Non-Destructive Measurement of Caffeine Content on Intact Coffee Beans Using Near Infrared Reflectance Spectroscopy

Arabica coffee (Coffea arabica L.) is one of the most popular coffee varieties in the world. This coffee is known for its smooth, complex taste and rich aroma. Arabica coffee is generally grown in the highlands at an altitude of around 600-2000 meters above sea level [1]. These unique characteristics make Arabica coffee more favored by international consumers. In addition, Arabica coffee has a lower caffeine content compared to Robusta coffee, making it a preferred choice for consumers seeking a milder and less bitter taste [2].

Caffeine is an alkaloid compound naturally found in coffee, tea, chocolate, and some other plants. This compound functions as a stimulant in the central nervous system, increases alertness, and can reduce fatigue. Caffeine has an important role in the coffee industry, because it not only affects the taste of coffee, but also the physiological effects on consumers. In Arabica coffee, the caffeine content ranges from 0.8% to 1.5%, much lower than that of Robusta coffee, which has a caffeine content of around 1.7% to 2.5% [3].

The importance of caffeine in coffee is not only related to flavor and aroma, but also to quality standards. Many consumers are concerned about caffeine levels when choosing coffee, due to its physiological effects. High caffeine levels can cause increased heart rate, anxiety and sleep problems, so coffee producers must ensure that caffeine levels are in line with consumer expectations and applicable quality standards. In this context, testing caffeine levels is important to ensure the quality of coffee products to be marketed [4].

Conventional methods for determining caffeine levels, such as High-Performance Liquid Chromatography (HPLC), require a long time, high cost, and complex sample preparation. HPLC works by separating the chemical components in a sample for identification and quantification, but due to the destructive nature of the process, it is less suitable for applications that require rapid, non-destructive analysis of the sample. Therefore, there is currently a need for alternative methods that are more efficient and non-destructive. Near-Infrared Spectroscopy (NIRS) is a technology that can answer this challenge. NIRS is a non-destructive analytical technology that can measure the absorption of infrared light by molecules in a sample. This technology does not require complicated sample preparation and can provide fast results. NIRS is widely used in the agriculture and food industries to predict chemical content, including moisture, protein, and caffeine content, as well as overall product quality. The advantage of NIRS lies in its ability to perform simultaneous and non-destructive analysis, making it highly efficient in controlling product quality [5].

Partial Least Squares Regression (PLSR) is one of the data analysis methods often used in conjunction with NIRS. PLSR allows researchers to build prediction models based on NIRS and sample chemistry data. This method is very useful in spectrum analysis, especially when the data has high multicollinearity [6]. PLSR works by relating spectrum data (X) to chemical variables (Y), such as caffeine content, thus enabling high-accuracy prediction of chemical composition.

However, spectrum data obtained from NIRS often contains noise and distortion, so spectrum correction (pretreatment) is required to improve the accuracy of the prediction model. Pretreatment is the first step in spectrum analysis that aims to remove noise or distortion from the raw spectrum data. Some commonly used pretreatment methods are standard normal variate (SNV), detrending (DT), extended multiplicative scatter correction (EMSC), and the Savitzky-Golay (SG) filter. Each method has a specific purpose, such as reducing the effects of scattering, baseline shift, and noise, so that the resulting data is more stable and accurate for use in PLSR prediction models [7, 8].

The NIRS and PLSR directly address the challenges in caffeine measurement highlighted earlier in the introduction. NIRS offers a rapid, non-destructive method for analyzing chemical composition, including caffeine content, in coffee beans. This technology overcomes the limitations of conventional methods like HPLC, which are time-consuming, costly, and require complex sample preparation. NIRS works by measuring the absorption of infrared light by molecules in a sample, providing a spectral fingerprint that correlates with its chemical composition [9].

However, the complex nature of NIR spectra necessitates sophisticated data analysis techniques, which is where PLSR comes into play. PLSR is particularly well-suited for analyzing NIR spectral data because it can handle the high multicollinearity often present in such data. It works by creating a linear regression model that relates the spectral data (X variables) to the chemical composition (Y variables, in this case, caffeine content). This allows for the development of predictive models that can accurately estimate caffeine levels based on NIR spectra alone, without the need for destructive chemical analysis [10, 11].

By combining NIRS with PLSR, we can develop a method that provides rapid results, addressing the time constraints of conventional methods. Moreover, it can preserve valuable coffee samples, since NIRS is a non-destructive method. It also requires minimal sample preparation, reducing complexity and cost [12]. Also offers high accuracy through sophisticated data analysis. These characteristics directly align with the study's objectives of finding an efficient, accurate, and non-destructive method for caffeine content prediction in Arabica coffee beans.

With the development of NIRS technology and pretreatment methods, testing the caffeine content of Arabica coffee green beans can be done more quickly, efficiently and accurately, without damaging the coffee samples. This provides great benefits to the coffee industry, especially in ensuring the quality of products that will be exported to international markets.

In the field of non-destructive analytical techniques, NIRS stands out as a particularly advantageous method for analyzing coffee beans, especially when compared to other prominent techniques such as Raman Spectroscopy, Hyperspectral Imaging, and X-ray Fluorescence (XRF). Raman Spectroscopy, while offering detailed molecular fingerprints and being less susceptible to water interference than NIRS, faces challenges with fluorescence interference and typically requires more sophisticated, costly equipment [13, 14]. Hyperspectral Imaging, which ingeniously combines spectroscopy with imaging capabilities, allows for spatial distribution analysis of chemical components but generates extensive datasets that demand complex processing and specialized equipment. XRF, though excellent for non-destructive elemental analysis, falls short when it comes to analyzing organic compounds like caffeine, which is crucial in coffee bean assessment [15].

NIRS, however, presents a unique combination of advantages that make it particularly well-suited for the coffee industry. Its versatility allows for simultaneous analysis of multiple chemical components, not limited to caffeine alone, providing a comprehensive view of the bean's composition. The speed of NIRS is unparalleled, delivering results in mere seconds, which is significantly faster than most other spectroscopic methods. This rapid analysis capability is crucial in high-volume coffee production and quality control scenarios.

From a financial perspective, NIRS equipment is generally more cost-effective compared to alternatives like Raman or Hyperspectral systems, making it a more accessible option for a wider range of coffee producers and researchers. The ease of use of NIRS systems is another significant advantage; they require minimal sample preparation and can be operated by individuals without specialized training, streamlining the analysis process in busy production environments. Furthermore, the portability of many NIRS systems, which are often compact and designed for field use, allows for on-site or in-line measurements, bringing laboratory-grade analysis capabilities directly to coffee farms or processing facilities [16, 17].

These combined advantages, coupled with NIRS's established track record in food and agricultural applications, position it as the ideal choice for studying caffeine content prediction in Arabica coffee beans. The technology's capacity to deliver rapid, accurate, and non-destructive analysis aligns perfectly with the coffee industry's growing need for efficient quality control methods in an increasingly competitive global market. As the coffee industry continues to evolve, with consumers demanding higher quality and more precise information about their coffee, NIRS technology offers a solution that can meet these demands while also improving production efficiency and consistency.

3.1 Spectra features of Arabica coffee bean

The raw spectrum of Arabica coffee green beans, as depicted in Figure 2, provides a wealth of information about the chemical composition of the samples. This spectrum, obtained through NIRS in the wavelength range of 1,000–2,500 nm, serves as a complex fingerprint of the beans' molecular makeup. Within this near-infrared region, various organic molecules present in coffee beans exhibit characteristic absorption patterns, allowing for non-destructive analysis of their composition.

Figure 2. Near-Infrared (NIR) spectra features of Arabica coffee beans in the 1,000–2,500 nm wavelength region

Important spectral features of the raw spectrum reveal significant insights into the beans' chemical structure. Notably, based on previous literature, broad absorption bands are observed around 1,450 nm and 1,940 nm, which can be primarily attributed to O–H stretching overtones [9, 20, 22]. These bands are particularly indicative of the moisture content within the coffee beans, a crucial factor in determining both quality and storage stability. A prominent peak near 2,270 nm is of particular interest, as it likely corresponds to C–H stretching and C=O combination bands, which are characteristic molecular vibrations associated with caffeine molecules [23, 24]. This peak serves as a potential marker for caffeine content, the primary focus of this study. The spectral region between 2,000 and 2,500 nm presents a complex landscape of overlapping bands, reflecting the presence of a diverse array of organic compounds within the coffee beans. These include proteins, lipids, and carbohydrates, each contributing to the overall spectral profile.

The intricate nature of these overlapping bands underscores the complexity of coffee bean composition and highlights the challenge in isolating specific components for analysis. While the raw spectrum provides valuable initial insights, it is important to note that this data is subject to various physical phenomena that can affect its interpretation. Scattering effects, caused by the physical structure of the coffee beans and variations in particle size, can introduce noise and baseline shifts in the spectrum [25]. These effects, if left uncorrected, could potentially mask or distort the chemical information of interest, particularly the subtle spectral features associated with caffeine content. Consequently, the raw spectral data, while informative, necessitate careful pretreatment to enhance their accuracy and reliability for quantitative analysis of specific compounds like caffeine.

3.2 Caffeine prediction using non-treated spectra

Preliminary results of the prediction of Arabica coffee caffeine content without pretreatment showed good results, with a correlation coefficient (r) of 0.95 and a coefficient of determination (R²) of 0.91, as shown in Figure 3.

Figure 3. Caffeine prediction using raw spectral data

This value indicates that the developed model has a very close correlation between the raw spectrum and the caffeine content measured using the HPLC method. The Root Mean Square Error of Calibration (RMSEC) value of 0.20238, which is lower than the standard deviation of the data, indicates that the model can provide fairly accurate predictions. However, while these results show that the raw spectrum data is good enough for caffeine prediction, it is important to remember that spectrum data often contains noise and distortion, which can reduce the accuracy of long-term predictions.

3.3 Caffeine prediction after standard normal variate spectra correction

The SNV pretreatment method demonstrated significant improvements in the prediction model's performance for caffeine content in Arabica coffee green beans. After applying SNV, the value of r increased to 0.96, and R² reached 0.92, indicating a stronger relationship between the spectral data and the actual caffeine content, as presented in Figure 4.

Figure 4. Caffeine prediction after standard normal variate (SNV) spectral data

SNV works by normalizing each spectrum individually, setting the mean to 0 and the standard deviation to 1. This normalization process is particularly effective in addressing variations caused by particle size differences and surface scattering effects, which are common challenges in NIRS analysis of solid samples like coffee beans.

The improvement in prediction accuracy after SNV pretreatment suggests that scattering effects were indeed masking some of the caffeine-related spectral information in the raw data. Coffee beans, being a heterogeneous material, can cause significant light scattering due to variations in particle size, shape, and surface characteristics [8, 20]. These scattering effects can introduce noise and baseline shifts in the spectra, potentially obscuring the subtle spectral features associated with caffeine content.

By reducing these scattering effects, SNV allows the PLSR model to focus more accurately on the spectral changes directly related to caffeine concentration. This enhanced focus results in a more robust and reliable prediction model. The higher R² value of 0.92 indicates that 92% of the variance in caffeine content can be explained by the spectral data after SNV pretreatment, a significant improvement over the raw spectral analysis.

It's worth noting that while SNV showed considerable improvement, it was not the best-performing pretreatment method in this study. The DT method achieved even better results with r = 0.99 and R² = 0.99. This suggests that while scattering effects were a significant source of spectral variation, other factors, such as baseline shifts, might also play a crucial role in the accuracy of caffeine prediction models for coffee beans. The effectiveness of SNV in this context underscores its value as a pretreatment method for NIRS analysis, particularly in applications involving solid agricultural products where particle size and surface scattering can significantly impact spectral quality.

3.4 Caffeine prediction after detrending spectra correction

DT pretreatment emerged as the most effective method for improving the prediction of caffeine content in Arabica coffee green beans using NIRS and PLSR. The results obtained with DT pretreatment were exceptional, with r = 0.99 and R² = 0.99, as presented in Figure 5, indicating an almost perfect linear relationship between the predicted and actual caffeine values.

The residual predictive deviation (RPD) value of 11.74 is particularly noteworthy. In NIRS analysis, RPD values above 3 are generally considered good for quantitative predictions, while values above 8 are excellent. The RPD of 11.74 achieved with DT pretreatment far exceeds these benchmarks, suggesting that the model is highly reliable and accurate for predicting caffeine content in unknown samples. DT works by removing linear or curvilinear trends in the spectral data [26-28]. These trends can arise from various sources unrelated to the chemical composition of interest, such as instrument drift over time, variations in sample thickness or density, differences in sample presentation or packing, and also environmental factors like temperature or humidity fluctuations.

Figure 5. Caffeine prediction after detrending (DT) spectral data

By eliminating these extraneous trends, DT allows the PLSR model to focus more precisely on the spectral features that are directly related to caffeine content. This is evident in the dramatically improved prediction accuracy compared to other pretreatment methods and the raw spectra. The effectiveness of DT in this study suggests that baseline variations were indeed a major source of error in the raw spectra of the coffee samples. These baseline shifts can mask or distort the subtle spectral changes associated with varying caffeine concentrations. By removing these variations, DT essentially levels the playing field for all spectra, making the caffeine-related features more prominent and easier for the PLSR model to interpret [29].

The superior performance of DT is further underscored by the RMSEC value of 0.06066, which is significantly lower than that obtained with other pretreatment methods. This low RMSEC indicates that the predictions made by the model have very small deviations from the actual caffeine values, further confirming the high accuracy and reliability of the DT-PLSR model.

The exceptional results achieved with DT pretreatment have significant implications for the coffee industry. They demonstrate that NIRS, when combined with appropriate data preprocessing techniques, can provide a rapid, non-destructive, and highly accurate method for determining caffeine content in Arabica coffee beans. This could potentially revolutionize quality control processes in coffee production, enabling real-time monitoring and adjustment of caffeine levels to meet specific standards or consumer preferences.

3.5 Caffeine prediction after extended multiplicative scatter correction spectra correction

The EMSC pretreatment method was applied to the NIRS data of Arabica coffee green beans to evaluate its effectiveness in improving the prediction accuracy of caffeine content. The results obtained from EMSC pretreatment showed a moderate improvement over the raw spectra, with r = 0.91 and R² = 0.91, as shown in Figure 6.

Figure 6. Caffeine prediction after extended multiplicative scatter correction (EMSC) spectral data

EMSC is a sophisticated pretreatment technique designed to separate chemical and physical information within the spectra. It corrects for both multiplicative and additive scatter effects, which are common issues in NIRS analysis. Multiplicative scatter effects arise from differences in sample thickness, particle size, and surface roughness, while additive scatter effects are due to baseline shifts caused by instrument drift or sample presentation variations. By addressing these scatter effects, EMSC aims to enhance the signal-to-noise ratio and improve the overall quality of the spectral data [30, 31].

The EMSC pretreatment resulted in an RPD value of 3.51, which indicates that the method is capable of providing good predictions. However, when compared to other pretreatment methods such as SNV and DT, EMSC's performance was somewhat lower. SNV achieved an RPD of 3.76, and DT yielded an exceptional RPD of 11.74, significantly outperforming EMSC.

The moderate improvement seen with EMSC suggests that while this method is effective in reducing some spectral variations, it may not be as well-suited for this particular dataset as simpler correction methods like SNV or DT. This discrepancy could indicate that the primary sources of spectral variation in these coffee samples are more effectively addressed by these simpler methods. For instance, the high performance of DT suggests that linear or curvilinear trends in the spectral data, possibly due to instrument drift or sample-to-sample variations, were a major source of error. SNV, on the other hand, effectively reduced scatter effects caused by particle size non-uniformity. In contrast, EMSC's more complex approach to correcting both multiplicative and additive scatter effects might not have been as targeted or efficient for the specific types of variations present in this dataset.

3.6 Caffeine prediction after peak normalization spectra correction

The PN demonstrated significant improvements in the prediction accuracy of caffeine content in Arabica coffee green beans, highlighting its utility in addressing specific types of spectral variations. With r = 0.96 and R² = 0.91, the PN pretreatment outperformed the raw spectra, indicating a stronger relationship between the normalized spectral data and the actual caffeine content, as presented in Figure 7.

Figure 7. Caffeine prediction after peak normalization (PN) spectral data

PN is a pretreatment technique designed to correct variations in the overall intensity of spectral peaks. This method is particularly useful when dealing with samples that exhibit differences in thickness or when there are variations in the instrument's response. By normalizing the intensity of the spectral peaks, PN ensures that the data is scaled to a uniform level, thereby minimizing the influence of these external factors on the spectral analysis.

The improvement seen with PN suggests that variations in peak intensity were indeed a significant factor affecting the model's ability to accurately quantify caffeine. In NIRS analysis, the intensity of spectral peaks can vary due to several reasons, including differences in sample thickness, packing density, and instrument calibration. These variations can introduce noise and distortions in the raw spectra, making it challenging for the PLSR model to accurately detect the subtle spectral changes associated with caffeine content.

By normalizing these peaks, the PN method allows the PLSR model to focus more precisely on the relative intensities of the caffeine-related absorption bands compared to other spectral features. This normalization process effectively reduces the impact of external factors, enabling the model to better capture the intrinsic chemical information related to caffeine. As a result, the model's ability to predict caffeine content is enhanced, as evidenced by the improved r and R² values [32, 33].

The effectiveness of PN in this context underscores its practical significance in NIRS applications, particularly in scenarios where sample presentation and instrument response can vary. In the coffee industry, where consistency in quality control is crucial, PN can be a valuable tool for ensuring that spectral data is reliable and consistent across different samples and measurement conditions.

Moreover, the PN method is relatively straightforward to implement and does not require complex computational resources, making it a practical choice for routine quality control processes. The RMSEC value of 0.20237, which is comparable to other pretreatment methods, further supports the efficacy of PN in providing accurate predictions with minimal error. While PN showed significant improvements, it is important to compare its performance with other pretreatment methods. DT and SNV pretreatments, for instance, achieved even higher R² values and RPD scores, indicating that they might be more effective in certain contexts.

However, PN's ability to correct for variations in peak intensity makes it a valuable addition to the arsenal of pretreatment techniques, especially when dealing with samples that exhibit notable differences in thickness or instrument response. The PN pretreatment method is a valuable tool for improving the accuracy of caffeine content predictions in Arabica coffee green beans using NIRS and PLSR. By normalizing the intensity of spectral peaks, PN effectively addresses variations that can affect the model's performance, allowing for more accurate and reliable predictions [34, 35]. This method, along with other pretreatment techniques, enhances the overall robustness and applicability of NIRS technology in the coffee industry, facilitating efficient and non-destructive quality control processes.

3.7 Caffeine prediction after median filter spectra correction

The MF pretreatment method was applied to the NIRS spectral data of Arabica coffee green beans to evaluate its effectiveness in improving the prediction accuracy of caffeine content. The results obtained from MF pretreatment showed a moderate improvement, with r = 0.95 and R² = 0.91. The RPD value of 3.51, as shown in Figure 8, indicates that this method is capable of providing good predictions, although it does not surpass the performance of other pretreatment methods like DT or SNV.

Figure 8. Caffeine prediction after median filter (MF) spectral data

MF is a spectral preprocessing technique designed to reduce noise in the spectra while preserving the sharp features that are crucial for accurate chemical analysis. This method works by replacing each data point in the spectrum with the median value of neighboring data points within a specified window. This approach is particularly effective at removing spike noise, which can arise from various sources such as instrumental errors or minor sample inconsistencies.

The moderate improvement seen with MF suggests that while random noise was not a major factor affecting the model's accuracy, some level of noise reduction was still beneficial. The similarity in performance to the raw spectra indicates that the primary sources of error in the model were not solely due to random noise but might be attributed to other factors such as baseline shifts, scattering effects, or sample-to-sample variations.

Despite this, the slight improvement achieved with MF highlights the importance of noise reduction in refining the prediction model. By smoothing out the spectra, MF helps to stabilize the data, making it more consistent and reliable for analysis. This is reflected in the RMSEC value of 0.20277, which, although not significantly lower than the raw spectra, still indicates a reduction in the error of calibration.

The use of MF in this context underscores its practical significance in NIRS applications where noise reduction is necessary but not the primary concern. In the coffee industry, where consistency and reliability are key, MF can be a useful tool for ensuring that spectral data is of high quality. This method is particularly advantageous because it is relatively simple to implement and does not require complex computational resources, making it accessible for routine quality control processes.

While MF showed moderate improvement, it is important to compare its performance with other pretreatment methods. DT and SNV pretreatments, for example, achieved higher R² values and RPD scores, indicating that they might be more effective in addressing the specific types of spectral variations present in this dataset. However, MF remains a valuable tool for reducing noise and preserving spectral features, especially in scenarios where other pretreatment methods may not be as effective [36].

The MF pretreatment method demonstrated its utility in improving the accuracy of caffeine content predictions in Arabica coffee green beans by reducing noise in the spectra. While it may not have been the most effective method in this study, its ability to preserve sharp spectral features and reduce spike noise makes it a valuable addition to the array of pretreatment techniques available for NIRS analysis. This method, combined with other pretreatment approaches, enhances the overall robustness and applicability of NIRS technology in the coffee industry, facilitating more accurate and efficient quality control processes.

3.8 Caffeine prediction after Savitzky-Golay filter spectra correction

The SG filter pretreatment method was employed to evaluate its effectiveness in improving the prediction accuracy of caffeine content in Arabica coffee green beans using NIRS and PLSR. The results obtained from SG filter pretreatment showed r = 0.95 and R² = 0.91, with an RPD value of 3.46, as shown in the scatter plot in Figure 9. These metrics indicate that the SG filter performed similarly to the MF method, suggesting that it is capable of providing fairly accurate predictions, although not significantly outperforming other pretreatment methods.

Figure 9. Caffeine prediction after Savitzky-Golay (SG) filter spectral data

SG filter is a sophisticated smoothing technique that aims to reduce noise in the spectral data while preserving the higher moments of the spectral peaks. This method is particularly useful for maintaining the shape and width of spectral features, which is crucial for accurate chemical analysis. By fitting a polynomial to a set of neighboring data points, the SG filter effectively smooths out high-frequency noise without distorting the underlying spectral information. This approach ensures that the intrinsic spectral features, including those related to caffeine content, remain intact and are not compromised by the smoothing process.

The results suggest that while SG filters were effective in reducing noise, they did not significantly improve the model's performance compared to other methods, such as DT or SNV. This could indicate that the spectral features related to caffeine content in the coffee beans are relatively broad and not significantly affected by high-frequency noise. Broad spectral features are less susceptible to the smoothing effects of the SG filter, as they are not as easily obscured by minor fluctuations in the data.

The performance of the SG filter, while comparable to MF, highlights the importance of selecting the most appropriate pretreatment method based on the specific characteristics of the dataset. DT emerged as the most effective method in this study, likely due to its ability to remove baseline variations that were masking the caffeine-related spectral information. The significant improvements seen with DT underscore the importance of addressing baseline shifts and other forms of spectral distortion in NIRS analysis.

The use of the SG filter in this context underscores its practical significance in NIRS applications where preserving spectral features is critical. In the coffee industry, where accurate and rapid analysis of chemical content is essential, the SG filter can be a valuable tool for ensuring that spectral data is of high quality. Although it may not be the most effective method for every dataset, the SG filter remains a useful technique for reducing noise while maintaining the integrity of the spectral signal.

In general, the various pretreatment methods demonstrated different levels of effectiveness in improving the PLSR model for caffeine prediction in Arabica coffee green beans. While the SG filter performed well in reducing noise and preserving spectral features, it did not significantly outperform other methods like DT or SNV [37-39]. The significant improvements seen with pretreatment underscore the importance of spectral correction in NIRS analysis of complex organic samples like coffee beans. By carefully selecting and applying the appropriate pretreatment method, researchers and industry professionals can enhance the accuracy and reliability of NIRS-based predictions, ultimately contributing to more efficient and effective quality control processes in the coffee industry. Performance comparison of various spectra correction approaches is presented in Table 1.

Table 1. Prediction performances of various spectra corrections using Partial Least Squares Regression (PLSR)

3.9 Statistical significance of prediction models

The reported coefficients r and R² provide valuable initial insights into model fit and explanatory power; however, these descriptive metrics alone do not establish the statistical reliability or significance of the developed calibration models. To rigorously validate that the observed relationships between the NIR spectral data and the reference caffeine measurements are not due to random chance, comprehensive statistical significance testing was performed on each PLSR model developed with the various pretreatment methods.

The statistical evaluation was conducted at two primary levels. First, an overall F-test was performed for each PLSR model to test the global null hypothesis that none of the latent variables have a significant linear relationship with the caffeine content, in other words, that all regression coefficients in the model are effectively zero. For every pretreatment method evaluated, the calculated F-statistic vastly exceeded the critical F-value at a stringent significance level of α = 0.01. The associated p-values were all less than 0.001. This decisive result allows for the rejection of the null hypothesis for every model, providing strong statistical evidence that each PLSR calibration explains a significant proportion of the variance in caffeine content. The probability that the high correlations r observed for models like DT (r = 0.99) and SNV (r = 0.96) arose by random coincidence is thus exceedingly low (p < 0.001).

Second, to ensure the models were robust and not overfit to noise in the calibration set, the significance of individual regression coefficients for the retained latent variables was examined. Analysis of the confidence intervals for these coefficients revealed that the key latent variables contributing to the prediction of caffeine were statistically significant (p < 0.01). This granular analysis confirms that the predictive structure captured by the models is based on meaningful spectral-caffeine relationships rather than spurious correlations with random variance. The exceptional results for the DT pretreatment are particularly noteworthy in this context. Its extraordinarily high F-statistic and the statistical significance of its model coefficients provide definitive mathematical support for its outstanding practical performance metrics (R² = 0.99, RPD = 11.74). This confluence of evidence confirms that the DT-PLSR model possesses not only exceptional predictive accuracy but also foundational statistical validity.

These formal significance tests affirm that the performance metrics reported throughout this study are grounded in statistically significant relationships. The findings confirm that NIRS-PLSR models, particularly when optimized with the DT pretreatment, produce reliable and significant calibrations for the non-destructive prediction of caffeine in intact Arabica coffee beans. This statistical rigor strengthens the conclusion that the method is a robust analytical tool suitable for further development and practical application.

3.10 Limitations and future work

This study, while demonstrating the significant potential of NIRS for caffeine prediction, is subject to limitations that must be transparently addressed to contextualize its findings and guide subsequent research. The most prominent limitation is the constrained sample size (n = 32) and its geographic scope. Although the beans were sourced from several distinct Indonesian regions (Gayo, West Java, South Sulawesi, Bali), providing valuable initial diversity, this set cannot encompass the full spectrum of genetic variability, agro-climatic conditions, soil compositions, post-harvest processing methods, and storage histories found across global Arabica coffee production.

Consequently, the high-performance calibration models developed here, particularly the exceptional model using DT pretreatment, are primarily validated for and most reliably applicable to bean populations similar in origin and character to those used in this calibration. Their predictive accuracy for beans from entirely different continents, novel cultivars, or unconventional processing methods remains untested and cannot be assumed.

To maximize model robustness within the available dataset, a rigorous internal validation protocol was implemented. The optimal complexity for each PLSR model was determined through an iterative process of leave-one-out cross-validation (LOOCV), which systematically holds out each individual sample for validation while calibrating the model on all remaining samples. This process minimizes the risk of overfitting and is reflected in the reported metrics, such as the RMSEC. However, it is critically important to distinguish internal cross-validation from external validation. LOOCV assesses performance on variations of the same dataset, whereas a true external validation tests the model on a completely independent set of samples, collected and measured separately.

The limited total number of samples in this study made the partitioning of a dedicated, statistically powerful external validation set impractical, as it would have severely weakened the stability and information content of the primary calibration model. Therefore, the outstanding metrics (R² = 0.99, RPD = 11.74 for DT) indicate a powerful explanatory model for this specific dataset, but its predictive power for new, unrelated populations requires further confirmation.

Given these considerations, the present work is most accurately positioned as a highly promising and methodologically rigorous proof-of-concept. It definitively establishes that a strong, statistically significant spectral-caffeine relationship exists and can be optimally accessed through DT pretreatment. To transition this finding into a universally robust, deployable analytical tool, focused future work is essential. The foremost requirement is the development of a vastly expanded and more diverse calibration library. This library should encompass hundreds of samples, representing major and minor Arabica growing regions worldwide, multiple harvest years to account for seasonal variation, different quality grades, and a range of processing methods. Following the creation of this comprehensive library, a rigorous validation protocol using a large, fully independent external test set is mandatory to establish true predictive accuracy and generalizability.

Further investigative avenues include exploring advanced machine learning algorithms like support vector regression, artificial neural networks, or ensemble methods that may capture non-linear relationships in the spectral data beyond the linear framework of PLSR. Research into model transferability, the ability of a calibration developed on one spectrometer to perform accurately on another instrument of the same or different model, is also crucial for practical industry adoption. Finally, integrating this NIRS method with rapid, online sorting systems represents a key engineering challenge for future implementation. Despite the need for these advancements, the core contribution of this study remains intact: it provides a statistically solid and exceptionally accurate foundation, identifying both the feasibility of the approach and the optimal pretreatment strategy, upon which a reliable, non-destructive quality control system for the global coffee industry can be built.