Material

Three brands of nam pla-ra, each with a different sodium chloride and protein content (low, medium and high), were chosen from regional and local supermarkets in Bangkok and Pathum Thani Province, Thailand. They were used without further treatment and mixed using the mixture design technique, resulting in a total of 60 samples (Table S1). An additional forty brands from the market were obtained. All samples (N=100) were used to develop the calibration and prediction models for the physicochemical values of nam pla-ra. Information on nam pla-ra manufacturers, is given in Table S2.

FT-NIR analysis

A volume of 30 mL of nam pla-ra from each sample was placed in separate plastic tubes and incubated in a water bath at 30 °C for 30 min. Each sample was then placed in a glass Petri dish with a transflectance cover and analysed using a Fourier transform near-infrared (FT-NIR) spectrophotometer (model NIR Flex N-500; Buchi, Flawil, Switzerland) in transflectance mode. All spectra were collected in the wavenumber range 10 000–4000 cm^-1 (1000–2500 nm) with 64 times per scan and a resolution of 8 cm^-1. A single spectrum was produced for each sample by averaging the three acquired spectra.

Physicochemical characterisation of nam pla-ra by reference methods

The physicochemical properties of nam pla-ra, namely total soluble solids (TSS), pH, colour (L*, a*, b*), sodium chloride and protein content, were determined using standard methods. Three measurements were carried out per sample.

TSS were measured using a pocket refractometer (model PAL-1; Atago, Tokyo, Japan). The pH value was determined at 25 °C using a pH meter (model CG842; Schott, Mainz, Germany). The colour (L*, a*, b*) of each sample was measured using a spectrophotometer (model CM-700d; Konica Minolta, Osaka, Japan).

To determine the NaCl content, a sample (1 g) was weighed into a beaker with 200 mL of deionized water. The mixture was adjusted to neutral pH with 0.1 M sodium hydrogen carbonate solution (Elago Enterprises Pty Ltd, Castle Hill, NSW, Australia) and the volume was brought to 250 mL with deionized water. A sample (10 mL) was pipetted into an Erlenmeyer flask, and 1 mL of 5 % potassium chromate solution (Elago Enterprises Pty Ltd) was added. The sample was titrated with 0.01 M silver nitrate solution (Avantor Performance Materials Poland S.A., Gliwice, Poland) until the end point (brick red sediment). The NaCl content was calculated using the following equation (15):

/1/

where V1 and V2 are the volumes of silver nitrate solution titrated with the sample (mL) and the blank (mL), respectively, c is the concentration of silver nitrate solution (M), and 0.05844 is the molar mass of NaCl expressed in g/mmol, which is used as a conversion factor to calculate the NaCl content.

The protein content was determined according to the AOAC method (16) using the Kjeldahl method. A sample (2 g) was weighed into a protein digestion tube. Then, 10 g of selenium mixture and 15 mL of sulfuric acid were added. A blank was prepared in parallel using distilled water instead of the sample. The samples were digested and allowed to cool to room temperature. After digestion, the samples were distilled by adding 25 mL of 4 % boric acid solution into a 250-mL flask containing 2 drops of indicator solution, which was used to collect the distillate. The distillate was then titrated with 0.1 M hydrochloric acid solution until the end point was reached. The volume of 0.1 M hydrochloric acid solution consumed was recorded. The protein content was calculated using the following equations:

/2/

/3/

where V1 and V2 are the volumes of hydrochloric acid solution used in the titration of the sample (mL) and the blank (mL), respectively, c is the concentration of hydrochloric acid solution (M), and m is the sample mass (g).

NIR spectroscopic model development

Physicochemical values of nam pla-ra by reference method

Table 1 shows the properties of nam pla-ra as determined by the reference technique, including total soluble solids (TSS), pH, colour (L*, a*, b*), sodium chloride and protein content. The physicochemical properties (TSS, pH, colour, NaCl and protein content) and the spectral results were divided into two groups to develop the predictive model, with a calibration set (N=75) and a prediction set (N=25). The ranges of all physicochemical reference values in the calibration set fully encompassed those observed in the prediction set.

Results of FT-NIR analysis

The NIR spectra of nam pla-ra samples were obtained using an FT-NIR spectrophotometer in transflectance mode and the spectral profiles are shown in Fig. 1a.

Fig. 1

NIR spectra of: a) Original nam pla-ra product, and b) after the pretreatment with second derivative Savitzky-Golay method

Fig. 1b shows the NIR spectra of samples with the highest and lowest protein content, recorded on FT-NIR and pretreated with the second derivative Savitzky-Golay method for important band assignment, while Table 2 (19-24) shows the band assignment data. Water is a key component in nam pla-ra. At 1403, 1463 and 1877–1900 nm, it exhibited strong absorption bands resulting from the combination of O-H overtone, stretching, and bending wavelengths (19). The bands associated with pH (lactic acid from the fermentation and the ionization of lactate from lactic acid) were observed at 1710, 1722, 1730, 1750, 1890, 2160 and 2470 nm as the first overtones of C-H stretching, O-H stretching combined with C=O stretching, C-H stretching, C=O stretching combination, and C-H stretching combined with C=O stretching (20–24). The absorption bands for protein in the spectral region 2050–2168 nm resulted from the N-H stretching combination and the N-H band second overtone (24). At 2242 and 2472 nm, combinations of C-H stretching, CH₂ deformation, C-C stretching combination, and C-O-C stretching combination were related to sugar (24). Furthermore, although NaCl lacks unique NIR absorption bands, the effect of salt on the water absorption band enabled the detection of changes in salt content. The strength of the water band absorption decreased as the salt content increased (19, 25). For TSS, this can be attributed to NaCl, as the main soluble solid dissolved in nam pla-ra. NIRS does not directly measure the L*, a*, b* colour parameters of fermented fish sauce as human vision or a colorimeter does. Instead, it measures light absorption or reflection in the near-infrared range, which is indirectly correlated with the chemical composition and physical structure affecting the colour of the product through precise chemometric modelling. In nam pla-ra, the brown to dark brown colour is due to compounds containing C-H, O-H, and N-H bonds, which are detectable by NIRS. Variations in these colour-forming compounds alter the NIR spectrum. Specifically, their concentrations, which determine L*, a* and b* colour values, correlate with light absorption in the NIR range.

Calibration models for physicochemical values were developed using PLS regression. The calibration and prediction statistics are shown in Table 3, and scatter plots comparing the measured and predicted values for each parameter are shown in Fig. 2 (models using the full spectrum in the range 1000–2500 nm) and Fig. 3 (models using the selected spectrum in the ranges 1000–1889 and 2031–2408 nm to avoid overlapping and over-absorption bands (26)). The performance of the PLS calibration models was evaluated using the coefficient of determination (R²), the standard error of calibration (SEC), and the standard error of prediction (SEP). As shown in Table 3, the number of factors indicates the optimal number of latent variables used in the PLS regression models, selected to balance model accuracy and complexity while avoiding overfitting. In this study, the optimal number of factors ranged from 4 to 22, depending on the physicochemical parameter and the spectral preprocessing method applied. The use of an appropriate number of factors helped minimise prediction errors, indicating good model stability and generalisation capability. Overall, the calibration models showed relatively high R² values, demonstrating strong agreement between the measured and predicted physicochemical properties. In addition, the SEC and SEP values were generally low and comparable for most parameters, suggesting satisfactory calibration accuracy and reliable prediction performance. For the full spectrum model, the original spectrum produced the best results for pH, colour (L* and a*) and NaCl. The second derivative of NIR spectra using the Savitzky-Golay method was optimal for colour (b*), while the standard normal variate (SNV) method was optimal for TSS and protein content. For the selected spectrum model, the original spectrum was optimal for TSS, colour (b*) and NaCl. The second derivative of NIR spectra using the Savitzky-Golay method was optimal for pH and protein content, while the SNV method was optimal for colour (L* and a*).

Fig. 2

Scatter plots for comparison of measured and predicted values (from full spectrum) for nam pla-ra qualities of prediction set for: a) TSS, b) pH, c) L*, d) a*, e) b*, f) sodium chloride content, and g) protein content. TSS=total soluble solids

Fig. 3

Scatter plots for comparison of measured and predicted values (from selected spectrum) for nam pla-ra qualities of prediction set for: a) TSS, b) pH, c) L*, d) a*, e) b*, f) sodium chloride content, and g) protein content. TSS=total soluble solids

The comparison between the full spectrum and selected spectrum models showed that the best predictive model from the full spectrum included pH and colour (L* and a*). In contrast, the best predictive model from the selected spectrum was developed using TSS, colour (b*), NaCl, and protein content. The full spectrum provided the best model for physical properties, since colour may arise from physical properties that do not have distinct absorption bands in the NIR. The best model for physical parameters was produced using the full spectrum (representative of all functions) predictive model, while the selected spectrum gave the best model for chemical properties (TSS, NaCl and protein content), possibly because these properties corresponded to specific absorption bands. Thus, the selected spectrum (without the broad band for water) provided the best model for specific absorption bands. Furthermore, the predictive models for pH and colour (L*, a*, b*) may be used for screening (R²=0.73–0.81); the protein content models can be used for research and general purposes (R²=0.86); and the TSS and NaCl models can be used for all types of work (R²=0.97–0.98) (27).

Compared with previous studies, Ritthiruangdej et al. (11) applied NIRS in the spectral range of 1100–2500 nm to predict total nitrogen (TN), pH, refractive index, density and Brix in Thai fish sauces. Their results showed that the informative spectral regions selected by searching combination moving window partial least squares (SCMWPLS) were strongly associated with specific chemical properties, including the regions of 2264–2428 nm for TN, 1698–1722 and 2222–2258 nm for pH, 1358–1438 nm for density, 1774–1846 and 2078–2114 nm for refractive index, and 1322–1442 and 2000–2076 nm for Brix. In addition, qualitative classification models based on TN content achieved correct classification rates exceeding 82 %. These findings are consistent with the results of the present study, demonstrating that variable selection and the use of informative spectral regions can improve the prediction of chemical attributes related to distinct NIR absorption bands, while full spectrum models are more suitable for predicting physical properties. Similarly, Jiang et al. (28) investigated the application of dual-band NIRS for the non-destructive determination of fat, protein, collagen, ash and sodium contents in meat stewed in soy sauce. Using spectral ranges of 650–950 and 960–1660 nm, PLS models were developed based on spectra collected from vacuum-packed ready-to-eat products from 97 different brands. The results demonstrated that fat and protein contents were predicted with the highest accuracy, with superior model performance obtained in the 960–1660 nm region compared to the 650–950 nm range. These findings are in agreement with the present study, emphasising that the selection of appropriate spectral regions enhances predictive performance for chemical constituents associated with specific absorption bands. Furthermore, Zhang et al. (29) collected NIR spectra in the range of 1000–2500 nm using a miniature fibre-optic NIR spectrometer to monitor different stages of soy sauce production, including steamed soybean, koji and moromi. Their study successfully developed predictive models for in situ and real-time assessment of the digestion rate of steamed soybean, protease activity in koji, and formaldehyde nitrogen content in moromi. These results further support the suitability of NIRS for predicting key chemical attributes in fermented food systems, which is consistent with the findings of the present study. Nevertheless, there have been no reported studies on the quantification of physicochemical properties of nam pla-ra using NIRS.

Predictive model validation

The predictive model validation showed that the developed model accurately predicted nam pla-ra physicochemical properties (TSS, colour (L*, a*, b*), NaCl and protein content). There was no significant difference between the physicochemical values obtained from the predictive model and those from the reference method, except for the pH value (Table 4) possibly because the pH value in the samples had a relatively narrow range. The pH value (Table 1) had the lowest standard deviation value (0.38) compared to other physicochemical values. Therefore, the accuracy of the pH value may still be affected by the prediction model, even if it has a low bias value. Consequently, the outputs of the prediction model and the reference method were different.