Monitoring the Composting Process of Olive Oil Industry Waste: Benchtop FT-NIR vs. Miniaturized NIR Spectrometer

Abstract

Miniaturized near-infrared (NIR) spectrometers are revolutionizing the agri-food industry thanks to their compact size and ultra-fast analysis capabilities. This work compares the analytical performance of a handheld NIR spectrometer and a benchtop FT-NIR for the determination of several parameters, namely, pH, electrical conductivity (EC₂₅), C/N ratio, and organic matter as LOI (loss-on-ignition) in compost. Samples were collected at different stages of maturity from a full-scale facility that processes olive mill semi-solid residue together with olive tree pruning residue and animal manure. Using an FT-NIR spectrometer, satisfactory predictions (RPD > 2.0) were obtained with both partial least squares (PLS) and support vector machine (SVM) regression, SVM clearly being superior in the case of pH (RMSEP = 0.26; RPD = 3.8). The superior performance of the FT-NIR spectrometer in comparison with the handheld spectrometer was essentially due to the extended spectral range, especially for pH. In general, when analyzing intact samples with the miniaturized spectrometer, sample rotation decreased RMSEP values (~20%). Nevertheless, a fast and simple assessment of compost quality with reasonable prediction performance can also be achieved on intact samples by averaging static measurements acquired at different sample positions.

1. Introduction

The olive oil industry plays a key socio-economic role in the Mediterranean countries, accounting for over 95% of global olive oil production. However, this sector faces severe environmental challenges due to the production of large quantities of wet olive mill pomace (OMP), the main by-product of the two-phase centrifugation method. OMP is rich in easily degradable organic matter, but it contains high levels of tannins and polyphenols which are responsible for antimicrobial activities and phytotoxicity [1], so it cannot be safely soil-applied, also being very difficult to dispose of. Composting, as a means of waste-to-fertilizer conversion, aids in the transformation from the traditional linear economic model that generates tremendous amounts of waste to a circular regenerative agriculture model that reduces waste by reusing by-products to generate additional value, thus restoring and enhancing soil health. This approach mimics the natural cycle of nutrient use and recycling, thereby promoting sustainability. OMP can be composted by mixing with other waste like animal manure, cereal straw, olive leaves, or pruning residue, being an efficient recycling solution, ensuring the biological degradation of potentially harmful compounds [2]. Therefore, composting is an effective alternative to managing large amounts of olive oil production waste in these regions, providing an organic amendment suitable for application to croplands, and is likely to improve poor and degraded soils [3,4]. To ensure successful field application, it is essential to analyze the chemical composition of OMP and establish criteria for compost maturity assessment, since the use of non-mature compost may cause phytotoxicity, oxygen reduction, and an increase in soil carbon-to-nitrogen (C/N) ratio. However, these methods often require chemical or biological analyses in a laboratory, which are also typically time- and money-consuming.

Vibrational spectroscopy in the near-infrared region (NIR) has been reported to have great potential for non-destructive quality control of several biological materials, mainly organics [5,6,7,8]. It provides rapid and reliable complex chemical and physical information related to the vibration of molecular bonds (such as C-H, C-C, C=C, N-H, or O-H), after little or even no sample pretreatment. For this reason, this technology is far more sustainable than the wet chemical analyses commonly used for compost assessment. This aligns with the 17 Sustainable Development Goals (SDGs); in particular, Goal 17 promotes “the development, transfer and diffusion of environmentally sound technologies” [9]. Interest in near-infrared (NIR) spectroscopy is currently widespread, and its combination with chemometric tools has significantly expanded its applications in many research fields. Since NIR spectra consist of weak and broad absorption bands, multivariate statistical approaches like regression are always required to extract useful information about the parameters of interest from the NIR spectra, and a straightforward interpretation is often difficult. Furthermore, one important aspect to consider when using any type of NIR device is the possible sources of variability that can influence the spectra obtained (environmental effects, instrumental noise, wavelength-dependent scattering effects, etc.).

In this sense, in most industrial spectroscopic applications, classical regression algorithms like partial least squares (PLS) are used to model spectral data because of their simplicity, speed, relatively good performance, and easy accessibility. However, a major limitation concerns the assumption of linear relationships between the spectral patterns and the properties to be predicted. Such an assumption is not always valid for industrial samples whose chemical composition or particle size may vary, or in systems with strong inter- or intramolecular interactions. Note that, by using PLS, non-linear relations can only be modeled in a limited way by considering a larger number of latent variables, leading to model overfitting and possible prediction errors [10]. As an alternative, non-linear machine learning methods have progressively gained interest in chemometrics in recent years, although they are still little known [11,12]. Among them, support vector machine regression (SVM) can be considered a good option for spectral regression purposes, also being able to efficiently deal with high-dimensional input vectors [12,13,14]. The principle of SVM consists of transforming the non-linear problem into a linear one by mapping the original data into a high-dimensional feature space. In this case, the linear problem is much simpler. The mapping is constructed using kernel functions, which can be regarded as the expression of similarities between samples [13].

Studies involving compost and organic waste have shown that NIR spectroscopy is highly effective in monitoring these processes and evaluating their quality [15,16,17,18]. However, very little is known about its potential for monitoring and tracing the composting process of olive mill pomace in full-scale facilities [19]. The size and low robustness of commercially available instruments limited the use of NIR spectroscopy under field conditions for many years [11], but ultra-miniaturized NIR spectrometers are nowadays a reality. They are either USB-powered or have a built-in battery, weigh less than 100 g, and can be operated by an application installed on a laptop or even a smartphone. Such devices are gaining increasing interest as an attractive alternative for agro-industries and small producers in quality control of various natural products [11,12,20,21,22,23]. Nevertheless, as recognized by the scientific and professional community, miniaturized equipment still requires an in-depth evaluation of performance characteristics and comprehensive validation studies performed in well-equipped laboratories to make these instruments suitable for everyday life applications by a non-expert user community [24]. So far, several studies have focused on comparing the performance of both benchtop and portable NIR spectrometers using chemometric regression tools, especially in the fruit and vegetable sectors [21,22,23,25], but feasibility studies in different scenarios with comparative analyses carried out on scientific-grade benchtop spectrometers are still essential.

It is worth noting that a careful evaluation for each particular application is required before these miniaturized instruments that are currently marketed can be effectively implemented, and none exists for the quality assessment of compost. Furthermore, one limitation of handheld spectrometers is their small probing area in comparison with conventional ones [20]. They only probe a few millimeters squared of solid, which may affect the representativeness of the spectral data acquired [26], mainly when heterogeneous samples, such as compost, are analyzed. Sample pretreatment through intensive grinding is a way to overcome this limitation. Thus Fernández-Pierna et al. [27] observed that the performance of portable NIR spectrometers in predicting grass silage quality was lower when dealing with non-ground samples. However, this preprocessing is time-consuming and not suited for field measurements.

This study aims to critically compare the performance of a miniaturized near-infrared (NIR) spectrometer to a benchtop Fourier transform (FT)-NIR spectrometer. Key physical–chemical parameters of compost quality are estimated during the maturation process within a full-scale composting facility by means of various regression algorithms. The benchtop FT-NIR spectrometer, renowned for its high-end capabilities, including exceptional wavelength accuracy, spectral quality, and reproducibility, served as a benchmark for performance accuracy. Furthermore, the possibility of working with intact (non-ground) compost samples, as a first approximation to direct applicability in composting plants, is explored. To carry this out, two different strategies are considered to increase the representativeness of the spectral data acquired with the miniaturized spectrometer: (a) the use of a windowed sapphire collar, which allows a direct analysis in contact with the sample and easy displacement at various measurement points, and (b) the use of a spinner accessory, which rotates the sample during spectra acquisition. The comparison of these commercial measurement accessories, carried out here for the first time, can aid in better comprehending the performance of the micro-NIR spectrometer and its potential usability in large-scale composting facilities.

2. Materials and Methods

2.1. Composting Plant and Sample Description

Compost samples were collected from INGNIA S.L., a composting facility located in Almedinilla (Córdoba, Andalucía, Spain). A mixture of wet olive mill pomace (68%) from local olive mills, goat manure mixed with straw (20%), olive tree pruning residue (9%), and chicken manure mixed with straw cattle bedding (3%) is used to produce their compost. Plant residue and straw cattle bedding are added as bulking agents to promote efficient decomposition.

Sampling. A total of 126 compost samples were collected throughout the entire composting process over two olive harvest seasons (2020/2021 and 2021/2022). In both cases, windrow composting was conducted in an open pile (80 m (L) × 6 m (W) × 4 m (H)), formed at the start of each olive harvest and oil production season. Each sample consisted of 5 sub-samples collected around specific points of the pile, considering different heights and depths to ensure representativeness.

Sample pretreatment. All the samples were desiccated at 65 °C, finely ground in a domestic dry grinder and sieved to 200 μm. Intact samples (non-finely ground) were also considered.

2.2. Physical–Chemical Analyses

Electrical conductivity at 25 °C (EC₂₅) and pH were measured in 1:10 (w/v) water extracts with an EC-meter CRISON Basic 30 (Hach Lange GmbH, Düsseldorf, Germany) and a HANNA pH meter (HANNA^® Instruments, Melbourne, Australia), respectively. The determination of carbon and nitrogen was conducted in a LECO TruSpec CHN 620-100-400 (LECO Corporation, St. Joseph, MI, USA) elemental analyzer. To determine the organic matter of compost, the “loss-on-ignition” (LOI) method was employed, involving combustion in a muffle furnace at 550 °C for 5 h [28].

2.3. Near-Infrared Spectra Acquisition

An Antaris Fourier transform near-infrared (FT-NIR) system (Thermo Fisher Scientific, Waltham, MA, USA), equipped with an integrating sphere module for diffuse reflectance measurements, was used for benchtop NIR spectra registration. For the measurements, approximately 5 g of each sample was placed directly into a 3 cm diameter capsule, which was placed on top of the integrating sphere window and rotated during the process. A total of 64 accumulations were conducted within a spectral range of 10000–4000 cm⁻¹ (1000–2500 nm) and at a resolution of 8 cm⁻¹ with a measurement time of 55 s. Each sample was measured 3 times by refilling the capsule to ensure representativeness and a single spectrum, was averaged from these three replicates, and was used for data analysis. A reference measurement was taken before each spectrum using the sphere’s internal gold reference.

For handheld NIR spectral data acquisition, an ultra-compact Micro-NIR 1700ES spectrometer was employed (VIAVI Solutions, Chandler, AZ, USA), using VIAVI Micro-NIR Pro software version 3.2. The micro-NIR spectrometer’s dimensions were 45 × 50 mm, weighing 64 g. It was equipped with two tungsten light bulbs as the radiation source and a linear variable filter (LVF) directly connected to a linear indium–gallium–arsenide (InGaAs) 128-pixel array detector. The spectra were registered in the range of 908–1600 nm (11,013–6250 cm⁻¹). An external white reference (>99% reflectance, achieved with a disk of compressed polytetrafluoroethylene) and a black reference (<1% reflectance) were scanned every 10 min. Spectra collection was performed using two micro-NIR accessories. (a) Static measurements were achieved using the windowed collar. It features an integrated anti-reflective coated sapphire window, making it ideal for non-rigid materials. The window surface is directly in contact with the sample and lies about 3 mm away from the micro-NIR input, ensuring optimum signal-to-noise ratio performance. Five spectra were registered with an integration time of 9 ms and 500 accumulations in different positions of each sample and averaged. (b) The micro-NIR sample spinner is an accessory intended for measuring large particle-sized materials and/or heterogeneous samples. The samples were placed into a 9 cm diameter polystyrene Petri dish placed over the accessory and rotated at 10 rpm. Spectra were registered during28 s (25 ms of integration time and 1000 accumulations); at least 4 complete rotations of the sample were achieved. Three replicate measurements were performed by refilling the Petri dish.

2.4. Chemometrics

Various spectral preprocessing techniques and their combinations were evaluated: baseline correction, mean centering, standard normal variate (SNV), and detrending (DT) for scatter correction, along with first- and second-derivative treatments. Data were divided into two sets: a calibration set containing 101 samples and a validation set with 25 samples. Modeling was carried out using partial least squares (PLS) and support vector machine (SVM) regression algorithms. All calculations were performed with Solo+MIA software version 8.9.1 (Eigenvector Research, Inc., Manson, WA, USA).

The PLS regression method was used to build predictive models by calculating a small number of orthogonal factors, named latent variables (LVs), as new inputs for predicting response Y. The LVs are linear combinations of such original predictor variables that explain the maximum covariance between X and Y. Here, the number of independent LVs was obtained using cross-validation by random subsets (5 groups and 5 iterations). To avoid model overfitting, the maximum potential LVs for the PLS models was set to 10. Outliers were identified and removed using the values of Hotelling T² and the Quadratic Sum of the Residuals.

The SVM is based on statistical learning theory and offers a non-linear correlation approach. The main idea of SVM is to generate a model hyperplane that separates data points into different classes. Thus, function f is estimated, for which the maximum deviation between the target and input variables does not exceed a given threshold, denoted by ε. The region enclosed by the upper and lower deviation values is named the ε tube, and the points that lie on its boundaries are the support vectors. As many regression problems are highly non-linear, kernel functions are applied to map data objects into a higher-dimensional feature space [29]. In SVM, several parameters are responsible for the model performance, i.e., the cost parameter (C, which allows for the penalizing of the error by determining the tradeoff between the training error and the model complexity), the choice of kernel function, and its corresponding parameters of gamma ($\gamma$). A Gaussian radial basis function (RBF) kernel was used, as it is a non-linear function and a more compact supported kernel. C and $\gamma$ were optimized with a random subset cross-validation process, set as C = 100 and $\gamma$ = 10.

The performance of the models was compared through the determination of calibration (Rc²) and prediction (Rp²) coefficients, and the root-mean-square error of calibration (RMSEC) and prediction (RMSEP). The residual predictive deviation (RPD) was calculated as the ratio between the standard deviation in the prediction set and the RMSEP. Values of RPD greater than 2 can be considered as good predictions [30]. Bland–Altman plots were used to evaluate the consistency of the models’ predictions against reference values [31]. Each plot displays the difference between predicted and reference values on the y-axis and the average of the two values on the x-axis. The Passing–Bablok regression was also employed to assess the accuracy of the predicted values in comparison with the reference measurements [32]. This non-parametric method evaluates whether the intercept of the regression lines is non-zero and whether the slope deviates from unity. Therefore, it provides an estimation of the potential systematic (intercept) and proportional (slope) biases.

3. Results

3.1. Physical–Chemical Characterization of Calibration and Prediction Sets

The 126 samples of olive mill pomace compost were initially subjected to physical–chemical characterization.

Table 1. Statistical analysis of the compost quality parameters in the calibration and prediction sets.

Quality Parameters	Calibration Set (101 Samples)		Prediction Set (25 Samples)
	Range	Mean ± SD	Range	Mean ± SD
pH	5.3–9.8	7.3 ± 0.9	5.7–8.7	7.1 ± 0.9
EC25 (dS m−1)	2.9–7.9	5.2 ± 0.8	2–8	5 ± 1
C/N	15–36	22 ± 4	15–37	22 ± 6
LOI (%)	62–87	76 ± 5	68–85	77 ± 4

EC₂₅: electrical conductivity (25 °C); C/N: carbon-to-nitrogen ratio; LOI: organic matter expressed as “loss-on-ignition”.

provides information on the ranges of the studied parameters for both the calibration and the prediction sets. Similar variability between the two olive harvest seasons (2020/2021 and 2021/2022) was also considered when building the two sets (Figure S1, Supplementary Materials). It is noteworthy that using samples from two harvest seasons allows for a wider range of values in the dataset for a more robust and useful calibration.

In all cases, the ranges of minimum and maximum values were quite wide, which can be attributed to the different stages of maturation and heterogeneity of the materials. It is the case of pH ranging from acidic to values above 9, which were sometimes reached in the early stages. However, at the end of the composting process, all the final materials stabilized between 6.5 and 8, being moderately or weakly alkaline, similar to those reported by previous research [33,34]. Values of electrical conductivity tended to grow in all cases versus composting weeks, as a consequence of organic matter mineralization. Maximum values were found during mid-stages ($~$8 dS m⁻¹), decreasing slightly until around 6.5–7 dS m⁻¹ at the end of the processes, probably due to leaching.

The C/N ratio is considered a significant factor that influences the composting process since it plays an important role in nutrient balance for microorganisms. Initial C/N values were around 30, being ideal to enable microorganisms’ growth and to guarantee the carbon energy supply. During the initial weeks of composting, the ratio consistently decreased, typically reaching values below 20 at the end of the processes, which agree with the values set by the current Spanish regulation for ready-to-use compost [35]. In addition, a slight decrease in LOI was observed, from levels above 80% to 60–70% during the two olive harvest seasons. These final values are significantly higher than those found in the experimentally produced compost of Albuquerque et al. [1]. This can be explained by the relatively large proportions of OTP and straw in the mixtures, possibly reflecting incomplete organic matter degradation of the oversized particles and their elevated lignin content.

3.2. Spectral Characteristics

The typical NIR spectra of compost samples obtained with benchtop and handheld instruments are shown in Figure 1. FT-NIR spectra were obtained under optimal laboratory settings using a benchtop instrument and finely ground samples. In turn, the analytical performance of the portable spectrometer was tested by recording the spectra under different conditions, thus studying the possible influence of sample pretreatment and the two micro-NIR accessories. Finely ground and intact samples were analyzed using the windowed sapphire collar, enabling direct contact analysis at the different areas of interest. Furthermore, the spinner accessory was used to analyze intact samples accounting for spectral differences due to material heterogeneity and differences in particle size.

It is evident that FT-NIR spectra present a wider spectral range involving more information than the handheld one. Thus, the two sharp bands located at 2308 and 2348 nm, together with those at 1726 and 1760 nm, are related to contributions of C-H aliphatic chains from hydrocarbons, cellulosic materials, and lipids [36]. In addition, the broadband located at around 2127 nm is attributed to combination bands involving C–H, C=C, and C=O stretching vibrations of lipids, phenols, and other aromatic compounds like lignin [37,38]. The band centered at 1933 nm has been also assigned to O-H vibrations of water and polysaccharides [36]. In the common spectral range (1000–1600 nm), similar broad features were observed for both instruments. This region is mainly dominated by water absorption bands due to O-H vibrations at 1205 and 1450 nm (first and second overtone), as well as contributions of the third overtone of O-H from non-hydrogen-bonded short-chain alkyl alcohols (around 960 nm), and aliphatic methylene groups from hydrocarbons (930 nm) [36]. Contributions of methyl groups appeared around 1190 nm.

As can be observed (Figure 1b), similar patterns were found between the spectra corresponding to finely ground and intact samples. Nevertheless, when replicates are considered, the variability increases in the case of static measurement of intact samples as can be seen in the higher standard deviation values (bottom of Figure 1b). When using a spinner accessory, the rotation helped to reduce the variability to similar levels to those obtained with the FT-NIR with finely ground samples (bottom of Figure 1a,c).

3.3. Quantitative Prediction of Compost Quality Parameters

3.3.1. Spectral Preprocessing

Before establishing the prediction models, preprocessing of the spectral data is of great importance for diminishing the systematic errors that can be caused by different factors unrelated to chemical composition [20], such as light scattering and differences in the effective path length, which may induce baseline shifts (multiplicative effects). Basic preprocessing of NIR spectra includes baseline offset correction (BL) together with mean centering (MC), which substantially enhance sample differences both in terms of concentration and spectral response by subtracting the average value of each variable. Furthermore, derivatives (DVs), mainly using Savitzky–Golay smoothing, are widely used to remove both additive and multiplicative effects in spectra and improve the resolution of overlapping bands. Finally, multiplicative scatter correction (MSC), detrending (Dtr), and standard normal variate (SNV) are the most common methods for the correction of light scattering effects. The main concept behind them is that artifacts or imperfections will be removed from the dataset before modeling [39].

Thus, PLS and SVM regression models for the prediction of compost quality parameters were built for both instruments using several combinations of the above-mentioned preprocessing methods. Table S1 (Supplementary Materials) shows some of the results obtained with PLS to explore the performance of the models when predicting C/N and LOI as illustrative examples. The use of the most basic preprocessing using BL+MC, even when combined with the first and second derivatives, resulted in RMSEP ranging from 0.8% to 2% of the mean value for the prediction set in the compost quality parameters. Specifically, the RMSEP for C/N was 21.6 ± 0.2 and, for LOI, 77.6% ± 0.6%. These results, coupled with RPD values consistently below 0.5, indicate inadequate performance for practical applications. Better results were obtained when using scattering correction methods like detrending (Dtr), typically employed for baseline shift and curvilinearity adjustment of powdered samples. MSC typically leads to a less pronounced baseline correction compared to the combination of SNV and Dtr, so this last combination was henceforth chosen. Furthermore, the performance improved after adding the first derivative (SNV + Dtr + 1st DV), since the NIR spectra presented large band overlapping. This optimal preprocessing provided the best RMSE and coefficients of determination values, with RPDs above 1.5 in most of the cases, so it was selected for the rest of the analysis.

3.3.2. Prediction Performance Using FT-NIR and Micro-NIR Using Different Regression Algorithms

FT-NIR spectra of the finely ground compost samples were obtained under optimal laboratory conditions.

Table 2. Results of PLS and SVM regression models using FT-NIR spectra for different compost quality parameters.

	PLS Model							SVM Model
	LVs	Rc2	RMSEC	Rp2	RMSEP	RPD	Bias	Rc2	RMSEC	Rp2	RMSEP	RPD	Bias
pH	6	0.81	0.44	0.72	0.51	1.9	0.07	0.99	0.13	0.91	0.26	3.8	0.05
EC25 (dS m−1)	7	0.88	0.39	0.85	0.48	2.4	0.2	0.89	0.38	0.79	0.55	2.1	0.1
C/N	7	0.81	1.7	0.52 (0.78)	3.8 (2.4)	1.5 (2.5)	0.4 (1.5)	0.83	1.7	0.54 (0.78)	3.7 (2.3)	1.5 (2.5)	0.5 (1)
LOI (%)	7	0.93	1.4	0.81	2.0	2.7	−0.4	0.92	1.7	0.82	2.1	2.5	−0.9

EC₂₅: electrical conductivity; C/N: carbon-to-nitrogen ratio; LOI: organic matter expressed as “loss-on-ignition”. PLS: partial least squares regression; SVM: support vector machine regression; LVs: latent variables; Rc²: determination coefficient of calibration; RMSEC: root-mean-square error of calibration; Rp²: determination coefficient of prediction; RMSEP: root-mean-square error of prediction; RPD: residual predictive deviation. Spectra preprocessing was performed using SNV + Dtr + 1st DV. For C/N ratios, the values in parentheses correspond to validation with 23 samples.

shows the comparison of the models obtained using PLS and SVM for the prediction of the different quality parameters. A Passing–Bablok test was used to compare reference measurements with predicted values for all the parameters and models. As both 1 and 0 were consistently within the confidence intervals for the slope and intercept, respectively, we conclude that the FT-NIR predicted values do not differ from the reference values (at 95% confidence). Thus, there is no evidence of systematic error (constant or proportional) in the FT-NIR predictions.

The prediction accuracy of the model for LOI with PLS was slightly higher compared with the others, achieving an RPD of 2.7. In general, RMSEP values were similar using both models except for pH. In this case, the SVM model showed a significant improvement in predictive accuracy with a substantial increase in Rp² (0.72 to 0.91), a nearly halved prediction error, and an excellent RPD value (>3) compared to the PLS model. The non-linear relationship between pH and concentration can explain this superior performance of the SVM model. Only in the case of the C/N ratio, poorer prediction performance was observed with an RPD of around 1.5 and RMSEP values much higher than RMSEC ones. This can indicate a possible model overfitting. A deeper analysis of this point was achieved by the inspection of the residuals using a Bland–Altman plot (see Figure 2), which shows the differences between the reference measurement and the predicted value against their averages for the validation set. This revealed no trends for all the parameters and models with the points randomly scattered around zero. However, the presence of two samples with anomalous residuals for both PLS and SVM models in the case of C/N prediction was detected.

These two samples present very high values of C/N ratio (>35) that were underestimated in the prediction. This suggested that samples with such high C/N ratios were under-represented in the calibration set, leading to an increase in RMSEP not caused by model overfitting. Nevertheless, as previously mentioned, the initial C/N ratio, a crucial factor impacting the composting process, should be ideally between 25 and 30 [40] and decreases during composting. Although there are some samples over 30 in the calibration set, these particular values of such high C/N ratio are quite unusual, and therefore, the possibility of errors in the reference measurements cannot be entirely ruled out. Removing these two outlier samples significantly improved the model’s performance, reducing the RMSEP by approximately 45% for both models. The new RMSEP values are now closer to the RMSEC values, as observed for the other parameters (see Table 2). Therefore, these two samples were excluded from further analysis for the C/N parameter.

When using portable instruments, a careful design of measurement protocols for a particular application is of importance. Here, static measurements for different sample presentations, intact and finely ground, together with the use of a rotation accessory for the measurement of intact samples, were studied (Figure S2). The results of the different regression models are presented in

Table 3. Results of PLS and SVM regression models using micro-NIR spectra for different compost quality parameters. Different sample statuses and accessories are included.

	Sample Status–Mode	PLS Model							SVM Model
		LVs	Rc2	RMSEC	Rp2	RMSEP	RPD	Bias	Rc2	RMSEC	Rp2	RMSEP	RPD	Bias
pH	FG-S	5	0.58	0.66	0.46	0.70	1.4	0.18	0.77	0.49	0.51	0.61	1.6	0.01
	I-S	9	0.66	0.59	0.46	0.66	1.5	0.09	0.77	0.49	0.49	0.72	1.4	0.18
	I-R	10	0.81	0.45	0.54	0.62	1.6	0.12	0.73	0.53	0.62	0.58	1.7	0.12
EC25 (dS m−1)	FG-S	5	0.64	0.67	0.69	0.56	2.1	0.15	0.71	0.62	0.69	0.61	2.0	0.30
	I-S	8	0.71	0.61	0.82	0.57	2.1	0.04	0.85	0.45	0.77	0.56	2.1	0.05
	I-R	8	0.74	0.58	0.57	0.82	1.4	−0.03	0.72	0.57	0.64	0.66	1.8	0.24
C/N *	FG-S	9	0.65	2.4	0.76	2.9	2.0	2.1	0.74	2.1	0.70	3.1	2.0	1.3
	I-S	7	0.62	2.5	0.53	2.8	2.4	1.4	0.66	2.4	0.55	2.6	2.2	1.0
	I-R	9	0.66	2.2	0.54	2.4	2.4	1.4	0.55	2.4	0.57	1.7	3.4	0.19
LOI (%)	FG-S	9	0.63	3.4	0.48	3.9	1.3	−2.1	0.83	2.3	0.57	3.5	1.5	−1.6
	I-S	7	0.43	4.3	0.42	4.0	1.3	−0.92	0.73	2.8	0.59	3.2	1.7	−1.5
	I-S	8	0.58	3.4	0.66	3.4	1.6	−2.0	0.63	3.2	0.52	3.4	1.6	−2.0

FG: finely ground; I: intact; S: static; R: rotating; EC₂₅: electrical conductivity; C/N: carbon-to-nitrogen ratio; LOI: organic matter expressed as “loss-on-ignition”. PLS: partial least squares regression; SVM: support vector machine regression; LVs: latent variables; Rc²: determination coefficient of calibration; RMSEC: root-mean-square error of calibration; Rp²: determination coefficient of prediction; RMSEP: root-mean-square error of prediction; RPD: residual predictive deviation. Spectra preprocessing was performed using SNV + Dtr + 1st DV. * C/N ratio results correspond to validation with 23 samples.

.

The RMSEP values for pH with the micro-NIR spectrometer are similar for both the PLS and SVM models, being slightly better in the case of rotating measurements. However, the prediction performance is notably inferior (with RMSEP increasing over 50%) than the one obtained with the laboratory instrument using the SVM model.

For electrical conductivity EC₂₅, similarly good results were obtained using the two chemometric algorithms (with RPD > 2). The measurement of intact samples implies more complex models (with a higher number of latent variables in PLS) dealing with different light effects for more heterogeneous samples. In this case, the use of rotation during measurement seemed to increase the prediction errors.

The use of micro-NIR succeeded for the prediction of C/N ratio, using the spinner accessory for the measurement of intact samples. Thus, the RMSEP values were comparable to or even better than those of the laboratory instrument, especially with the SVM model. Nevertheless, the inspection of the residuals, with the corresponding Bland–Altman plots (Figure S3, Supplementary Materials) reveals, in fact, a wider 95% limit of agreement interval for micro-NIR measurements in comparison to FT-NIR spectra (Figure 2), even when the RMSEP value is lower. This suggests higher discrepancies in measured predicted values. For LOI prediction, relative RMSEP values, although higher than FT-NIR, are still under 5%.

In addition, all the predictions were evaluated with the Passing–Bablok test, comparing to reference measurements. Only in the case of the C/N ratio when spectra were measured on static mode and using the SVM model, the results revealed the underestimation of the value of the parameter for high C/N ratios (slope < 1). The rest of the models for all the parameters have no signs of any systematic error (see examples in Figure 3).

4. Discussion

In this study, the predictive capacity of NIR spectroscopy, using benchtop FT-NIR and handheld micro-NIR spectrometers, was evaluated to assess pH, EC₂₅, C/N, and LOI during the composting maturation process of OMP. For this purpose, models have been developed using both PLS and SVM regression tools. The analytical performance of the micro-NIR spectrometer using two different measurement accessories, the sapphire windowed collar and the spinner (Figure S2), is compared here for the first time. In addition, the influence of the sample grinding was evaluated.

For FT-NIR, satisfactory results were obtained for all the parameters considered, namely, pH, EC₂₅, C/N, and LOI, with RPD values higher than 2.0 in all cases. Similar PLS models were obtained by Galvez-Sola et al. [15] from several compost samples using waste from the winery and distillery industry, covering the 1000–2000 nm range using a benchtop FT-NIR spectrometer. Using the handheld spectrometer for samples measured in similar conditions (finely ground) led to higher RMSEP for all the parameters. Taking into account that the spectral range registered with the portable instrument is considerably narrower, it is interesting to study if the drop-in prediction performance is substantially due to this fact. To carry this out, models with FT-NIR spectra restricted to the spectral range covered by the micro-NIR spectrometer (1000–1600 nm) were also built (Table S2, Supplementary Materials). In general, the higher the wavelengths, the more sensitive and specific are the bands, thus the full range of the FT-NIR instrument range always provided more accurate results, whereas other factors related to portability like lower spectral resolution were not so relevant. Comparing the RMSEP values, it can be concluded that the reduction in the spectral range only had the most important effect in the decrease in the prediction performance of pH. This behavior could be attributed to the influence of carbonates, which have been found to increase during the compositing process of this type of material [19]. In comparison with other functional groups like organic acids or ammonium, affecting the pH of compost, carbonates have weak absorption bands in the NIR region, and all are located above 1600 nm. Thus, for calcite, the strongest is about 2335 nm, and some weaker absorptions occur near 2160 nm, 1990 nm, and 1870 nm [41]. The availability of the chemical information in the spectral region observed by a given sensor has already been reported as one of the most important factors influencing predictive performance [23]. Nevertheless, the influence of the narrowed spectral range was not so evident for the rest of the parameters. Thus, our findings suggest that a limited spectral range in the overtone region, where C-H, N-H and O-H bonds absorb in multiple locations, can still yield highly accurate predictions. In particular, the best prediction models obtained with micro-NIR for EC₂₅ resemble those obtained with FT-NIR and are similar to data reported in previous studies on compost monitoring with benchtop NIR spectroscopy [15,16]. EC₂₅, which is related to dissolved salt concentration, varies significantly in this material type between 1.5 and 10 dS m⁻¹ [42,43]. This property is probably estimated indirectly, using spectral information over the wavelength range covered by the micro-NIR instrument.

Furthermore, support vector machine (SVM) regression provided, in general, slightly superior performance than partial least squares (PLS) regression, showing lower values of RMSEP for most of the parameters determined, although the difference is not very significative. In fact, the performance of SVM in comparison with PLS was only significantly better for the prediction of pH using the full spectral range in FT-NIR. SVM can perform non-linear regression efficiently for high-dimensional datasets. Thus, the ability of a non-linear SVM algorithm to compensate for non-linearities in the spectral response by implicit modeling of non-linear effects is probably the reason for this, as reported by other authors [10]. In comparison with the benchtop spectrometer, the PLS models for the micro-NIR spectra required a greater number of latent variables. Such complexity could be explained, for example, by effects derived from the background acquisition. In FT-NIR, the reference is measured automatically using the internal gold reference of the sphere, while the portable instrument requires the operator’s assistance for the measurement of an external reference, which can affect the spectral quality. Whereas SVM non-linear modeling can easily deal with these effects, PLS models must increase their complexity (using more latent variables) to achieve a good fitting [12].

Despite these drawbacks and the higher values of RMSEP in comparison with the benchtop FT-NIR spectrometer, all the models obtained with micro-NIR offer a fast and easy prediction of the studied parameters, providing values comparable to those obtained by reference measurements, as demonstrated by the results of the Passing–Bablok regression, and are consistent with those found in the literature [15,16,17].

Finally, when using handheld spectrometers, the possibility of acquiring information on samples with minimal sample preparation in order to perform analysis in situ for fast process control is of particular interest. Olive mill pomace-based compost samples present high physical and chemical heterogeneity, so the spectral data acquired on unmilled samples may not represent the average composition of the material, which harms the performance of multivariate regression models [26]. Here, three typical strategies for increasing the representativeness of the final spectrum have been considered: (i) reduction in the heterogeneity of the sample by intensive grinding, which is laborious and time-consuming; (ii) the use of average spectra obtained from five static measurements collected at different sample positions; and (iii) the use of a spinner accessory to rotate the sample during spectral acquisition. We have demonstrated that it is possible to obtain accurate predictions without grinding, as can be seen in Figure 3, which shows the results for the measurement of intact samples both in the static mode (with the windowed sapphire collar) and when rotating with the spinner accessory. The Passing–Bablok regression provides similar results for both static and rotating modes, disregarding the presence of either constant or proportional systematic errors. However, RMSEP values do not present significant differences. Probably, the use of the sapphire collar for the static measurements reduces spectral variability, since it sets the optimum working distance, avoiding errors due to incorrect sample positioning as observed for other handheld devices [44]. In addition, the direct contact with the sample can also minimize effects due to ambient light influence. In these conditions, the use of the average spectrum of five replicate measurements in different positions compensates for sample heterogeneities and yields reasonable prediction accuracy.

5. Conclusions

Effective waste management is a crucial issue for the olive oil industry, composting being an interesting way to contribute to a circular regenerative agricultural model. As shown here, NIR spectroscopy offers a potential solution for rapid and efficient compost quality assessment, avoiding the time-consuming analyses of the different physical–chemical parameters with the additional benefit of being eco-friendly (since it does not use chemical reagents). Furthermore, our results demonstrated the feasibility of using a miniaturized spectrometer to determine several physical–chemical parameters in either intact or ground compost samples. Although the handheld spectrometer showed lower predictive performance than the benchtop high-end instrument using the full spectral range, the models presented a reasonable fit. Therefore, based on the evaluation of the statistical parameters (RMSEP, RPD, and Passing–Bablok regression), all models could be considered good and reliable for analytical tasks. Thus, miniaturized NIR spectrometers can become true spectral sensors that can help in the compost process management, particularly for large-scale composting facilities. The feasibility of this approach has been here demonstrated on the example of olive mill waste compost, but the technology could be easily implemented for other agricultural or even urban waste. In fact, the use of NIR sensors could also be of great interest for evaluating possible contaminations in raw waste (for example, with plastics) as well as for decision-making during the composting process and effective in situ discrimination between materials in optimal stages (sanitized and stabilized) to be used as organic amendments.