Previous Article in Journal
Upgraded Protocol for Microplastics’ Extraction from the Soil Matrix by Sucrose Density Gradient Centrifugation
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Soil Systems
Volume 9
Issue 3
10.3390/soilsystems9030067
soilsystems-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Field Evaluation of a Portable Multi-Sensor Soil Carbon Analyzer: Performance, Precision, and Limitations Under Real-World Conditions

by
Lucas Kohl
1,2,*,
Clarissa Vielhauer
3,
Atilla Öztürk
4,
Eva-Maria L. Minarsch
1,
Christian Ahl
3,
Wiebke Niether
1,
John Clifton-Brown
2 and
Andreas Gattinger
1
1
Department of Agronomy and Plant Breeding II, Organic Farming with Focus on Sustainable Soil Use, Justus-Liebig University Giessen, Karl-Gloeckner-Strasse 21 C, 35394 Giessen, Germany
2
Department of Agronomy and Plant Breeding I, Crop Biomass and Bioresources, Justus-Liebig University Giessen, Karl-Gloeckner-Strasse 21 C, 35394 Giessen, Germany
3
Institute of Soil Science, University of Göttingen, Von-Siebold-Str. 4, 37075 Göttingen, Germany
4
Institute of Ecology, Chair of Ecology and Environmental Planning, Technical University of Berlin, Straße Des 17. Juni 145, 10623 Berlin, Germany
*
Author to whom correspondence should be addressed.
Soil Syst. 2025, 9(3), 67; https://doi.org/10.3390/soilsystems9030067 (registering DOI)
Submission received: 5 May 2025 / Revised: 9 June 2025 / Accepted: 19 June 2025 / Published: 27 June 2025
Browse Figures
Versions Notes

Abstract

Soil organic carbon (SOC) monitoring is central to carbon farming Monitoring, Reporting, and Verification (MRV), yet high laboratory costs and sparse sampling limit its scalability. We present the first independent field validation of the Stenon FarmLab multi-sensor probe across 100 temperate European arable-soil samples, benchmarking its default outputs and a simple pH-corrected model against three laboratory reference methods: acid-treated TOC, temperature-differentiated TOC (SoliTOC), and total carbon dry combustion. Uncorrected FarmLab algorithms systematically overestimated SOC by +0.20% to +0.27% (SD = 0.25–0.28%), while pH adjustment reduced bias to +0.11% and tightened precision to SD = 0.23%. Volumetric moisture had no significant effect on measurement error (r = −0.14, p = 0.16). Bland–Altman and Deming regression demonstrated improved agreement after pH correction, but formal equivalence testing (accuracy, precision, concordance) showed that no in-field model fully matched laboratory standards—the pH-corrected variant passed accuracy and concordance evaluation yet failed the precision criterion (p = 0.0087). At ~EUR 3–4 per measurement versus ~EUR 44 for lab analysis, FarmLab facilitates dense spatial sampling. We recommend a hybrid monitoring strategy combining routine, pH-corrected in-field mapping with laboratory-based recalibrations alongside expanded calibration libraries, integrated bulk density measurement, and adaptive machine learning to achieve both high-resolution and certification-grade rigor.

1. Introduction

Soil organic carbon (SOC) constitutes the principal fraction of soil organic matter (SOM) and underpins critical soil functions such as nutrient cycling, water retention, and crop productivity, while also sustaining biodiversity and ecosystem resilience [1]. As a dynamic reservoir within the global carbon cycle, the accumulation of SOC has the capacity to sequester atmospheric carbon dioxide (CO2), thereby contributing to climate change mitigation [2,3]. In light of this potential, the European Commission’s Carbon-Farming Initiative under the European Green Deal now offers incentives to land managers for practices that increase SOC stocks, provided that their field measurements are both reliable and cost-effective for robust certification [4,5,6,7].
Conventionally, the quantification of SOC is performed through the utilization of laboratory-based approaches, including dry-combustion elemental analysis and mid-infrared (MIR) and near-infrared spectroscopy (NIRS). These methods are known for their high accuracy (R2 ≥ 0.98), and once samples are dried, homogenized, and ground, they enable non-destructive, high-throughput SOC analysis—e.g., MIR spectroscopy can process samples in ≤1 min [8,9,10]. However, they require extensive sample preparation, homogenization, and calibration against large reference datasets. Laboratory MIR/NIRS protocols have been developed to minimize spectral interference from moisture, texture, and mineralogy. This is achieved through the stringent control of sample moisture content and particle size [11,12,13]. Despite the fact that these techniques remain the gold standard, their laborious workflows and per-sample cost of approximately EUR 44 limit their feasibility for frequent, large-scale on-farm monitoring [14].
In order to address these limitations, a considerable number of studies have explored the potential of portable in situ sensing platforms that integrate optical, electrochemical, and environmental measurements. Early on-the-go visible/NIR systems have been shown to possess both potential and limitations with regard to the mapping of soil clay and SOC [15,16]. Subsequent reviews have documented the progression from rudimentary sensors to sophisticated benchtop and mobile instruments enhanced by machine learning algorithms [17,18]. Field evaluations under disturbance-reduced protocols have reported improved estimates of SOC [19], and multi-sensor probes, coupled with advanced chemometrics, have yielded robust predictions of soil profile properties [20,21]. Recent investigations into miniaturized spectrometers have confirmed their potential for rapid soil property assessment [22], while the application of unsupervised learning to regional Vis-NIR spectral libraries has further enhanced the prediction of organic carbon [23]. Innovations such as moisture- and salinity-correction algorithms [24,25,26] and rapid in situ CO2-sensor methods [27] continue to expand the field.
The multi-sensor FarmLab device from Stenon integrates visible/NIR reflectance, electrical impedance spectroscopy (EIS), and environmental sensors—including soil moisture, temperature, and volatile organic compounds (VOCs)—within a handheld spade probe for measurements in the upper 15 cm of soil [28,29]. Moreover, FarmLab offers dramatic cost savings—approximately EUR 3–4 per measurement versus roughly EUR 44 per laboratory sample—enabling much higher sampling densities at a fraction of the cost. Despite these economic advantages, independent assessments have indicated that its accuracy and precision remain lower than laboratory standards. Residual biases have been shown to be influenced by soil pH and texture [30,31,32]. A recent comparison to another Vis-NIR multi-sensor platform has further highlighted its current limitations [33].
Although the manufacturer reports deployments of the FarmLab device in Germany, Brazil, Kazakhstan, Kyrgyzstan, and California—including over 450 calibration sites in Germany and more than 6000 calibration measurements in Brazil [34,35]—no independent field evaluation has yet assessed the performance of the FarmLab device under temperate European arable conditions. In particular, there is a clear gap in the literature for research comparing its SOC estimates to established laboratory methods—acid-treated total organic carbon (TOC-acid), temperature-differentiated TOC (SoliTOC), and total carbon analysis—and examining how its integrated moisture and pH sensors mitigate field-sensor artifacts. Accordingly, the present study aims to (1) quantify the accuracy and precision of FarmLab SOC measurements against these laboratory standards, (2) evaluate the effectiveness of its onboard moisture and pH corrections, (3) determine its suitability for carbon-farming applications.

2. Materials and Methods

2.1. Sampling Sites

Soil sampling was conducted at nine temperate arable sites in Lower Saxony (May 2021) and Hesse (August–September 2021). Site selection was based on their inclusion in broader research projects and their representation of a wide range of soil types typical of central Germany’s arable land, while also falling within the SOC measurement range specified by the FarmLab device manufacturer [28]. At each location, one to four soil samples were collected from agricultural fields. Three sites (Sites 1–3) were located in Lower Saxony and sampled in May 2021, while six sites (Sites 4–9) were situated in Hesse and sampled between August and September 2021 (see Figure 1 for details).
Sites 1 and 2 were located near Braunschweig and lay only about 0.6 km apart. Site 3 was situated approximately 97 km southwest of Site 2, in the northwestern vicinity of Kassel.
In Hesse, Site 4 and Site 5 were located about 85 km and 94 km south/southeast of the preceding sites, respectively. Site 6 lay roughly 63 km west of Site 5, and Site 7 was located less than 1 km from Site 6. Site 8 was situated approximately 34 km southwest of Site 6, and Site 9 about 66 km southwest of Site 8.
A total of 20 plots were sampled, with each plot encompassing 5 georeferenced subplots, resulting in a total of 100 subplots. Within each subplot, two adjacent FarmLab measurements were made and five soil cores (0–30 cm) were collected within 0.5 m of the probe position. To facilitate the subsequent laboratory analysis, the cores were amalgamated into composite samples for each subplot. At the Hesse sites, undisturbed cores were also taken at two depths to determine bulk density; in Lower Saxony, bulk density was estimated from texture and SOC following Rawls [36].

2.2. Sampling Scheme

At each site, the sampling plots were laid out according to the following protocol. In Hesse (Sites 4–9), one square plot measuring 2 m2 contained five subplots arranged in a 2 × 2 m grid. In Lower Saxony (Sites 1–3), where sampling had been conducted independently and prior by the research group of C. Ahl (University of Göttingen, unpublished data), larger fields featured transects of 20 subplots arranged linearly across four plots at Sites 1 and 2 and seven plots at Site 3 (Figure 2). Within each subplot, two independent in-field FarmLab measurements were taken, each consisting of three rapid sub-readings. Subsequently, five soil cores (0–30 cm) were collected with an auger within a 0.5 m radius of the probe insertion point and amalgamated into a composite sample for subsequent laboratory analysis.

2.3. FarmLab In-Field Measurements

The in-field soil moisture content measurements were conducted utilizing the Stenon FarmLab portable multi-sensor probe (software version d-1.3.0; calibration model P-2.1.0). The device integrates visible/NIR reflectance (400–2500 nm), electrical impedance spectroscopy (EIS), soil temperature, volumetric moisture, atmospheric humidity, volatile organic compound (VOC) resistance, and GPS reference at the probe tip [28,29]. Operators inserted the spade-mounted probe vertically until a laser-etched collar marked that a 15 cm depth was reached, in order to ensure consistency. In accordance with the manufacturer’s protocol, the instrument underwent zero-calibration prior to each subplot measurement [33,35]. It is important to note that each point measurement was the mean of three sub-readings. Furthermore, the two point measurements per subplot were then averaged to yield one SOC value. FarmLab’s proprietary algorithm applies internal moisture correction based on its on-board soil-moisture sensor, thereby effectively compensating for moisture-induced spectral artifacts [30]. Due to the confidential nature of the calibration functions, no independent spectral modeling was conducted [19,37].

2.4. Laboratory Analysis

Samples composed of composite materials were subjected to air-drying at a temperature of 40 °C. Thereafter, they were sieved to a size of less than 2 mm and subdivided into four aliquots. Two aliquots were analyzed for total carbon (TC) by dry combustion at 1140 °C using a Vario Max Cube elemental analyzer with thermal conductivity detection and helium carrier gas, once at Justus-Liebig University Giessen (TC-Gi) and once at Georg-August University Göttingen (TC-Goe) (DIN 13878; [38]). A third aliquot was analyzed by the SoliTOC Cube elemental analyzer (Elementar, Langenselbold, Germany) using temperature-dependent oxidation to separate thermally labile organic carbon (<400 °C) and residual oxidizable carbon (500–600 °C). The results were summed as TOC (DIN 19539; [39]). Inorganic carbon (TIC900) was quantified on the same instrument via thermal decomposition at 900 °C. The fourth aliquot underwent inorganic-carbon removal via HCl fumigation, followed by combustion at up to 1500 °C on an Eltra Helios C/S device at Agrolab GmbH (TOC-acid; [40]). This step ensured the complete removal of carbonates, so that only organic carbon was quantified during combustion. The arithmetic mean of SoliTOC and TOC-acid was defined as the Standard-TOC reference for in-field comparisons.
Soil pH from each subplot was measured in a 0.01 M CaCl2 solution [39]. Clay content was determined for a subset of sampling positions (n = 22) using the hydrometer method [41], based on data from earlier, unpublished measurement campaigns.
Bulk density in Hesse was determined from undisturbed stainless steel cylinder cores at two depths per subplot that were oven-dried at 105 °C and weighed [42]. In Lower Saxony, the estimation of bulk density was conducted through the utilization of the Rawls pedotransfer function, which utilizes clay content and SOC as the primary variables [36].

2.5. Methods Used for Comparison

In order to evaluate FarmLab’s performance against established approaches, eight different SOC determination methods were compared (see Table 1). Throughout the manuscript, “SOC” refers to soil organic carbon as the target parameter. “TOC” and “TC” are used only when referring to specific laboratory methods, whose outputs serve as SOC estimates but retain their original names for methodological clarity. Two laboratory dry-combustion methods (TC-Gi, TC-Goe) measured total carbon (TC), whilst two temperature-differentiated methods (SoliTOC, TOC-acid) quantified total organic carbon (TOC) with distinct separation of organic fractions. Furthermore, four FarmLab in-field outputs (In-field-TOC-1 through In-field-TOC-4) represent successive algorithm versions, including a pH-adjusted model (In-field-TOC-4). The arithmetic mean of SoliTOC and TOC-acid was defined as Standard-TOC for all pairwise comparisons. Statistical equivalence, bias, and precision were assessed via Bland–Altman plots, Deming regression, and equivalence tests (eirasBA package [43]).

2.6. Data Analysis

All data were analyzed in R 4.2 [44]. FarmLab SOC outputs were matched to subplot Standard-TOC values via their unique FarmLab ID. Measurement error (FarmLab—Standard-TOC) was computed for each subplot. Pearson’s correlation coefficients were calculated for error versus soil moisture (measured with FarmLab) as well as pH (CaCl2 method [41]). In addition, Pearson correlation was used to quantify the linear association between the absolute SOC values from the in-field models and the Standard-TOC reference.
Bland–Altman analyses (blandr package [45]) quantified the mean bias and 95% limits of agreement, and Deming regression (deming package [46]) assessed structural accuracy and precision. In addition, we computed three complementary error metrics for each method pair: the mean absolute error (MAE), root mean square error (RMSE), and Nash–Sutcliffe efficiency coefficient (NSE) [47].
We further refined our method-comparison framework by applying the extended Bland–Altman tests in three formal steps, implemented via the eirasBA package with 30,000 bootstrapped resamples for robust confidence intervals [43]. In brief, for each pairwise comparison of SOC methods we tested the following:
Structural Mean Equality (Accuracy).
We tested whether the average difference between two methods was statistically zero by fitting an analysis of covariance (ANCOVA) with the reference method as the covariate and treatment method as the factor, incorporating measurement error as described by Hedberg & Ayers [48]. The rejection of the null hypothesis indicated a systematic bias in mean SOC estimates.
Structural Variance Equality (Precision).
We examined whether the variability of measurement errors differed between methods by comparing error variances via regression-based variance tests [49,50]. A significant result denoted unequal precision.
Agreement with the True Bisector (Concordance).
Using Deming regression [46], we tested whether the intercept equaled zero and the slope equaled one—i.e., whether the two methods lied on the 1:1 identity line. Failure to reject either parameter’s null hypothesis indicated concordance in both scale and location.
All three tests were run at a 5% significance level, and 95% confidence intervals for means, variances, and regression parameters were obtained via bootstrapping [51]. The outcomes from these tests were synthesized to determine which methods met all the criteria for equivalence in accuracy, precision, and concordance.

3. Results

Standard TOC ranged from 0.87% to 1.76% (1.17% ± 0.02 SE), and soil pH from 5.7 to 7.6 (6.57 ± 0.05 SE). The skewness and kurtosis were 1.13 and 1.73 for Standard-TOC and 0.20 and −1.07 for pH, respectively. The distribution and spread of these and other variables are shown in Figure 3.

3.1. Descriptive Statistics of SOC Methods

The descriptive statistics for all eight SOC determination methods are summarized in Table 2. Across the three uncorrected FarmLab algorithms (In-field-TOC-1–3), SOC was on average overestimated by +0.24% relative to Standard-TOC (mean bias +0.20–0.27%). Incorporating pH correction (In-field-TOC-4) cut that bias roughly in half (to +0.11%) and reduced the pooled standard deviation from 0.27% to 0.23%. By comparison, the two dry-combustion labs (TC-Gi, TC-Goe) differed from Standard-TOC by only +0.04–0.06% (SD ≈ 0.20%). These results confirm that In-field-TOC-4 is the most unbiased and precise in-field algorithm under our conditions.

3.2. Correlation of SOC Error with Soil Properties

We assessed whether the subplotwise SOC measurement error of the baseline FarmLab algorithm (In-field-TOC-1), defined as In-field-TOC-1 SOC minus Standard-TOC (SOC_error), was influenced by soil pH, carbonate content (TIC900), or volumetric soil moisture, all recorded in our dataset. Pearson’s correlation coefficients (n = 100) are presented in Table 3.
The negative correlation between SOC_error and pH (r = −0.39, p < 0.01) indicates that lower-pH soils tend to produce larger positive errors (overestimation by FarmLab), whereas higher-pH soils yield smaller biases. In contrast, no significant relationship was found between SOC_error and soil moisture (r = −0.14, p > 0.05), suggesting that FarmLab’s integrated moisture sensor effectively compensated for moisture-induced spectral artifacts.
A visual comparison of predicted and measured SOC values across the pH gradient (Figure 4) reveals distinct patterns related to soil acidity. The correlation between Standard-TOC and In-field-TOC-1 was r = 0.13, while the pH-corrected In-field-TOC-4 showed a stronger association, with r = 0.39.
In the uncorrected model (Figure 4A), SOC tends to be overestimated at low pH and low SOC levels (below 1.2%), consistent with the negative correlation reported in Table 3. After applying the pH correction (Figure 4B), these deviations are reduced, indicating improved accuracy in more acidic soils.
At higher SOC concentrations (>1.5%), overestimation persists in samples with higher pHs. Overall, the highest agreement between predicted and measured SOC is observed at moderate to low pH values (approximately 6.0–6.8) and SOC concentrations between 1 and 1.5%.

3.3. Method Agreement Evaluation Using Regression, Error Metrics, and Bland–Altman Analysis

To assess structural agreement and systematic bias between each in-field algorithm and the Standard-TOC reference, we first conducted Deming regression analyses (Table 4). The intercept and slope of a Deming analysis fit quantify location and scale agreement: an ideal method lies exactly on the identity line (intercept = 0, slope = 1). In our comparisons, In-field-TOC-4 showed the closest proximity to these ideal parameters (intercept = 0.05 ± 0.06%, slope = 0.97 ± 0.04), followed by In-field-TOC-1 (intercept = 0.18 ± 0.07%, slope = 1.10 ± 0.05). The uncorrected algorithms (TOC-2, TOC-3) exhibited larger intercepts and slopes further from unity, indicating both constant and proportional bias. By contrast, the laboratory methods TC-Gi and TC-Goe yielded intercepts and slopes statistically indistinguishable from (0, 1), confirming strong equivalence between the two dry-combustion laboratories.
In addition to structural comparison, absolute and squared deviations from the Standard-TOC reference were evaluated by computing the MAE, RMSE, and NSE coefficient (Table 5). In-field-TOC-4 yielded lower MAE and RMSE values than In-field-TOC-1, indicating reduced deviation magnitudes. The NSE values for all the in-field methods and one lab replicate (TC-Gi) were negative, reflecting systematic deviations from the Standard-TOC reference across the full range of observations. Only TC-Goe achieved an NSE near zero, consistent with the minimal bias observed in the Bland–Altman analysis.
Next, Bland–Altman plots quantify the mean bias and 95% limits of agreement (LoAs) between methods (Figure 5; Table 6).
In-field-TOC-1 exhibited a mean positive bias of +0.20% SOC and a wide LoA (−0.35 to +0.75%), whereas In-field-TOC-4 reduced both the bias (+0.11%) and LoA (−0.27 to +0.49%), reflecting improved equivalence. Laboratory replicates TC-Gi and TC-Goe had negligible bias (+0.05%) and narrow LoAs (−0.12 to +0.22%), underscoring their mutual consistency.
Together, these pairwise comparisons demonstrate that the pH-adjusted algorithm (In-field-TOC-4) achieves the best overall alignment with laboratory standards, substantially reducing both constant and proportional errors, while uncorrected in-field methods retain significant biases.

3.4. Inferential Comparison of SOC Methods

The equivalence between each SOC method and the Standard-TOC reference was evaluated using the all.structural.tests function in the eirasBA package with 30,000 bootstrap resamples [43]. Three criteria were evaluated at α = 0.05: structural mean equality (accuracy), variance equality (precision), and bisector agreement (concordance). The results (Table 7) reveal that no method satisfied all three tests simultaneously.
Both laboratory comparisons (TC-Gi vs. TC-Goe and SoliTOC vs. TOC-acid) achieved precision (p > 0.05) and concordance (intercept/slope p > 0.05) but failed the accuracy test (p < 0.01). The baseline FarmLab algorithm (In-field-TOC-1) similarly passed precision and concordance yet exhibited significant bias (accuracy p < 0.0001). The two uncorrected in-field updates (In-field-TOC-2/3) managed to pass only concordance (p > 0.05) but failed in both accuracy and precision (p < 0.0001). Although the pH-corrected model (In-field-TOC-4) met the accuracy criterion (p = 0.3250) and showed concordance (p > 0.05), it failed the precision test (p = 0.0087), indicating unequal variance between FarmLab and the reference measurements.

4. Discussion

4.1. Key Insights and Their Implications

Our comparison of eight SOC methods confirms that the default FarmLab chemometric model consistently overestimates soil carbon, while a straightforward pH correction cuts that bias roughly in half and yields precision approaching laboratory standards. Crucially, the absence of a moisture effect demonstrates that FarmLab’s integrated humidity sensor successfully neutralizes one of the biggest hurdles in field spectroscopy. In contrast, the persistent pH–error relationship highlights acidity as a primary driver of spectral artifacts and reinforces findings from others that acidity must be explicitly accounted for in proximal sensing models [12,52].
From a methodological standpoint, our Deming and Bland–Altman analyses show that even the pH-adjusted variant cannot fully replicate laboratory precision. This “residual variance” underscores the need for further model refinement—perhaps by integrating additional soil covariates such as texture or carbonate content—and suggests that any in-field SOC sensor must be embedded in a broader calibration framework.
Practically, these insights point the way to a hybrid approach: users can leverage FarmLab’s low cost and moisture robustness for the high-density mapping of relative SOC patterns, but anchor critical change-detection decisions to periodic laboratory benchmarks and expanded covariate calibration. Such a strategy marries the speed and affordability of field sensors with the accuracy and rigor required for carbon farming Monitoring, Reporting, and Verification (MRV).

4.2. Accuracy and Precision of FarmLab

While laboratory-grade SOC measurements via dry combustion typically achieve accuracies within ±0.15% SOC [49], our uncorrected FarmLab outputs exceeded this threshold, reflecting the well-documented challenge of translating benchtop calibrations into field environments. Loria et al. [31] reviewed numerous handheld in situ SOC probes and emphasized that multi-sensor fusion—such as combining VIS–NIR with electrical conductivity or impedance—can substantially improve prediction accuracy by compensating for individual sensor limitations, provided that extraneous influences are removed via preprocessing (e.g., EPO or SNV) and that key covariates are incorporated into calibration models. In our study, a simple pH adjustment halved the mean bias (from +0.20–0.27% down to +0.11%) and tightened variability (SD from ~0.27% to 0.23%), bringing FarmLab closer to the Giessen and Göttingen dry-combustion laboratories. Additional error metrics support this interpretation: the RMSE for In-field TOC 4 dropped to 0.32%, and MAE to 0.24%, compared to 0.38% (MAE = 0.16%) for TC Gi and 0.17% (MAE = 0.10%) for TC Goe. These values are similar to those reported by Sanderman et al. [53], who achieved RMSE values of approximately 0.24% using mid-infrared spectroscopy on homogenized and dried soil samples in controlled laboratory conditions. The low or negative Nash–Sutcliffe efficiency (NSE) values for all method pairs except TC Goe (NSE ≈ 0.00) suggest that even between laboratories using identical dry-combustion procedures on homogenized samples, small systematic differences may persist. The Pearson correlation coefficients between the in-field and laboratory values remained low (r = 0.13 for In-field-TOC-1; r = 0.39 for TOC-4), although it is important to note that correlation can be misleading when methods differ in scale or when the observed range is narrow [54,55,56]—as is likely the case here given the compressed SOC range (Figure 3). Notably, Pearson correlation (Figure 4) suggests that FarmLab tends to overestimate lower SOC values and underestimate higher ones, whereas Bland–Altman plots and residual trends (Figure 5) indicate that the largest deviations for in-field methods occur at higher SOC levels (>1.5%), suggesting decreasing model reliability in this upper range. This pattern may help explain why Stenon GmbH specifies a technical upper limit of 3% SOC for FarmLab’s measurement range [30]—although our results imply that practical reliability may already decline beyond ~1.5%, potentially warranting a more conservative threshold.
Such non-uniform performance across the SOC spectrum complicates equivalence testing, and indeed, formal assessments still flagged a precision shortfall (p = 0.0087). While accuracy and concordance were achieved, the spread of individual predictions by FarmLab remained slightly more variable than the laboratory reference. In practical terms, this could limit confidence in single-point estimates under heterogeneous field conditions, although systematic bias appears to be corrected for. This pattern aligns with Angelopoulou et al. [13], who found that even multi-sensor fusion schemes require the explicit inclusion of pH, texture, and other site-specific covariates to achieve laboratory-grade concordance.

4.3. Influence of Soil pH and Moisture on In-Field SOC Estimates

Our results confirm that FarmLab’s integrated moisture sensor and internal correction algorithm effectively alleviate the common spectral artifacts caused by variable soil water content. The non-significant correlation between In-field-TOC-1 error and volumetric moisture (r = −0.14, p = 0.16) aligns with findings by Vikuk et al. [30], who reported minimal moisture bias when combining NIR and EIS sensors in situ [30]. This robustness to moisture fluctuations reduces the need for extensive field drying or gravimetric moisture correction, streamlining the sampling workflow and lowering operational costs.
In contrast, soil pH emerged as a dominant driver of residual bias. The strong negative correlation between SOC error and pH (r = −0.39, p < 0.01) reflects the well-known effects of acidity on optical and electrochemical sensor responses [12,51,57]. Low-pH soils likely alter the chemical speciation and light-scattering properties of organic matter, resulting in systematic overestimation by unadjusted models. By incorporating an empirically derived pH correction, we reduced mean bias by nearly 50% (from +0.20% to +0.11%) and narrowed the limits of agreement—an improvement consistent with recommendations to include pH as a calibration covariate in proximal sensing applications [58].
Moreover, Vogel et al. [59] demonstrated that the performance of in situ pH sensors—and hence the effectiveness of pH-based SOC corrections—depends critically on calibration sample size and the spatial and temporal proximity of reference samples [59]. This implies that FarmLab’s pH correction factors should be periodically updated using locally collected calibration samples to maintain optimal accuracy. However, the residual pH-related bias observed in the color-coded scatterplots (Figure 4) suggests that the empirically derived correction, while effective, may not fully capture the underlying causes of systematic error—particularly at higher SOC and pH values. This could reflect complex interactions between organic matter and the mineral matrix that are not adequately modeled using pH alone. In particular, elements such as Fe, Ca, and Al—which have been associated with both pH buffering and organo-mineral stabilization processes under specific soil conditions [60,61]—may influence sensor response patterns in ways not captured by vis-NIR spectroscopy, as these elements are not directly detectable in this range. Their omission may therefore limit the model’s ability to fully correct for pH-related distortions. O’Rourke et al. [62] demonstrated that integrating X-ray fluorescence-derived elemental information with vis-NIR data significantly improved predictive accuracy for both SOC and pH, reducing RMSE by up to 21% for pH and 18% for SOC through model averaging techniques. Incorporating such complementary sensor data may therefore offer a more robust correction strategy by explicitly accounting for the underlying geochemical drivers of bias, rather than relying solely on empirical adjustment.
Where available, soil buffering capacity and carbonate content (TIC900) could further refine these corrections [63], although our data showed no significant direct influence of TIC900 alone (r = −0.10, p = 0.31). Likewise, soil porosity or water-filled pore space (WFPS) may impact both optical and impedance measurements by altering light scattering and conductive pathways in the pore network [30]. Given that salinity is also known to distort spectral SOC estimates by affecting the ionic strength and dielectric properties of the soil matrix, the inclusion of EC data from FarmLab’s impedance sensor, combined with an optimal band combination algorithm, may offer a means to detect or correct for such effects in future model iterations [26]. These considerations further underscore the need for algorithm development to incorporate multivariate calibration frameworks that jointly account for pH, texture, porosity/WFPS, salinity, and other site-specific factors, as demonstrated by multisensory fusion approaches in the literature [18,31]. This targeted adjustment strategy preserves FarmLab’s moisture resilience while addressing its remaining pH-driven limitations.

4.4. Applicability in Carbon-Farming Frameworks

Accurate and cost-effective SOC measurements are essential to MRV protocols, which typically require annual change detection at around ±0.3% SOC [64]. Although pH-corrected In-field-TOC-4 narrows its Bland–Altman limits to −0.27/+0.49%—approaching this benchmark—it still slightly exceeds the target range and fails the formal precision test (p = 0.0087), indicating that its variance remains higher than acceptable for standalone certification. Its per-sample cost (~EUR 3–4) in this study—comprising FarmLab rental fees and field labor (see Section 2.3)—is an order of magnitude lower than conventional laboratory analysis (EUR 44 per sample, including GPS-referenced composite sampling and Agrolab GmbH service charges; Section 2.2 and Section 2.3). This cost advantage enables much denser sampling, which is critical for capturing spatial heterogeneity and reducing overall uncertainty in carbon farming MRV systems [65].
Since soil carbon sequestration is inherently non-permanent and requires ongoing monitoring [66], and because no in-field method yet meets all three equivalence criteria (accuracy, precision, concordance), we advocate a hybrid MRV framework. Routine, high-frequency pH-corrected in-field measurements can track spatial and temporal dynamics, anchored by periodic laboratory benchmarks to ensure certification-grade accuracy, although the optimal recalibration frequency remains to be determined. Such an approach balances the need for spatial and temporal resolution with the rigor of laboratory benchmarks, providing both practical monitoring density and certification-grade accuracy.

4.5. Methodological Limitations

Despite the advances demonstrated here, several methodological constraints temper our conclusions. First, the proprietary nature of FarmLab’s chemometric algorithms prevents full transparency and independent recalibration [33]. Without access to raw spectral coefficients, our pH adjustment represents a pragmatic workaround rather than a fundamental model re-development. Second, although our subplot-level dataset (n = 100) provides robust statistical power for the tests applied, it remains moderate in size and spatial scope; longer-term, multi-season trials across diverse soil types will be needed to generalize findings [55].
Moreover, Roper et al. [67] showed that even established SOM assays can diverge substantially across sites and methods, underscoring the need to maintain a consistent SOC protocol over time or to apply site-specific calibration when switching assays [67].
Third, the FarmLab probe measures directly at ~10–15 cm depth and extrapolates to 30 cm via its internal models, rather than sampling the full profile physically. This may misrepresent SOC distribution in stratified soils, particularly where root-zone carbon differs markedly with depth [63]. Fourth, while moisture artifacts appear well-controlled, extreme moisture gradients or surface-crusting conditions may still challenge the EIS and NIR sensors under real field conditions [30].
Finally, our assessment focused on SOC concentration without direct SOC stock calculations; accurate stock estimation requires concurrent bulk density measurements in situ, which FarmLab currently does not provide [68]. Integrating a soil-compaction sensor or spatially explicit bulk density maps would therefore strengthen the device’s utility for carbon-stock monitoring. Acknowledging these limitations guides future research and underscores that, while FarmLab shows promise, it should be deployed within a hybrid framework that retains laboratory validation for high-stakes carbon accounting.

4.6. Comparison with Other In-Situ Sensor Platforms

Our results for FarmLab reflect a common theme in proximal-sensing research: while combining multiple sensors can substantially improve SOC estimates, achieving laboratory-level precision still hinges on site-specific calibration. For example, Dhawale et al. [69] showed that integrating visible/NIR reflectance with electrical conductivity yields R2 values up to 0.85 in Canadian fields—but only after extensive wet- and dry-season recalibration [69]. Likewise, the Yardstick probe, which extends its spade-mounted array to a 50 cm depth, achieves ±0.5% bias only once soil moisture and texture corrections are applied, yet still requires laboratory anchoring of its proprietary models [70].
More recently, Gyawali et al. [71] demonstrated that a handheld Vis-NIR sensor can reliably profile SOC down to 45cm with accuracy on par with conventional field campaigns, underscoring the promise of deep-profiling spade probes [71]. Reviews by Mokere et al. [17] and Gowera et al. [22] similarly report that most mobile and miniaturized spectrometers only attain ±0.3–0.7% SOC errors after incorporating covariates such as pH and clay content [17,22]. Sanderman et al. [53] further highlight that fusing benchtop MIR libraries with proximal spectra can detect multi-year SOC changes across diverse U.S. trials, boosting change-detection sensitivity in field sensors [53]. Even unsupervised learning on regional Vis-NIR libraries can reduce errors to ±0.3%, though ancillary soil data remain essential [23]. In this context, FarmLab’s fusion of visible/NIR, EIS, and environmental sensing—augmented by our pH-based adjustment—delivers performance at the lower end of this error spectrum (±0.27–0.49% LoA), demonstrating that thoughtful covariate integration can bring in-field estimates close to laboratory standards. Crucially, FarmLab’s low per-sample cost (~EUR 3–4) and rapid deployment strike a practical balance between affordability and precision when accompanied by robust calibration protocols.

4.7. Future Directions

Building on our findings, we identify several opportunities to enhance in-field SOC sensing and integrate FarmLab more fully into carbon-farming practice. First, longitudinal field trials are needed to assess temporal stability and repeatability at fixed monitoring points, as demonstrated for mineral-N sensing [30]. Such repeated measurements under varying environmental conditions will help quantify device drift and inform automated Quality Control routines.
Second, the integration of soil bulk density measurement into the probe design or the acquisition of concurrent bulk density maps (e.g., via proximal gamma-ray attenuation) would permit the direct calculation of SOC stocks rather than concentrations alone, addressing a key limitation for carbon accounting [68].
Third, expanding the calibration library with diverse soil types and management histories—potentially through federated spectral databases like LUCAS—would improve model transferability across regions [11,72]. In particular, the incorporation of texture and mineralogy covariates in a multivariate calibration framework could further reduce pH- and carbonate-related biases [12,18].
Fourth, data fusion combining FarmLab measurements with drone- and satellite-based spectral imagery offers the promise of scaling plot-level readings to field and landscape scales, as explored by others [14,73]. Hybrid models could leverage high-resolution in-field data for ground-truthing remote predictions, yielding robust SOC maps for MRV systems.
Finally, embedding machine learning pipelines that dynamically update calibration models based on ongoing field data (e.g., via active learning) could maintain accuracy in the face of seasonal and management-induced soil changes. Such adaptive approaches align with the vision of precision carbon farming as a continuously optimized system [74].
Implementing these recommendations will move FarmLab, and similar in situ SOC sensors, toward becoming reliable, scalable tools for high-density, cost-effective soil carbon monitoring, fulfilling both the scientific and practical requirements of carbon-farming initiatives.

5. Conclusions

Our first independent field validation of the FarmLab multi-sensor probe under temperate European arable conditions shows that its default model overestimates SOC by +0.20–0.27% (SD 0.25–0.28%), while a simple pH correction halves that bias (+0.11%, SD 0.23%) and moisture effects are effectively neutralized. However, formal equivalence testing confirms that even the pH-corrected algorithm cannot yet match laboratory precision and concordance, and predictive reliability is limited to SOC concentrations between ~1% and ~1.5% under low to moderate pH. Outside this range, the model tends to overpredict lower and underpredict higher SOC values.
Economically, FarmLab’s per-sample cost of ~EUR 3–4 (versus ~EUR 44 for GPS-referenced lab analysis) enables high-density mapping essential for carbon farming MRV. We therefore advocate a hybrid approach: the use of routine, pH-corrected in-field measurements to capture spatial and temporal trends, anchored by periodic laboratory benchmarks to ensure certification-grade accuracy.
Looking forward, improving FarmLab’s performance will depend on expanding calibration across diverse soils, integrating additional sensing technologies to capture mineralogical, texture, and bulk density data, and adopting adaptive, data-driven calibration algorithms—steps that together can elevate low-cost in-field sensing to near-laboratory standards and support scalable, cost-effective soil carbon monitoring.

Author Contributions

Conceptualization, validation, formal analysis, methodology, and investigation, L.K., A.G., C.V., C.A., E.-M.L.M. and A.Ö.; data curation, L.K., C.V., A.Ö. and C.A.; writing—original draft preparation, L.K.; writing—review and editing, L.K., E.-M.L.M., W.N., J.C.-B. and A.G.; visualization, L.K., C.V. and A.Ö.; supervision, A.G.; project administration, A.G.; funding acquisition, L.K. and A.G. All authors have read and agreed to the published version of the manuscript.

Funding

This research was partially funded by the European Innovation Partnership for Agricultural Productivity and Sustainability (EIP-AGRI) and the Rural Development Programme of Hesse 2014–2020 (EPLR), within the project “Humuvation—Innovative cultivation systems to promote yield stability and humus formation”. The APC was not funded.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

Data from this study are available from the corresponding author upon request.

Acknowledgments

We would like to express our gratitude to Stenon GmbH (Stenon GmbH, Zeppelinstraße 10, 14471 Potsdam, Germany) for providing the FarmLab device and granting access to the different measurement data, which were instrumental in conducting this study. Their support enabled us to critically assess the performance of the FarmLab device in comparison with established laboratory methods. Furthermore, we sincerely acknowledge the financial support from EIP-Agri Humuvation, which was funded by the European Union and the State of Hesse. This funding facilitated the soil sampling and laboratory analysis, without which this research would not have been possible.

Conflicts of Interest

The authors declare no conflicts of interest.

Abbreviations

The following abbreviations are used in this manuscript:
ANCOVAAnalysis of covariance
EISElectrical impedance spectroscopy
MIRMid-infrared spectroscopy
MRVMonitoring, Reporting, and Verification
NIRSNear-infrared spectroscopy
SOCSoil organic carbon
SOMSoil organic matter
TCTotal carbon
TOCTotal organic carbon
VOCVolatile organic compound
WFPSWater-filled pore space

References

  1. Bünemann, E.K.; Bongiorno, G.; Bai, Z.; Creamer, R.E.; De Deyn, G.; De Goede, R.; Fleskens, L.; Geissen, V.; Kuyper, T.W.; Mäder, P.; et al. Soil quality—A critical review. Soil Biol. Biochem. 2018, 120, 105–125. [Google Scholar] [CrossRef]
  2. Oldfield, S.F.; Olwell, P.; Shaw, N.; Havens, K. Conservation of Plant Species. In Seeds of Restoration Success; Springer: Cham, Switzerland, 2019; pp. 41–57. [Google Scholar] [CrossRef]
  3. Lal, R. Soil carbon management and climate change. Carbon Manag. 2013, 4, 439–462. [Google Scholar] [CrossRef]
  4. European Commission 2021. The Green Deal. 2025. Available online: https://eur-lex.europa.eu/legal-content/DE/TXT/?uri=CELEX:52019DC0640 (accessed on 19 April 2025).
  5. Fuss, S.; Canadell, J.G.; Ciais, P.; Jackson, R.B.; Jones, C.D.; Lyngfelt, A.; Peters, G.P.; van Vuuren, D.P. Moving toward Net-Zero Emissions Requires New Alliances for Carbon Dioxide Removal. One Earth 2020, 3, 145–149. [Google Scholar] [CrossRef]
  6. Schenuit, F.; Colvin, R.; Fridahl, M.; McMullin, B.; Reisinger, A.; Sanchez, D.L.; Smith, S.M.; Torvanger, A.; Wreford, A.; Geden, O. Carbon Dioxide Removal Policy in the Making: Assessing Developments in 9 OECD Cases. Front. Clim. 2021, 3, 638805. [Google Scholar] [CrossRef]
  7. Fridahl, M.; Schenuit, F.; Lundberg, L.; Möllersten, K.; Böttcher, M.; Rickels, W.; Hansson, A. Novel carbon dioxide removals techniques must be integrated into the European Union’s climate policies. Commun. Earth Environ. 2023, 4, 459. [Google Scholar] [CrossRef]
  8. McCarty, G.W.; Reeves, J.B.; Reeves, V.B.; Follett, R.F.; Kimble, J.M. Mid-Infrared and Near-Infrared Diffuse Reflectance Spectroscopy for Soil Carbon Measurement. Soil Sci. Soc. Am. J. 2002, 66, 640–646. [Google Scholar] [CrossRef]
  9. Reeves, J.B. Near-versus mid-infrared diffuse reflectance spectroscopy for soil analysis emphasizing carbon and laboratory versus on-site analysis: Where are we and what needs to be done? Geoderma 2010, 158, 3–14. [Google Scholar] [CrossRef]
  10. O’Rourke, S.M.; Holden, N.M. Optical sensing and chemometric analysis of soil organic carbon—A cost effective alternative to conventional laboratory methods? Soil Use Manag. 2011, 27, 143–155. [Google Scholar] [CrossRef]
  11. Stevens, A.; Nocita, M.; Tóth, G.; Montanarella, L.; van Wesemael, B. Prediction of Soil Organic Carbon at the European Scale by Visible and Near InfraRed Reflectance Spectroscopy. PLoS ONE 2013, 8, e66409. [Google Scholar] [CrossRef]
  12. Dorantes, M.J.; Fuentes, B.A.; Miller, D.M. Subsetting reduces the error of MIR spectroscopy models for soil organic carbon prediction in the U.S. Great Plains. Soil Secur. 2024, 16, 100145. [Google Scholar] [CrossRef]
  13. Angelopoulou, T.; Balafoutis, A.; Zalidis, G.; Bochtis, D. From Laboratory to Proximal Sensing Spectroscopy for Soil Organic Carbon Estimation—A Review. Sustainability 2020, 12, 443. [Google Scholar] [CrossRef]
  14. Croft, H.; Kuhn, N.J.; Anderson, K. On the use of remote sensing techniques for monitoring spatio-temporal soil organic carbon dynamics in agricultural systems. CATENA 2012, 94, 64–74. [Google Scholar] [CrossRef]
  15. Bricklemyer, R.S.; Brown, D.J. On-the-go VisNIR: Potential and limitations for mapping soil clay and organic carbon. Comput. Electron. Agric. 2010, 70, 209–216. [Google Scholar] [CrossRef]
  16. Debaene, G.; Bartmiński, P.; Siłuch, M. In Situ VIS-NIR Spectroscopy for a Basic and Rapid Soil Investigation. Sensors 2023, 23, 5495. [Google Scholar] [CrossRef]
  17. Mokere, R.; Ghassan, M.; Barra, I. Soil Spectroscopy Evolution: A Review of Homemade Sensors, Benchtop Systems, and Mobile Instruments Coupled with Machine Learning Algorithms in Soil Diagnosis for Precision Agriculture. Crit. Rev. Anal. Chem. 2024, 1–20. [Google Scholar] [CrossRef] [PubMed]
  18. Viscarra Rossel, R.A.; Lobsey, C.R.; Sharman, C.; Flick, P.; McLachlan, G. Novel Proximal Sensing for Monitoring Soil Organic C Stocks and Condition. Environ. Sci. Technol. 2017, 51, 5630–5641. [Google Scholar] [CrossRef] [PubMed]
  19. Biney, J.K.M.; Blöcher, J.R.; Bell, S.M.; Borůvka, L.; Vašát, R. Can in situ spectral measurements under disturbance-reduced environmental conditions help improve soil organic carbon estimation? Sci. Total Environ. 2022, 838, 156304. [Google Scholar] [CrossRef]
  20. Pei, X.; Sudduth, K.A.; Veum, K.S.; Li, M. Improving In-Situ Estimation of Soil Profile Properties Using a Multi-Sensor Probe. Sensors 2019, 19, 1011. [Google Scholar] [CrossRef]
  21. Grunwald, S.; Murad, M.O.F.; Farrington, S.; Wallace, W.; Rooney, D. Multi-Sensor Soil Probe and Machine Learning Modeling for Predicting Soil Properties. Sensors 2024, 24, 6855. [Google Scholar] [CrossRef]
  22. Gowera, G.T.; Sorenson, P.; Bedard-Haughn, A.; Rivard, B.; Peak, D.; Feng, J. Evaluating performance of miniaturized spectrometers in predicting soil properties using multivariate statistical analysis. Can. J. Soil. Sci. 2024, 104, 216–226. [Google Scholar] [CrossRef]
  23. Miloš, B.; Bensa, A. Organic carbon estimation in a regional soil Vis-NIR database supported by unsupervised learning and chemometrics techniques. Soil Adv. 2024, 2, 100013. [Google Scholar] [CrossRef]
  24. Yang, P.; Wang, Y.; Hu, B.; Li, S.; Chen, S.; Luo, D.; Peng, J. Predicting soil organic carbon content using simulated insitu spectra and moisture correction algorithms in southern Xinjiang, China. Geoderma Reg. 2024, 37, e00783. [Google Scholar] [CrossRef]
  25. Shi, L.; Daly, K.; O’Rourke, S. Estimating mineral-associated organic carbon saturation and sequestration potential using MIR spectral based local quantile regression. Geoderma 2025, 454, 117181. [Google Scholar] [CrossRef]
  26. Zhang, Z.; Ding, J.; Zhu, C.; Wang, J.; Ma, G.; Ge, X.; Li, Z.; Han, L. Strategies for the efficient estimation of soil organic matter in salt-affected soils through Vis-NIR spectroscopy: Optimal band combination algorithm and spectral degradation. Geoderma 2021, 382, 114729. [Google Scholar] [CrossRef]
  27. Zhou, T.; Jia, C.; Zhang, K.; Yang, L.; Zhang, D.; Cui, T.; He, X. A rapid detection method for soil organic matter using a carbon dioxide sensor in situ. Measurement 2023, 208, 112471. [Google Scholar] [CrossRef]
  28. Veris Technologies. Soil Scanning Sensors for Agriculture. 2025. Available online: https://www.veristech.com/sensing (accessed on 17 April 2025).
  29. Laskar, S.; Mukherjee, S. Optical Sensing Methods for Assessment of Soil Macronutrients and other Properties for Application in Precision Agriculture:A review. ADBU J. Eng. Technol. 2016, 4, 206–210. [Google Scholar]
  30. Vikuk, V.; Spirkaneder, A.; Noack, P.; Duemig, A. Validation of a sensor-system for real-time measurement of mineralized nitrogen in soils. Smart Agric. Technol. 2024, 7, 100390. [Google Scholar] [CrossRef]
  31. Loria, N.; Lal, R.; Chandra, R. Handheld In Situ Methods for Soil Organic Carbon Assessment. Sustainability 2024, 16, 5592. [Google Scholar] [CrossRef]
  32. Xie, A.; Zhou, Q.; Fu, L.; Zhan, L.; Wu, W. From Lab to Field: Advancements and Applications of On-The-Go Soil Sensors for Real-Time Monitoring. Eurasian Soil Sci. 2024, 57, 1730–1745. [Google Scholar] [CrossRef]
  33. Steiger, A.; Qaswar, M.; Bill, R.; Mouazen, A.M.; Grenzdörffer, G. Comparing the handheld Stenon FarmLab soil sensor with a Vis-NIR multi-sensor soil sensing platform. Smart Agric. Technol. 2025, 10, 100717. [Google Scholar] [CrossRef]
  34. Stenon GmbH. Product Information and Calibration Use Cases. Available online: https://www.stenon.io (accessed on 8 June 2025).
  35. Stenon. Sachgerechte Validierung von FarmLab. 2024. Available online: https://blog.stenon.io/wp-content/uploads/2021/11/2021-09-16-Sachgerechte-Validierung-von-FarmLab_final.pdf (accessed on 3 February 2025).
  36. Rawls, W.J. Estimating Soil Bulk Density From Particle Size and Organic Mater Content. Soil Sci. 1983, 135, 123–125. [Google Scholar] [CrossRef]
  37. Dai, L.; Xue, J.; Lu, R.; Wang, Z.; Chen, Z.; Yu, Q.; Shi, Z.; Chen, S. In-situ prediction of soil organic carbon contents in wheat-rice rotation fields via visible near-infrared spectroscopy. Soil Environ. Health 2024, 2, 100113. [Google Scholar] [CrossRef]
  38. DIN 13878:2014-09; Soil Quality—Determination of Total Carbon After Dry Combustion (Elementary Analysis). Beuth Verlag: Berlin, Germany, 2014.
  39. Natali, C.; Bianchini, G.; Carlino, P. Thermal stability of soil carbon pools: Inferences on soil nature and evolution. Thermochim. Acta 2020, 683, 178478. [Google Scholar] [CrossRef]
  40. DIN EN ISO/IEC 17025:2018-03; Allgemeine Anforderungen an Die Kompetenz von Prüf- und Kalibrierlaboratorien. Deutsches Institut für Normung: Berlin, Germany, 2018.
  41. ASTM D422-63(2002); Standard Test Method for Particle-Size Analysis of Soils. ASTM International: West Conshohocken, PA, USA, 2002. [CrossRef]
  42. VDLUFA. Methodenbuch Band 1: Die Untersuchung von Böden; VDLUFA: Darmstadt, Germany, 1991. [Google Scholar]
  43. Silveira, P.S.P.; Siqueria, J.d.O. 2022 eirasBA: Extended Bland-Altman Statistical Tests. R Package Version 0.1.0. 2022. Available online: https://CRAN.R-project.org/package=eirasBA (accessed on 11 November 2024).
  44. R Core Team. R: A Language and Environment for Statistical Computing; R Core Team: Vienna, Austria, 2024. [Google Scholar]
  45. Datta, D. Blandr: A Bland-Altman Method Comparison Package for R. 2025. Available online: https://github.com/deepankardatta/blandr (accessed on 17 April 2025).
  46. Therneau, T. Deming: Deming, Theil-Sen, Passing-Bablock and Total Least Squares Regression. 2018. Available online: https://CRAN.R-project.org/package=deming (accessed on 12 December 2024).
  47. Nash, J.E.; Sutcliffe, J.V. River flow forecasting through conceptual models part I—A discussion of principles. J. Hydrol. 1970, 10, 282–290. [Google Scholar] [CrossRef]
  48. Hedberg, E.C.; Ayers, S. The power of a paired t-test with a covariate. Soc. Sci. Res. 2015, 50, 277–291. [Google Scholar] [CrossRef] [PubMed]
  49. Creasy, M.A. Confidence Limits for the Gradient in the Linear Functional Relationship. J. R. Stat. Soc. Ser. B Stat. Methodol. 1956, 18, 65–69. [Google Scholar] [CrossRef]
  50. Shukla, G.K. Some Exact Tests of Hypotheses about Grubbs’s Estimators. Biometrics 1973, 29, 373. [Google Scholar] [CrossRef]
  51. Silveira, P.S.P.; Vieira, J.E.; Siqueira, J.O. Is Bland-Altman plot method useful without inference for accuracy, precision, and agreement? arXiv 2021, arXiv:2108.12937. [Google Scholar] [CrossRef]
  52. Metzger, K.; Zhang, C.; Daly, K. From benchtop to handheld MIR for soil analysis: Predicting lime requirement and organic matter in agricultural soils. Biosyst. Eng. 2021, 204, 257–269. [Google Scholar] [CrossRef]
  53. Sanderman, J.; Savage, K.; Dangal, S.R.S.; Duran, G.; Rivard, C.; Cavigelli, M.A.; Gollany, H.T.; Jin, V.L.; Liebig, M.A.; Omondi, E.C.; et al. Can Agricultural Management Induced Changes in Soil Organic Carbon Be Detected Using Mid-Infrared Spectroscopy? Remote Sens. 2021, 13, 2265. [Google Scholar] [CrossRef]
  54. Jo, J.; Son, Y.; Park, M.-K.; Lee, J.Y.; Chu, H.; Ahn, Y.G. Statistical comparison for assessing agreement between two mass spectrometric methods for the analysis of polychlorinated dibenzo-p-dioxins and furans (PCDDs/Fs) in contaminated soils. Chemosphere 2024, 363, 142806. [Google Scholar] [CrossRef] [PubMed]
  55. Grahmann, K.; Zwink, M.; Barkusky, D.; Verch, G.; Sommer, M. The dilemma of analytical method changes for soil organic carbon in long-term experiments. Eur. J. Soil Sci. 2023, 74, e13362. [Google Scholar] [CrossRef]
  56. Bland, J.M.; Altman, D.G. Statistical methods for assessing agreement between two methods of clinical measurement. Lancet 1986, 327, 307–310. [Google Scholar] [CrossRef]
  57. Steinegger, A.; Wolfbeis, O.S.; Borisov, S.M. Optical Sensing and Imaging of pH Values: Spectroscopies, Materials, and Applications. Chem. Rev. 2020, 120, 12357–12489. [Google Scholar] [CrossRef] [PubMed]
  58. Bönecke, E.; Meyer, S.; Vogel, S.; Schröter, I.; Gebbers, R.; Kling, C.; Kramer, E.; Lück, K.; Nagel, A.; Philipp, G.; et al. Guidelines for precise lime management based on high-resolution soil pH, texture and SOM maps generated from proximal soil sensing data. Precis. Agric 2021, 22, 493–523. [Google Scholar] [CrossRef]
  59. Vogel, S.; Gebbers, M.; Schröter, I.; Schwanghart, W.; Bönecke, E.; Rühlmann, J.; Kramer, E.; Gebbers, R. Towards site-independent calibration of in situ soil pH sensor data: Relevance of spatial and temporal proximity, sample size and data spread for calibration model performance. Geoderma 2025, 456, 117261. [Google Scholar] [CrossRef]
  60. Kleber, M.; Johnson, M.G. Advances in understanding the molecular structure of soil organic matter: Implications for interactions in the environment. Adv. Agron. 2010, 106, 77–142. [Google Scholar] [CrossRef]
  61. von Lützow, M.; Kögel-Knabner, I.; Ekschmitt, K.; Matzner, E.; Guggenberger, G.; Marschner, B.; Flessa, H. Stabilization of Organic Matter in Temperate Soils: Mechanisms and Their Relevance under Different Soil Conditions—A Review. Eur. J. Soil Sci. 2006, 57, 426–445. [Google Scholar] [CrossRef]
  62. O’Rourke, S.M.; Stockmann, U.; Holden, N.M.; McBratney, A.B.; Minasny, B. An Assessment of Model Averaging to Improve Predictive Power of Portable vis-NIR and XRF for the Determination of Agronomic Soil Properties. Geoderma 2016, 279, 31–44. [Google Scholar] [CrossRef]
  63. Ciani, A.; Goss, K.-U.; Schwarzenbach, R.P. Light penetration in soil and particulate minerals. Eur. J. Soil Sci. 2005, 56, 561–574. [Google Scholar] [CrossRef]
  64. Smith, P.; Soussana, J.-F.; Angers, D.; Schipper, L.; Chenu, C.; Rasse, D.P.; Batjes, N.H.; van Egmond, F.; McNeill, S.; Kuhnert, M.; et al. How to measure, report and verify soil carbon change to realize the potential of soil carbon sequestration for atmospheric greenhouse gas removal. Glob. Change Biol. 2020, 26, 219–241. [Google Scholar] [CrossRef] [PubMed]
  65. Barbato, C.T.; Strong, A.L. Farmer perspectives on carbon markets incentivizing agricultural soil carbon sequestration. Npj Clim. Action 2023, 2, 26. [Google Scholar] [CrossRef]
  66. Leifeld, J. Carbon farming: Climate change mitigation via non-permanent carbon sinks. J. Environ. Manag. 2023, 339, 117893. [Google Scholar] [CrossRef] [PubMed]
  67. Roper, W.R.; Robarge, W.P.; Osmond, D.L.; Heitman, J.L. Comparing Four Methods of Measuring Soil Organic Matter in North Carolina Soils. Soil Sci. Soc. Am. J. 2019, 83, 466–474. [Google Scholar] [CrossRef]
  68. Wiesmeier, M.; Mayer, S.; Paul, C.; Helming, K.; Don, A.; Franko, U.; Steffens, M.; Kögel-Knabner, I. CO2 Certificates for Carbon Sequestration in Soils: Methods, Management Practices and Limitations: BonaRes Data Centre (Leibniz Centre for Agricultural Landscape Research (ZALF). 2020. Available online: https://doi.org/10.20387/BonaRes-NE0G-CE98 (accessed on 5 May 2025).
  69. Dhawale, N.M.; Adamchuk, V.I.; Prasher, S.O.; Viscarra Rossel, R.A. Evaluating the Precision and Accuracy of Proximal Soil vis–NIR Sensors for Estimating Soil Organic Matter and Texture. Soil Syst. 2021, 5, 48. [Google Scholar] [CrossRef]
  70. Klein, J. Yard Stick Provides Measurement Technology to Combat Climate Change; TechCrunch: Boston, MA, USA, 2021. [Google Scholar]
  71. Gyawali, A.J.; Wiseman, M.; Ackerson, J.P.; Coffman, S.; Meissner, K.; Morgan, C.L. Measuring in situ soil carbon stocks: A study using a novel handheld VisNIR probe. Geoderma 2025, 453, 117152. [Google Scholar] [CrossRef]
  72. van der Voort, T.S.; Verweij, S.; Fujita, Y.; Ros, G.H. Enabling soil carbon farming: Presentation of a robust, affordable, and scalable method for soil carbon stock assessment. Agron. Sustain. Dev. 2023, 43, 22. [Google Scholar] [CrossRef]
  73. Seidel, M.; Hutengs, C.; Ludwig, B.; Thiele-Bruhn, S.; Vohland, M. Strategies for the efficient estimation of soil organic carbon at the field scale with vis-NIR spectroscopy: Spectral libraries and spiking vs. local calibrations. Geoderma 2019, 354, 113856. [Google Scholar] [CrossRef]
  74. Minasny, B.; McBratney, A.B.; Arrouays, D.; Chabbi, A.; Field, D.J.; Kopittke, P.M.; Morgan, C.L.S.; Padarian, J.; Rumpel, C. Soil Carbon Sequestration: Much More Than a Climate Solution. Environ. Sci. Technol. 2023, 57, 19094–19098. [Google Scholar] [CrossRef]
Soilsystems 09 00067 g001
Figure 1. Sampling sites with soil type, texture, and coordinates in Lower Saxony (1–3) and Hesse, Germany (4–9).
Figure 1. Sampling sites with soil type, texture, and coordinates in Lower Saxony (1–3) and Hesse, Germany (4–9).
Soilsystems 09 00067 g001
Soilsystems 09 00067 g002
Figure 2. Scheme of soil sampling. Left: 2 m2 plot with subplots arranged in square at sites in Hesse; right: plots (black frame) in line with subplots arranged in line at sites in Lower Saxony. Numbers from 1 to 5 indicate subplots.
Figure 2. Scheme of soil sampling. Left: 2 m2 plot with subplots arranged in square at sites in Hesse; right: plots (black frame) in line with subplots arranged in line at sites in Lower Saxony. Numbers from 1 to 5 indicate subplots.
Soilsystems 09 00067 g002
Soilsystems 09 00067 g003
Figure 3. Distribution of key soil parameters measured across different sampling sites. Histograms displaying frequency distribution: Standard-TOC, pH, moisture, and total inorganic carbon (TIC900) (n = 100); clay content (n = 22); and bulk density (n = 46).
Figure 3. Distribution of key soil parameters measured across different sampling sites. Histograms displaying frequency distribution: Standard-TOC, pH, moisture, and total inorganic carbon (TIC900) (n = 100); clay content (n = 22); and bulk density (n = 46).
Soilsystems 09 00067 g003
Soilsystems 09 00067 g004
Figure 4. Scatterplots of predicted versus measured soil organic carbon (SOC) values for In-field-TOC-1 (A) and pH-corrected In-field-TOC-4 (B) compared to Standard-TOC. Each point represents one subplot (n = 100), colored by soil pH. Dashed lines indicate 1:1 identity line; solid lines show linear regression fit. Color gradients illustrate relationship between prediction deviation and soil pH.
Figure 4. Scatterplots of predicted versus measured soil organic carbon (SOC) values for In-field-TOC-1 (A) and pH-corrected In-field-TOC-4 (B) compared to Standard-TOC. Each point represents one subplot (n = 100), colored by soil pH. Dashed lines indicate 1:1 identity line; solid lines show linear regression fit. Color gradients illustrate relationship between prediction deviation and soil pH.
Soilsystems 09 00067 g004
Soilsystems 09 00067 g005
Figure 5. Bland–Altman plots comparing In-field-TOC measurements ((A): In-field-TOC-1; (B): In-field-TOC-4) to Standard-TOC measurements. The x-axes show the mean SOC content (%) between the two methods, while the y-axes show their differences. The central dashed lines represent the mean bias between the methods, and the outer dashed lines denote the limits of agreement (±1.96 × standard deviation of the differences). The vertical bars on the left indicate the 95% confidence intervals for the mean bias (n = 100).
Figure 5. Bland–Altman plots comparing In-field-TOC measurements ((A): In-field-TOC-1; (B): In-field-TOC-4) to Standard-TOC measurements. The x-axes show the mean SOC content (%) between the two methods, while the y-axes show their differences. The central dashed lines represent the mean bias between the methods, and the outer dashed lines denote the limits of agreement (±1.96 × standard deviation of the differences). The vertical bars on the left indicate the 95% confidence intervals for the mean bias (n = 100).
Soilsystems 09 00067 g005
Table 1. Overview of soil organic carbon (SOC) measurement methods, including category, method description, and institution.
Table 1. Overview of soil organic carbon (SOC) measurement methods, including category, method description, and institution.
AcronymCategoryMethod DescriptionInstitution
TC-GiLaboratory (TC)Dry combustion at 1140 °C (DIN 13878) with thermal conductivity detection (N2) and He carrier gas to measure total carbon (organic + inorganic).Justus Liebig University, Giessen, Germany
TC-GoeLaboratory (TC)An identical dry-combustion protocol to TC-Gi, performed independently to assess inter-laboratory reliability.University of Göttingen, Göttingen, Germany
SoliTOCLaboratory (TOC)Temperature-differentiated oxidation (DIN 19539): thermally labile OC < 400 °C + residual OC 500–600 °C; late-stage carbonate breakdown > 650 °C.Justus Liebig University, Giessen, Germany
TOC-acidLaboratory (TOC)Acid fumigation to remove inorganic carbon, then combustion (900–1500 °C; DIN EN ISO/IEC 17025) to quantify organic C via CO2 detection.Agrolab GmbH, Leinefelde, Germany
Standard-TOCReferenceThe arithmetic mean of the two laboratory TOC methods (SoliTOC + TOC-acid).—
In-field-TOC-1In-field (FarmLab)The baseline SOC estimate from the FarmLab multi-sensor probe, combining visible/NIR spectroscopy and electrical impedance spectroscopy (EIS).Stenon GmbH, Potsdam, Germany
In-field-TOC-2In-field (FarmLab)The SOC estimate from FarmLab using the first updated calibration algorithm provided by Stenon.Stenon GmbH, Potsdam, Germany
In-field-TOC-3In-field (FarmLab)The SOC estimate from FarmLab using the second updated calibration algorithm provided by Stenon.Stenon GmbH, Potsdam, Germany
In-field-TOC-4In-field (FarmLab)The baseline In-field-TOC-1 output adjusted by empirically derived pH-based correction factors (authors’ modification).Stenon GmbH, Potsdam, Germany/Authors
Table 2. Mean soil organic carbon (SOC), standard deviation, and bias relative to Standard-TOC (n = 100).
Table 2. Mean soil organic carbon (SOC), standard deviation, and bias relative to Standard-TOC (n = 100).
MethodMean SOC ± SD (%)Bias vs. Standard-TOC (%)
SoliTOC1.26 ± 0.22−0.03
TOC-acid1.32 ±0.20+0.03
TC-Goe1.33 ± 0.21+0.04
TC-Gi1.35 ± 0.19+0.06
Standard-TOC1.29 ± 0.210.00
In-field-TOC-11.49 ± 0.28+0.20
In-field-TOC-21.56 ± 0.27+0.27
In-field-TOC-31.54 ± 0.25+0.25
In-field-TOC-41.40 ± 0.23+0.11
Table 3. Correlation of FarmLab SOC error with soil pH and soil moisture.
Table 3. Correlation of FarmLab SOC error with soil pH and soil moisture.
RelationshipPearson’s rp-Value
In-field-TOC-1 error vs. soil pH−0.39<0.01 **
In-field-TOC-1 error vs. TIC900−0.100.31 (n.s.)
In-field-TOC-1 error vs. soil moisture−0.140.16 (n.s.)
Note: ** = p < 0.01; n.s. = not significant.
Table 4. Deming regression intercepts, slopes, and coefficients of determination (R2) for selected method pairs (n = 100).
Table 4. Deming regression intercepts, slopes, and coefficients of determination (R2) for selected method pairs (n = 100).
Method PairIntercept (±SE)%Slope (±SE)R2
In-field-TOC-1 vs. Standard-TOC0.18 ± 0.071.10 ± 0.050.83
In-field-TOC-4 vs. Standard-TOC0.05 ± 0.060.97 ± 0.040.79
TC-Gi vs. Standard-TOC0.06 ± 0.021.01 ± 0.020.92
TC-Goe vs. Standard-TOC0.04 ± 0.021.00 ± 0.020.93
Table 5. Mean absolute error (MAE), root mean square error (RMSE), and Nash–Sutcliffe efficiency (NSE) for selected method pairs compared to Standard TOC reference (n = 100).
Table 5. Mean absolute error (MAE), root mean square error (RMSE), and Nash–Sutcliffe efficiency (NSE) for selected method pairs compared to Standard TOC reference (n = 100).
Method PairMAE%RMSE%NSE
In-field-TOC-1 vs. Standard-TOC0.2820.349−3.21
In-field-TOC-4 vs. Standard-TOC0.2440.317−2.478
TC-Gi vs. Standard-TOC0.1620.377−3.915
TC-Goe vs. Standard-TOC0.0960.170.001
Table 6. Mean biases and 95% limits of agreement (LoAs) from Bland–Altman analyses (n = 100).
Table 6. Mean biases and 95% limits of agreement (LoAs) from Bland–Altman analyses (n = 100).
Method PairMean Bias (%)95% LoA (%)
In-field-TOC-1 vs. Standard-TOC+0.20−0.35 to +0.75
In-field-TOC-4 vs. Standard-TOC+0.11−0.27 to +0.49
TC-Gi vs. TC-Goe+0.05−0.12 to +0.22
Table 7. Equivalence tests for SOC method pairs (n = 100), showing both pass/fail and underlying p-values for accuracy (mean equality), precision (variance equality), and concordance (bisector agreement: intercept/slope). Methods either passed (“✔”) or failed (“✘”).
Table 7. Equivalence tests for SOC method pairs (n = 100), showing both pass/fail and underlying p-values for accuracy (mean equality), precision (variance equality), and concordance (bisector agreement: intercept/slope). Methods either passed (“✔”) or failed (“✘”).
Method PairAccuracy p-ValuePrecision p-ValueConcordance p-Value (Int/Slope)AccuracyPrecisionConcordance
TC-Gi vs. TC-Goe0.00300.08420.1055/0.0613✘✔✔
SoliTOC vs. TOC-acid0.00280.00850.1147/0.1212✘✔✔
Std-TOC vs. In-field-TOC-1<0.00010.08420.2401/0.1116✘✔✔
Std-TOC vs. In-field-TOC-2<0.0001<0.00010.0116/0.2401✘✘✔
Std-TOC vs. In-field-TOC-3<0.0001<0.00010.1690/0.2299✘✘✔
Std-TOC vs. In-field-TOC-40.32500.00870.1157/0.1212✔✘✔
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Kohl, L.; Vielhauer, C.; Öztürk, A.; Minarsch, E.-M.L.; Ahl, C.; Niether, W.; Clifton-Brown, J.; Gattinger, A. Field Evaluation of a Portable Multi-Sensor Soil Carbon Analyzer: Performance, Precision, and Limitations Under Real-World Conditions. Soil Syst. 2025, 9, 67. https://doi.org/10.3390/soilsystems9030067

AMA Style

Kohl L, Vielhauer C, Öztürk A, Minarsch E-ML, Ahl C, Niether W, Clifton-Brown J, Gattinger A. Field Evaluation of a Portable Multi-Sensor Soil Carbon Analyzer: Performance, Precision, and Limitations Under Real-World Conditions. Soil Systems. 2025; 9(3):67. https://doi.org/10.3390/soilsystems9030067

Chicago/Turabian Style

Kohl, Lucas, Clarissa Vielhauer, Atilla Öztürk, Eva-Maria L. Minarsch, Christian Ahl, Wiebke Niether, John Clifton-Brown, and Andreas Gattinger. 2025. "Field Evaluation of a Portable Multi-Sensor Soil Carbon Analyzer: Performance, Precision, and Limitations Under Real-World Conditions" Soil Systems 9, no. 3: 67. https://doi.org/10.3390/soilsystems9030067

APA Style

Kohl, L., Vielhauer, C., Öztürk, A., Minarsch, E.-M. L., Ahl, C., Niether, W., Clifton-Brown, J., & Gattinger, A. (2025). Field Evaluation of a Portable Multi-Sensor Soil Carbon Analyzer: Performance, Precision, and Limitations Under Real-World Conditions. Soil Systems, 9(3), 67. https://doi.org/10.3390/soilsystems9030067

Article Metrics

Back to TopTop