top 20 Pearson correlations between EM38-MK2-based zonation layers (C1 and C0.5) and corresponding VI/BSI classifications at a 100 m grid resolution. All input layers used to generate this figure are documented in Supplementary Table S2, which provides full names and metadata for traceability.

These results confirm the strong spatial coherence between temporally aggregated VI/BSI layers and EM38-derived zones, supporting their integration for constructing robust and scalable data-driven management zones in precision agriculture.

3.4.2. Zonation Consistency Assessment Based on Kappa Coefficient

Zonation consistency between EM38-MK2 conductivity maps and VI/BSI indices was assessed using Cohen’s Kappa coefficient. On average, agreement was low, with mean values across class counts (2–8) near or below zero (

Table 7. Summary of the Kappa metrics for EM38-MK2 vs. VI/BSI zonation pairs across class counts (2–8).

Number of Classes	Mean Kappa	Max Kappa	Min Kappa	Std. Dev.
2	−0.06	0.3672	−0.3548	0.078
3	−0.031	0.1736	−0.1857	0.0506
4	−0.024	0.15	−0.1275	0.0318
5	0.02	0.0913	−0.09	0.0238
6	−0.02	0.0731	−0.0613	0.019
7	0.001	0.0611	−0.0503	0.016
8	−0.001	0.0504	−0.0629	0.0136

). Slightly higher averages appeared in 5–8 class configurations (e.g., mean Kappa = 0.02 for 5 classes), while 2–4 class schemes showed greater variability and a higher frequency of negative values, indicating divergence in spatial structure. However, selected two-class combinations—especially C1C05 vs. multi-date VI/BSI—achieved localized alignment, suggesting targeted integration may be effective in specific scenarios.

Despite overall low agreement, analysis of the top 100 highest Kappa values revealed specific patterns. The C1C05 EM38 configuration dominated (39% of top matches), followed by C1C05L1L05 (25%) and C1 alone (24%). Additionally, 78 of the top 100 values were based on 2-class zonations, confirming their potential for strong alignment under optimal conditions.

BSI layers appeared in 70% of top-ranked matches, particularly those from 2018 and 2019 (together 64%), highlighting their spatial stability. Though VI had a slightly higher average Kappa, BSI proved more prominent among the best-scoring combinations. Notably, all top 10 individual pairings involved 2-class C1C05 EM38 zonations combined with multi-date VI or BSI layers (

Table 8. The 10 representative high-performing pairings of EM38-MK2 and VI/BSI zonation layers.

EM38-MK2 Channel Combination	VI/BSI Combination	Max Kappa
C1C05	3 dates VI 2018	0.3672
C1C05	4 dates VI 2019	0.3231
C1C05	5 dates VI 2019	0.3262
C1C05	4 dates VI 2018	0.3093
C1C05	1 date VI 2019	0.3183
C1C05	3 dates VI 2019	0.3196
C1C05	2 dates VI 2019	0.3174
C1C05	5 dates VI 2018	0.2485
C1C05	2 dates BSI 2021	0.2677
C1C05	2 dates VI 2018	0.286

).

These results indicate that, while general agreement remains weak, specific configurations—especially 2-class C1C05 zones aligned with temporally rich VI/BSI layers—can achieve meaningful spatial consistency. This emphasizes the value of targeted data combinations and temporal aggregation for robust zone delineation in precision agriculture.

3.4.3. Multi-Class Zonation Similarity Analysis Using the Jaccard and Dice Coefficients

Spatial similarity between zonation layers derived from EM38-MK2 conductivity and remote sensing indices (VI, BSI) was evaluated using the Jaccard and Dice coefficients across all agricultural sites, with class counts ranging from 2 to 8. A clear inverse relationship was observed: similarity decreased with increasing class number (Figure 13).

Two-class zonations based on C1C05 profiles exhibited the strongest agreement with surface indices. When compared to multi-date VI from 2018 or two-date BSI from 2021, the Jaccard values exceeded 0.62 and the Dice values reached 0.77, indicating that binary conductivity zones effectively reflect dominant surface patterns. Three-class schemes also performed well, with a Dice value up to 0.71 and a Jaccard value around 0.55, particularly when paired with early-season or aggregated VI/BSI layers.

Although similarity declined in four-class schemes, selected combinations—such as C1C05 with 3-date VI 2018 or 2-date BSI 2021—still produced moderate agreement (Dice ≈ 0.67; Jaccard ≈ 0.51). For class counts above five, both metrics dropped significantly, reflecting increased fragmentation and reduced coherence.

Across all scenarios, the Dice coefficients consistently exceeded the Jaccard coefficient due to higher sensitivity to shared class members, though both followed the same trends. Overall, simpler zonations with 2–4 classes—especially those based on C1C05 conductivity—showed the highest compatibility with VI/BSI-derived zones. While overall agreement was moderate (typically Dice = 0.1–0.4), results underscore the complementary nature of subsurface and surface data for spatial soil variability assessment.

3.5. Evaluation of EM38-MK2, BSP, VI and BSI Data

3.5.1. Correlation Analysis of Raster Layers for Zonation Stability

Pearson’s correlation analysis was conducted to evaluate spatial agreement among zonation layers derived from EM38-MK2, BSP, VI, and BSI datasets at 10, 20, 50, and 100 m resolutions. As presented in

Table 9. Summary statistics of Pearson’s correlations for combined EM38-MK2, BSP, VI, and BSI raster layers.

Grid Resolution	|r| > 0.4	|r| > 0.6	|r| > 0.8	Mean	Median
10 m	1379	486	187	0.583	0.517
20 m	1373	482	188	0.584	0.519
50 m	1477	496	188	0.579	0.522
100 m	1660	534	210	0.576	0.521

, mean absolute correlation values remained consistently above 0.57 across all grids, with minimal variation, and over 530 raster pairs exceeded |r| > 0.6 at each resolution. The 100 m grid yielded the highest number of very strong correlations (210 with |r| > 0.8), suggesting enhanced large-scale coherence due to reduced local noise.

Notably, hybrid layers combining EM38 and BSP data—especially humus, P₂O₅, K₂O, and pH—correlated strongly with multi-date VI and BSI layers, often surpassing r = 0.85 (R² > 0.70). The most stable alignments were observed in classified rasters with 3 to 5 classes and in composites integrating multi-source and seasonal data. Kriging-interpolated rasters were excluded to avoid correlation bias.

Although average correlation across all combinations remained modest (mean ~0.03–0.036), a large number of pairings exhibited strong alignment. At 100 m resolution, 619 raster pairs exceeded |r| > 0.8, compared to 571 at 10 m, highlighting the advantage of spatial aggregation (Figure 14). Full raster names corresponding to the 100 m correlation matrix in Figure 14 are available in Supplementary Table S2.

These results confirm that moderate-to-coarse resolutions, combined with weighted integration of EM38-MK2, BSP, and VI/BSI data, provide a solid foundation for generating consistent and scalable management zones in precision agriculture.

3.5.2. Zonation Consistency Assessment Based on the Kappa Coefficient

Zonation agreement among different data sources was assessed using Cohen’s Kappa coefficient across raster layers derived from EM38-MK2 (EM38_KL), weighted overlays with BSP (EM38_KNT), standalone BSP (BSP_KL, BSP_KNT), and VI/BSI indices (VI/BSI_KNT), covering class counts from 2 to 8. Each group included 10,000–31,000 comparisons depending on parcel size.

As shown in

Table 10. Summary of the Kappa statistics comparing different data groups.

Group Combination	Mean Kappa	Std. Dev.	Min Kappa	Max Kappa
BSP_KNT vs. BSP_KNT	0.948	0.033	0.858	0.997
EM38_KNT vs. EM38_KNT	0.808	0.117	0.483	0.985
EM38_KNT vs. BSP_KNT	0.738	0.124	0.488	0.990
VI/BSI_KNT vs. VI/BSI_KNT	0.552	0.131	0.368	0.995
BSP_KNT vs. VI/BSI_KNT	0.427	0.071	0.304	0.653
EM38_KL vs. BSP_KL	0.353	0.160	−0.037	0.852

, the highest internal agreement was observed within BSP_KNT (mean Kappa = 0.948), followed by EM38_KNT (0.808), and their mutual comparison (0.738), confirming the reliability of WOA-based zonations. VI/BSI_KNT layers showed moderate consistency (mean = 0.552), while quantile-classified layers (EM38_KL and BSP_KL) yielded much lower values (mean < 0.36), indicating weaker internal robustness.

Class count strongly influenced agreement (Figure 15). The best results were consistently observed with 2-class schemes—mean Kappa up to 0.99—especially for EM38_KNT, BSP_KNT, and VI/BSI_KNT. However, consistency decreased with increasing class number; for instance, Kappa for EM38_KNT vs. VI/BSI_KNT dropped from 0.516 (2 classes) to 0.373 (8 classes).

Key findings include the following:

• EM38_KNT vs. BSP_KNT had the strongest cross-source alignment (mean = 0.738), supporting integrated geophysical–chemical zonation.
• Simple classifications (EM38_KL vs. BSP_KL) had much lower agreement (mean = 0.353), underscoring the limitations of univariate approaches.
• VI/BSI_KNT comparisons with other groups were only moderately consistent and mostly in 2–3 class schemes.

These results highlight the benefits of multi-variable weighted models for producing stable, interpretable zonations. Careful calibration of input diversity and class complexity is essential for effective spatial decision-making in precision agriculture.

3.5.3. Multi-Class Zonation Similarity Analysis Using Jaccard and Dice

Spatial similarity between zonation layers from five data sources—EM38, BSP, VI-BSI, BSP weighted, and composite EM38+BSP overlays—was assessed using the Jaccard and Dice coefficients. As presented in

Table 11. Summary of the Jaccard and Dice similarity coefficients across raster group comparisons.

Group Comparison	Jaccard Mean	Jaccard Max	Jaccard Std	Dice Mean	Dice Max	Dice Std
EM38 vs. EM38	0.0959	0.9775	0.1692	0.1409	0.9886	0.2238
EM38 vs. BSP	0.0831	0.6054	0.0976	0.14	0.7542	0.1501
EM38 vs. BSP weighted	0.0863	0.7343	0.105	0.1436	0.8468	0.1594
EM38 vs. Other	0.0827	0.6081	0.0893	0.1414	0.7563	0.1396
EM38 vs. VI/BSI	0.0626	0.6254	0.0693	0.1106	0.7696	0.1125
EM38+BSP vs. EM38	0.085	0.7452	0.1066	0.1411	0.854	0.1599
EM38+BSP vs. EM38+BSP	0.5629	1	0.2507	0.6844	1	0.227
EM38+BSP vs. BSP	0.0871	0.924	0.1473	0.1335	0.9605	0.2
EM38+BSP vs. BSP-weighted	0.0888	0.9445	0.1458	0.1369	0.9715	0.1986
EM38+BSP vs. Other	0.4476	0.9978	0.2606	0.5738	0.9989	0.2509
EM38+BSP vs. VI-BSI	0.0738	0.4732	0.0622	0.1317	0.6424	0.0994
BSP vs. BSP	0.1008	1	0.2187	0.1315	1	0.2635
BSP vs. BSP-weighted	0.1136	0.98	0.2025	0.1582	0.9899	0.2553
BSP vs. VI-BSI	0.0664	0.5591	0.0723	0.1166	0.7172	0.1175
BSP-weighted vs. BSP-weighted	0.1316	0.9984	0.2589	0.166	0.9992	0.2961
BSP-weighted vs. VI-BSI	0.064	0.5953	0.0716	0.1125	0.7463	0.1165
Other vs. BSP	0.0943	0.9538	0.1839	0.1331	0.9764	0.2361
Other vs. BSP-weighted	0.0932	0.9445	0.1712	0.136	0.9715	0.2238
Other vs. Other	0.844	1	0.1202	0.9104	1	0.0772
Other vs. VI-BSI	0.078	0.3935	0.0641	0.1387	0.5648	0.1019
VI-BSI vs. VI-BSI	0.1504	1	0.1648	0.2339	1	0.1958

, the EM38+BSP group achieved the highest internal consistency (mean Dice = 0.6844; Jaccard = 0.5629), indicating that integrated geophysical and chemical data via weighted overlays yield the most coherent spatial outputs. Moderate similarity was also observed within VI-BSI and BSP groups (Dice means of 0.2339 and 0.1660, respectively).

Cross-group agreement was substantially lower. EM38 vs. VI-BSI comparisons showed the weakest spatial alignment (mean Dice ≈ 0.11; Jaccard < 0.07), reflecting fundamental differences between surface reflectance and subsurface conductivity. However, EM38+BSP composites aligned better with OHS-weighted rasters (Dice = 0.5738), suggesting improved compatibility through multi-source integration.

The maximum Dice and Jaccard values approached 1.0 in several within-group and hybrid configurations, especially in the “Other vs. Other” group (Dice = 0.9104), indicating near-identical zones when classification structures are well-calibrated (Figure 16). Key findings include the following:

• EM38+BSP overlays deliver the highest internal similarity.
• VI-BSI zones are consistent internally but poorly aligned with geophysical and chemical layers.
• Cross-domain similarities remain limited, underscoring the unique contributions of each data type.
• High similarity is achievable through optimized weighting and classification design.

These results underscore the importance of multi-source fusion for robust zonation and suggest that optimal models balance internal coherence with inter-source compatibility, tailored to specific agro-environmental contexts.

4. Discussion

4.1. Interpretation of Results in the Context of Existing Literature and Practice

The obtained results clearly demonstrate that different data sources and zonation approaches yield distinct patterns of stability, consistency, and spatial uniformity. Consistent with previous studies [15,17,32], classifications using a smaller number of zones (ranging from two to four) provide a higher degree of spatial similarity and lower inter-annual and inter-method variability. This is particularly important for practical implementation in precision agriculture.

The high stability observed in zones derived from the integration of multiple data sources (EC and BSP) through Weighted Overlay Analysis reinforces earlier findings on the effectiveness of multi-criteria approaches in soil zonation [17]. The results also show that giving more weight to humus leads to better alignment with reference zones, supporting earlier findings [19] that highlight the importance of soil organic matter.

Vegetation indices derived from Sentinel-2 data demonstrated good consistency, particularly in multi-year zonation analyses. This suggests that, with careful selection of acquisition dates and the combination of multiple indices, Sentinel-2 data can serve as a valid alternative for assessing the spatial structure of agricultural land—an observation also supported by previous studies [11,15,18].

Additionally, similar approaches to multi-year zonation evaluation have been applied by Heidari and Samavati [33], who introduced the Zoning Dissimilarity Metric (ZDM) to identify years that significantly disrupt the temporal consistency of zones derived from Sentinel-2 NDVI time series. Their analysis shows that excluding so-called “abnormal years,” identified through statistical deviations in their impact on zonation, can substantially enhance the accuracy and robustness of resulting management zones. This underscores the importance of careful date selection and temporal window definition when using satellite-derived indices for zonation purposes.

Although this study primarily employed the Kappa, Jaccard, and Dice coefficients to assess zone stability, future research could incorporate metrics such as ZDM to provide a more quantitative measure of the impact individual years have on zonation patterns.

4.2. Advantages and Limitations of Different Zonation Methods

The use of interpolation methods such as kriging enables the generation of smooth, continuous surfaces suitable for spatial analysis, but their performance is highly sensitive to the distribution and density of sampling points. In contrast, percentile-based classification provides a straightforward categorization of values but does not account for spatial context.

Fuzzy k-means proved to be flexible, letting each pixel belong partly to more than one class while still preserving spatial continuity. Its key advantage lies in its ability to express gradual class membership, though it requires longer processing times and careful parameter tuning.

Weighted Overlay Analysis (WOA) offers a high degree of adaptability and the capacity to integrate multiple data sources. However, the resulting zones are highly sensitive to the selection of weight coefficients, which, if not properly calibrated, can lead to significant variation in the final outputs. For this reason, cautious application and, where possible, validation using reference laboratory data or field measurements are recommended.

Vegetation and bare soil indices (BSIs) enable rapid and cost-effective zonation but are subject to seasonal and interannual variability, as well as the influence of cloud cover and atmospheric conditions. Therefore, their use is best complemented with additional data sources to enhance reliability.

4.3. Recommendations for the Optimal Selection of Data Sources and Methods

Based on the conducted analysis, the following recommendations can be made:

• For practical implementation in precision agriculture, an optimal number of management zones lies between three and four, as this provides a balanced trade-off between spatial informativeness and zone stability.
• The use of EM38-MK2 data—particularly channel C1 or the combination of C1 and C0.5—proved to be the most effective in detecting soil variability.
• BSP data, especially when all analyzed parameters are integrated, yield reliable results and can serve as a reference for calibrating other methods.
• VI and BSI derived from Sentinel-2 imagery are useful as complementary inputs, particularly when generated from multiple acquisition dates and weighted appropriately.
• The Weighted Overlay Analysis (WOA) method demonstrated high flexibility and produced the most robust results when humus played a dominant role in the weighting scheme.
• Raster normalization and the use of spatial resolutions between 20 and 50 meters provide an optimal balance between spatial detail and zone stability.

These recommendations offer concrete guidelines for selecting appropriate methodologies based on specific objectives, data availability, and budget constraints in the context of precision soil management.

4.4. Added Value of Multi-Source Correlation Analysis

The correlation matrices between raster layers—particularly those integrating EM38, BSP, and VI/BSI data—revealed important synergies and limitations in the spatial structure captured by each source. While normalized and interpolated EM38 layers consistently showed strong internal coherence, their correlation with satellite-derived indices was the highest when both data types were aggregated over multiple temporal images. Although time series data require more processing, they help identify stable biophysical patterns beyond just seasonal effects.

A critical observation was that correlation matrices can inform not only zonation stability, but also highlight complementary data sources, as low cross-correlations may reflect distinct underlying spatial processes. Careful matrix interpretation is thus essential to avoid over-integration or misrepresentation of spatial coherence.

4.5. Toward More Interoperable and Transferable Zonation Models

An open question arising from this study is the transferability of the delineation approach across geographies and farming systems. Given the strong internal agreement in weighted overlays (especially humus-dominant schemes), there is an opportunity to adapt this methodology to other regions, such as Dutch polder landscapes, where organic matter and moisture gradients also play a key agronomic role.

To support future replication, the methodological pipeline presented here—combining proximal sensing, remote sensing, and fuzzy clustering—could be adapted into a decision-support module.

4.6. Broader Implications for Decision-Making in Data-Rich and Data-Poor Contexts

This study reinforces the emerging understanding that precision agriculture requires context-specific approaches rather than universal models. In data-rich environments, fully integrated models using all three sources (EC, BSP, VI/BSI) can generate nuanced and robust zones. Conversely, in data-scarce settings, carefully selected VI/BSI composites or single-depth EC readings (e.g., C1) can still yield sufficiently stable zonation maps, enabling broader adoption in smallholder or transitional farming systems.

4.7. Limitations of the Study

While the results presented here demonstrate consistent and interpretable management zone delineation across multiple data sources, several limitations must be acknowledged. First, Sentinel-2-based indices are sensitive to atmospheric interference, cloud cover, and vegetation phenology, which can introduce variability in inter-annual comparisons. Although multi-date compositing helps reduce these effects, some residual noise may persist. Second, interpolation of BSP data is constrained by the spatial density and distribution of samples. In areas with heterogeneous soil structure or sparse sampling (e.g., >5 ha per sample), kriging may oversmooth the data or misrepresent local variability. Finally, classification results depend on selected parameters (e.g., number of classes, weights in WOA), which may limit transferability to other contexts unless locally calibrated. Future work should further explore uncertainty quantification and sensitivity analysis in zonation models.

4.8. Overcoming Operational and Methodological Constraints

This study addressed several key operational and methodological challenges commonly encountered in precision agriculture. By combining data from EM38-MK2 sensors, basic soil properties (BSPs), and Sentinel-2-derived indices, we demonstrated that stable and interpretable management zones can be delineated without the need for high-density sampling or complex ground campaigns. The use of optimized weighting schemes, fuzzy clustering, and temporal composites further enhanced the reproducibility and cost-efficiency of the zonation process.

Moreover, this study identified that a zonation scheme using three to four classes represents an optimal trade-off between spatial detail and operational feasibility. This level of complexity aligns with field-level management capacity, ensuring that resulting maps are both agronomically informative and practically implementable. These findings support the development of practical, scalable methods for soil and crop management that can be adapted to different production settings.

5. Conclusions

This study evaluated multiple approaches for delineating within-field management zones by integrating geophysical (EM38-MK2), chemical (basic soil properties), and spectral (VI/BSI from Sentinel-2) data. Through a multi-step analytical framework—encompassing interpolation, classification, and validation using the Kappa, Jaccard, Dice, and correlation metrics—we demonstrated that the most stable and internally consistent zonations were achieved using weighted combinations, particularly those dominated by humus content. Two- and three-class configurations generally provided the best trade-off between spatial coherence and agronomic interpretability.

A critical observation was that correlation matrices can inform not only zonation stability but also highlight complementary data sources, as low cross-correlations may reflect distinct underlying spatial processes. Careful matrix interpretation is thus essential to avoid over-integration or misrepresentation of spatial coherence.

The findings also reinforce the evolving perspective that precision agriculture must be tailored to specific agronomic and data contexts. In data-rich environments, robust multi-source models are feasible and desirable, allowing for complex, high-resolution zonation. In contrast, in data-scarce scenarios—such as smallholder or transitional farming systems—simpler inputs like two-date BSI composites or single-depth EC readings (e.g., C1) still offer viable pathways toward spatial differentiation and site-specific management.

Additionally, the strong internal agreement within humus-weighted overlays suggests potential for adapting this framework beyond the current study area. The approach may be transferable to other regions with similar soil-climate conditions, such as lowland agricultural regions characterized by high groundwater tables, organic-rich soils, and structured drainage systems common in parts of Northwestern Europe. To support future replication, the presented methodology—based on proximal sensing, remote sensing, and fuzzy clustering—can be modularized into open-source decision-support tools, enabling calibration across sites and farming systems.

Overall, this research highlights the value of integrating heterogeneous data sources and applying scalable analytical workflows for generating actionable management zones. Continued development of flexible, context-aware zonation models remains essential for advancing precision agriculture across diverse landscapes.