2026 Volume 19 Issue 2 Pages 96-104
Predicting sugar quality in sugarcane stalks using in-field measurements is difficult because the outer layer is covered with wax, water, and soil. This layer contains fibers and acts as a strong protective barrier with high thermal conductivity, hindering light transmission and masking signals related to the internal sugar content, increasing difficulties in measurement due to the difference in outer layers between sugarcane cultivars. Our research presents a solution to address the varying outer layer conditions of sugarcane cultivars by peeling away the outer layers to achieve uniformity. The findings indicate that this peeling method enhances the quality of the absorbance spectra and that focusing on the significant and reliable wavelength range of 800–950 nm further improves the model prediction accuracy.
Sugarcane is a globally important crop and is crucial to many economies. In Japan, the southern islands generate significant income for local sugarcane farmers from their 16 scattered raw sugar mills. These mills are vital to the livelihoods of many rural communities, and the economic stability of these regions is closely tied to the success of the sugarcane industry (Matsuoka, 2006).
Sugarcane quality, or sugar quality in sugarcane, known as polarimetric sucrose (Pol) and total soluble solids (Brix), is a key primary factor in price estimation for farmers and breeding programs. Pol and Brix are measured in sugar mills using a conventional polarimetric determination method. However, the extraction and preparation of juice for Pol measurement involve skilled, labor-intensive processes, including cutting, crushing, and clarification. Although Brix measurement can shorten the process, it includes sugars other than sucrose, such as fructose and glucose. Thus, new, rapid, accurate, and cost-effective methods are needed.
A previous study developed a measurement method for both on-site and laboratory settings using visible near-infrared (Vis-NIR) spectroscopy, offering a rapid assessment of sugarcane quality by measuring the absorption of NIR light in intact stalks and direct squeezing of sugarcane juice from a 10-cm cut of the stalk. It predicts sugar quality based on key wavelengths that are related to the reference sugar quality measured using conventional methods (Aparatana et al., 2022). This allows real-time monitoring and decision-making, potentially transforming management practices in sugarcane farming. However, the accuracy of the in-field prediction model using a portable Vis-NIR spectrometer remains inadequate because of known issues related to the instrument and sugarcane samples.
The portable Vis-NIR spectrometer used in this study operates in the short-wave NIR range (600–1,000 nm) and was designed to be cost-effective, which results in complicating the identification of the key wavelengths needed for precise predictions due to multiple information overlaps (Ozaki et al., 2021). The sugarcane rind, known as the outer layer of sugarcane, is reflective, fibrous, and mostly covered with a combination of wax, water, and soil residue. In addition, these combinations act as a strong protective barrier with high thermal conductivity. This thermal property varies among different sugarcane cultivars, potentially hindering light transmission and weakening the signals indicating internal sugar content. Consequently, the light source is crucial because some cultivars require adjustments in light intensity owing to their thicker diameters and sugarcane rind conditions (e.g., color, age, and wax). The issue of waxy materials can be mitigated by removing wax (Maraphum et al., 2018; Phuphaphud et al., 2019), but sugarcane rind issue remains a challenge. Developing a robust model that can be used for all cultivars is challenging because the sugarcane rind causes scattering and noise, complicating the identification of key wavelengths needed for a reliable model.
This study presents a method for addressing the issue of varying sugarcane rind conditions across sugarcane cultivars that involves removing the outer layers of sugarcane to reduce interference from different sugarcane rinds, wax, and impurities specific to each cultivar. Therefore, the aim of this study is to analyze the spectra of impurities and peeled sugarcane stalks in the late maturity phase using a custom portable Vis-NIR spectrometer, develop a sugar quality calibration model based on the analyzed information, and select an optimal model.
In this study, 48 sugarcane stalks aged 10–11 months were selected from sugarcane plants. Each stalk was subjected to 2 experiments, one at the bottom and the other at the top, resulting in 96 sugarcane billets. These plants were cultivated across 16 distinct fields on the remote island of Minamidaitōjima in Okinawa, Japan. The samples were collected during the 2023 harvesting season, specifically in February and March. The selected samples included six common and widely grown Japanese sugarcane varieties: Ni22, Ni28, Ni31, RK9714, RK1029, and Harunooki. The experimental and analytical procedures are briefly outlined in Fig. 1.
Spectral acquisitions were conducted in a sugarcane field using a custom portable Vis-NIR spectrometer (H-NIR-SC-01, HKN Engineering Co., Ltd., Japan), as shown in Fig. 2. The spectrometer was equipped with a built-in 25-W halogen lamp capable of measuring spectra up to 5,000 times and a silicon array detector. This spectrometer was used to scan the impurity surface (without removing wax and cleaning the surface, as shown in Fig. 3 (a)) and the undersurface after peeling, as shown in Fig. 3 (b), of each internode of the sugarcane stalk in the field.
Fig. 3 Photographs of sugarcane stalk and each stage after preparation
Temperature is commonly utilized to evaluate crop health. Measurement of stalk temperature can be a useful indicator for improving the accuracy of predictive models for sugar content, particularly when models are developed in field settings. Using the same tool for direct measurement may help reduce the time spent on data collection. To support this, a waterproof thermometer (SN3000, NETSUKEN Co. Ltd., Japan) was used during each spectral acquisition to measure temperature at the same scanned spot.
The spectral scanning and temperature measurement process involved the following steps. First, the impurity surface of each internode was scanned four times, twice at the top and twice at the bottom of the sugarcane stalk, for duplicates. The temperature was measured four times at each location. Next, the sample skin was peeled away (approximately 2 mm thick) using a Y-peeler and scanned, and the temperature was measured again, twice at the top and twice at the bottom, in duplicate. The temperature was measured four times at each location. These measurements were consistently conducted between 10 AM and 12 PM to reduce the reflection of light rays. The spectrometer detected diffuse reflected light, and the recorded data were automatically transferred via Bluetooth to a Panasonic Toughbook (FZ-T1, Panasonic Corp., Japan).
After the on-site measurements, the sugarcane was harvested with an axe, and only the scanned internode parts were cut into 10 cm billet pieces, as shown in Fig. 3 (c), then packaged in plastic zip-lock bags, and transported to the laboratory at the Faculty of Agriculture, University of the Ryukyus, Okinawa, Japan, for quality measurement. A total of 384 spectra and temperature values were collected in the field (48 stalks × 2 positions × 2 conditions × 2 consecutive), covering the 570 to 1031 nm wavelength range with 2 nm increments. The spectra were measured with an integration time of 200 ms. Subsequently, all consecutive measurements were averaged to obtain 96 data points each for the intact and peeled cane spectra, as well as for the intact and peeled cane stalk temperatures.
2.3. Acquisition of NIR spectra: benchtop NIR spectrometer and reference analysis of sugar qualityThe samples were kept temporarily in the refrigerator for 12 h and then transferred to an incubator set at a constant temperature of 20 °C for 18 h to ensure uniform conditioning before starting the quality measurement session. After incubation, each sample was carefully unpacked and cut into transverse sections. Each half was sectioned longitudinally twice, resulting in 8 smaller pieces of each original sample, as shown in Fig. 3 (d). The smaller pieces were then pressed using a hand press tool.
The 2 mL of the juice extracted from the pressed pieces were collected in a quartz slurry cup, which was subsequently covered with a gold reflector. Juice extraction and measurement processes were repeated twice for all samples to ensure data consistency and reliability. A benchtop NIR spectrometer (DS2500, FOSS A/S, Denmark) was utilized for measuring spectra of extracted juice. Distilled water was used to clean the measuring cup to reset the baseline for each sample and dried with tissue paper. The measured NIR spectra of each juice sample were used to predict the percentages of Pol and Brix using pre-established calibration models from previous studies. These calibration models achieved a coefficient of determination (R2) of 0.99 and a root mean square error of prediction (RMSEP) of 0.2 % for Brix and an R2 of 0.99 with an RMSEP of 0.3 % for Pol (Aparatana et al., 2022). Subsequently, all consecutive measurements were averaged, resulting in a total of 96 Brix and 96 Pol values.
2.4. Data processing and analysisMATLAB R2024b (ver. R2024b Update 3, MathWorks Inc., USA) with the PLS_Toolbox (ver. 9.5, Eigenvector Research Inc., USA) was used for data processing and analysis. All analyses were conducted using standard procedures in the toolbox.
2.4.1. NIR absorbance spectra pretreatmentThe raw spectra collected from the portable Vis-NIR spectrometer were transformed into absorbance spectra using Eq. (1).
| (1) |
Where Sa represents the absorbance spectrum, Sr is the raw spectrum, Sd is the dark reference spectrum obtained by covering the detector and turning off the light source, and Sw is the white reference spectrum obtained by scanning the Teflon white reference plate.
Subsequently, Savitzky–Golay (Savitzky et al., 1964) smoothing filters and differentiation techniques were applied to the spectra. In this study, only the second-order derivative (D2) was used. A previous study found that D2 exhibited the best outcome in correcting the baseline, enhancing key peaks, and maintaining the original peaks. First-order derivatives show similar results, but they are more difficult to identify the key peak, and standard normal variates only improve the result minimally, but are less effective than D2 (Aparatana et al., 2022). Thus, D2 was selected as the pretreatment in this study.
A window size of 21 data points with a second-order polynomial was selected for the intact cane absorbance spectra, and a window size of 19 data points with a second-order polynomial was selected for the peeled cane absorbance spectra after conducting trials in regression modelling.
2.4.2. Dataset preparationTo reduce instrument noise, the absorbance spectral wavelengths were trimmed at both ends, resulting in data analysis within the range of 600–1,000 nm. The spectral dataset included 192 raw spectra comprising 96 absorbance spectra from intact canes and 96 from peeled canes, which were used for principal component analysis (PCA). Data pre-treated with D2 were separated from each dataset using stratified random sampling based on monthly data. Specifically, 70 % of these data (42 from February and 25 from March) were allocated for calibration and cross-validation, while the remaining 30 % (18 from February and 11 from March) were reserved for validation in the development of multiple linear regression (MLR) and partial least squares regression (PLSR) models. Consequently, 18 MLR models were established (factorial: 3 wavelength sets, 600–1,000 nm, 800–1,000 nm, and 800–950 nm; 2 conditions, intact and peeled; and 3 references, Brix, Pol, and temperature [Temp.]). 18 PLSR models were established using the same parameters. The justification for the 3 selected wavelength regions is provided in the PCA and regression results sections.
2.4.3. Multivariate data analysisPCA is a powerful and versatile method that provides insight into complex multivariate data (Bro et al., 2014). This study used PCA to analyze the relationships between the spectra and samples, including intact and peeled cane varieties, various temperature ranges, and different months. The interpretation of principal component scores was based on the loadings of the principal components. Subsequently, a specific wavelength range was selected for a more detailed examination using an additional PCA model applied to the chosen spectral data.
For regression modeling, both MLR and PLSR were chosen because they are effective statistical tools widely used in NIR spectroscopy. MLR is more straightforward than PLSR because it directly models the linear relationship between the independent variables (each spectral wavelength) and the response variable. However, this approach may overlook important spectral information. In contrast, PLSR leverages the entire spectrum and employs pretreatment methods, allowing it to capture the most relevant spectral variations for accurate predictions and more effective spectral analysis, although with increased complexity relative to MLR.
This study employed the MLR with a step-up search method to select the wavelengths that exhibited the highest regression coefficients. All regression models utilized Venetian-blind cross-validation with 4-fold, each incorporating one blind to reduce local correlations and the risk of model overfitting. This 4-fold configuration was selected based on the structure of the calibration set, ensuring that each fold retained a sufficient and balanced number of samples from both sets while preserving stratification and supporting reliable model evaluation. Sugar quality values can be predicted using Eq. (2) for MLR and Eq. (3) for PLSR.
| (2) |
Where M is the predicted sugar quality value, b0 is the intercept regression coefficient, b is the regression coefficient, x represents the selected wavelengths, E is the model error, and n is the number of factors selected using the step-up method.
| (3) |
Where Y is the predicted sugar quality value, V is the sugarcane spectrum, K is the regression vector, and G is the residual.
The factors for MLR and the latent variables for PLSR were chosen based on the lowest root mean square errors of calibration (RMSEC) and cross-validation (RMSECV) before they began to increase. To assess model performance, we used the RMSEP, standard deviation of reference data from the validation sample, and bias to calculate the ratio of performance to deviation (RPD) and interpret the model performance alongside the R2. The William RPD statistical table for the forage, feed, and functionality parameters guided the interpretation of the model in this study. A classification score from 0 to 1.9 indicates poor performance and is not recommended; scores from 2 to 2.4 are considered poor but can be applicable for rough screening, and scores from 2.5 to 2.9 are considered fair for screening. (Williams, 2014).
The average spectra for intact and peeled sugarcane within the 600–1,000 nm range, collected in February and March (Fig. 4), show that the crucial wavelengths for sugarcane are approximately 680 nm, associated with chlorophyll or the outer skin, and near 970 nm, which represents the second overtone of the OH stretching mode representing water content (Williams et al., 2019). Both spectra displayed noise at both ends (approximately between 600–650 nm and 950–1,000 nm) owing to the limitations of the spectrometer, the rough skin surface of intact samples, and the irregular surfaces of peeled cane samples. The averaged spectra for peeled canes indicate lower absorbance than the intact spectra because the absence of the skin barrier facilitates more light penetration into the sugarcane's internal components. Additionally, the peeled cane spectra demonstrated a more distinct chlorophyll peak, as the removal of various pigments and wax from the skin made the water content peak more visible, indicating higher moisture levels within the cane compared with the intact skin. Notably, a small peak around 760 nm appeared specifically in the March spectra, which may suggest an increase in cellulose or fiber content (Williams et al., 2019), likely due to the beginning of sugarcane over-ripeness.
In total, 192 raw absorbance spectra, including 96 normal stalk spectra and 96 peeled spectra, were used to develop a PCA model with 4 principal components. The score plots are shown in the clustering plots of principal component 1 (PC1) and principal component 4 (PC4) in Fig. 5. PC1 represented the direction of maximum variance (89.23 %) in the dataset, capturing the most significant trend across all samples. In contrast, PC4, although accounting for only minimal variance (0.65 %), still offers a noteworthy source of variability that is orthogonal (uncorrelated) to PC1. Although PC2 (8.57 %) and PC3 (1.16 %) were part of the PCA, they contributed very little to the overall variance and did not yield significant insights into the objectives of our analysis.
Fig. 5 Score plots and loading plot of PCA (PC1 vs PC4) using raw data of sugarcane in the 600–1,000 nm wavelength range
In Fig. 5 (f), PC1 shows that the main pattern in the spectral data aligns with the characteristic absorbance of sugarcane. Meanwhile, despite its low explained variance, PC4 highlights a particular structure or signal of interest, focusing primarily on chlorophyll and fiber content between 680 and 800 nm (Gitelson et al., 1996; Williams et al., 2019), as well as sugar and water-related content from 800 to 1,000 nm, which corresponds to the locations of a second overtone of OH stretching and its combinations (Golic et al., 2003). The score plot in Fig. 5 (a) demonstrates that distinguishing varieties purely based on spectra is difficult because differences may be caused by numerous factors, including field weather conditions, nutrient levels, and varietal characteristics. However, Fig. 5 (b) shows that the intact and peeled cane spectra can be identified by PC1, which is interpreted as the absorbance signature of sugarcane. Monthly variations, as illustrated in Fig. 5 (c), are barely perceptible through PC4, which is related to chlorophyll, fiber, and water content. Fig. 5 (d) indicates similarity between the top and bottom sections. Fig. 5 (e) shows that PCA does not reveal a distinct spectral pattern linked to temperature, likely because of the narrow temperature range and the influence of the aforementioned factors, which can more significantly affect the spectra, resulting in weak or unclear correlations between temperature and sugarcane stalk spectra. Further classification does not alter the results and would only reduce clarity.
Building upon the previous PCA, we concentrated on sugar production from sugarcane, which mainly contains sugar and water. In this study, we particularly analyze the 800–1,000 nm region, acknowledging that the 700–800 nm range can also aid in identifying sugars. Still caution is required as the 680–800 nm region presents significant peaks in chlorophyll (Gitelson et al., 1996) and fiber (Williams et al., 2019) within sugarcane, overlapping with areas where both sugar and water are abundant. Consequently, we perform a second PCA again using the spectral data in the wavelength range of 800–1,000 nm. As shown in Fig. 6, the results were consistent with those of the initial PCA model. PC1 represents the direction of maximum variance (97.97 %) in the dataset, emphasizing the most significant trend across all samples. In contrast, although PC2 accounted for only a minor fraction of the variance (1.75 %), it remained an essential source since PC2 capturing unique information not captured by PC1.
Fig. 6 Score plots and loading plot of PCA (PC1 vs PC2) using raw data of sugarcane in the 800–1,000 nm wavelength range
The score plot presented in Fig. 6 (a) shows results consistent with the previous PCA analysis, indicating no notable differences among the sugarcane varieties. In Fig. 6 (b) and (f), PC1 indicates that the primary pattern in the spectral information aligns with the characteristic absorbance of sugarcane, which is similar to the findings from the first PCA model. Additionally, PC2, similar to PC4 from the earlier PCA, revealed a distinct structure associated with sugar and water content, excluding details of chlorophyll and fiber in the 680–800 nm range. This clarity facilitated differentiation based on sampling time, Fig. 6 (c). However, spectral variations due to plant health are influenced by field conditions, including weather, nutrient levels, varietal characteristics, and their interactions.
Consequently, these data present an opportunity for further exploration of field control conditions, aiming to devise a standardized vitality model for sugarcane that can assist in assessing their vitality levels. Fig. 6 (d) shows no difference between the spectra of the bottom and top parts of the sugarcane. Temperature is connected to water and OH stretching (Golic et al., 2003), leading to a discernible temperature distinction in the spectral data; however, the age of sugarcane and the previously mentioned factors still significantly impact the spectra, resulting in less clear interpretability, Fig. 6 (e).
3.3. Results of regression modelThe predicted % Brix, % Pol, and Temp. values for the sample set are listed in Table 1. The results of the MLR and PLSR models developed for Brix, Pol, and Temp., based on the 3 selected wavelength regions, 2 different conditions, and 3 reference values are presented in Tables 2 and 3.
| Component | Indicators | n | Average | Minimum | Maximum | StDev |
|---|---|---|---|---|---|---|
| Brix (% Brix) | Calibration set | 67 | 20.2 | 5.9 | 24.6 | 3.5 |
| Validation set | 29 | 20.6 | 11.6 | 23.9 | 2.8 | |
| Pol (% Pol) | Calibration set | 67 | 17.8 | 3.9 | 23.1 | 3.6 |
| Validation set | 29 | 18.2 | 9.8 | 21.9 | 2.9 | |
| Skin Temp. (°C) | Calibration set | 67 | 22.5 | 20.9 | 25.5 | 1.3 |
| Validation set | 29 | 22.5 | 20.9 | 25.1 | 1.2 | |
| Peeled Temp. (°C) | Calibration set | 67 | 22.2 | 20.5 | 25.8 | 1.1 |
| Validation set | 29 | 22.0 | 20.8 | 24.2 | 1.0 |
n: number of samples, StDev: standard deviation, Temp.: temperature.
Regarding the 3 selected wavelength regions, the first comprises the full wavelength range, and the second is the region chosen after PCA interpretation, which contains the most reliable information regarding sugar and water. The last set was derived after applying the second order derivative D2, as detailed in the absorbance spectra section. The beginning and end of the absorbance spectra were noisy for both cane conditions. It is recognized that applying D2 typically enhances the key peaks while increasing the noise, which justifies the selection of this wavelength range for the final set.
3.3.1. MLR model resultsWavelength selection has no effect on the intact cane, yielding the highest predictive coefficient of determination (Rp2) values of 0.63, 0.58, and 0.65 for Brix, Pol, and Temp., respectively, along with RMSEP values of 1.7 %, 1.8 %, and 0.7 °C for the same categories (Table 2).
Conversely, for peeled cane, eliminating unreliable information—except in the temperature models—leads to improved accuracy within the 800–1,000 nm range, especially between 800–950 nm, achieving Rp2 values of 0.70 and 0.73 for Brix and Pol and an RMSEP of 1.5 %. The temperature model for the peeled cane did not yield superior results, likely because of an insufficient temperature gap, and the key temperature wavelengths in the peeled cane spectra may have overlapped with other data. Regarding intact canes, because sugarcane skin has a high thermal conductivity, further investigation could enable the development of a temperature-prediction method for plant energy assessment.
All the models for intact samples exhibited RPD values ranging from 1.6 to 1.7. According to Williams’ criteria, these models are unsuitable, as they fall short of the minimum requirement of 1.9 and need enhancement. In contrast, the best model for the peeled sample operates within the 800–950 nm wavelength range, with Rp2 values of 0.70, 0.73, and 0.49 for Brix, Pol, and Temp., respectively, and RMSEP values of 1.5 %, 1.5 %, and 0.7 °C for Brix, Pol, and Temp., respectively, along with RPD values of 1.9, 2.0, and 1.5 for Brix, Pol, and Temp.. Although the Brix and Pol models slightly exceeded Williams’ rough screening requirements, the temperature model did not meet the expected standards, as previously indicated.
| Model | Selected wavelength (nm) | Rc2 | RMSEC (%), (°C) | Rcv2 | RMSECV (%), (°C) | Rp2 | RMSEP (%), (°C) | Bias (%), (°C) | RPD |
|---|---|---|---|---|---|---|---|---|---|
| Intact | |||||||||
| 600–1,000 nm | |||||||||
| Brix | 904, 866, 914, 800, 888 | 0.70 | 1.9 | 0.66 | 2.0 | 0.63 | 1.7 | 0.10 | 1.7 |
| Pol | 904, 866, 914 | 0.63 | 2.1 | 0.55 | 2.4 | 0.58 | 1.8 | 0.10 | 1.6 |
| Temp. | 778 | 0.68 | 0.7 | 0.66 | 0.7 | 0.65 | 0.7 | −0.03 | 1.7 |
| 800–1,000 nm | |||||||||
| Brix | 904, 866, 914, 800, 888 | 0.70 | 1.9 | 0.66 | 2.0 | 0.63 | 1.7 | 0.10 | 1.7 |
| Pol | 904, 866, 914 | 0.63 | 2.1 | 0.55 | 2.4 | 0.58 | 1.8 | −0.15 | 1.6 |
| Temp. | 800 | 0.66 | 0.8 | 0.64 | 0.8 | 0.56 | 0.8 | −0.17 | 1.6 |
| 800–950 nm | |||||||||
| Brix | 904, 866, 914, 800, 888 | 0.70 | 1.9 | 0.66 | 2.0 | 0.63 | 1.7 | 0.10 | 1.7 |
| Pol | 904, 866, 914 | 0.63 | 2.1 | 0.55 | 2.4 | 0.58 | 1.8 | −0.15 | 1.6 |
| Temp. | 800 | 0.66 | 0.8 | 0.64 | 0.8 | 0.56 | 0.8 | −0.17 | 1.6 |
| Peeled | |||||||||
| 600–1,000 nm | |||||||||
| Brix | 910, 850, 862, 928 | 0.82 | 1.5 | 0.80 | 1.6 | 0.70 | 1.5 | 0.23 | 1.9 |
| Pol | 908, 846, 746, 862, 744 | 0.83 | 1.5 | 0.81 | 1.6 | 0.74 | 1.4 | −0.22 | 2.0 |
| Temp. | 868 | 0.47 | 0.8 | 0.44 | 0.8 | 0.49 | 0.7 | 0.12 | 1.5 |
| 800–1,000 nm | |||||||||
| Brix | 910, 850, 862, 928 | 0.82 | 1.5 | 0.80 | 1.6 | 0.70 | 1.5 | 0.23 | 1.9 |
| Pol | 908, 846, 802, 886, 866 | 0.83 | 1.5 | 0.79 | 1.7 | 0.73 | 1.5 | −0.27 | 2.0 |
| Temp. | 868 | 0.47 | 0.8 | 0.44 | 0.8 | 0.49 | 0.7 | 0.12 | 1.5 |
| 800–950 nm | |||||||||
| Brix | 910, 850, 862, 928 | 0.82 | 1.5 | 0.80 | 1.6 | 0.70 | 1.5 | 0.23 | 1.9 |
| Pol | 908, 846, 802, 886, 866 | 0.83 | 1.5 | 0.79 | 1.7 | 0.73 | 1.5 | −0.27 | 2.0 |
| Temp. | 868 | 0.47 | 0.8 | 0.44 | 0.8 | 0.49 | 0.7 | 0.12 | 1.5 |
R2: coefficient of determination, RMSEC: root mean square error of calibration, RMSECV: root mean square error of cross-validation, RMSEP: root mean square error of prediction, RPD: ratio of performance to deviation.
Table 3 presents the outcomes of the PLSR model. For intact samples, these models yield Rp2 values ranging from 0.15 to 0.42 for Brix, 0.17 to 0.65 for Pol, and 0.34 to 0.68 for Temp., alongside RMSEP values of 2.2 % to 2.6 %, 1.7 % to 2.7 %, and 0.7 °C to 1.0 °C for these parameters. This suggests that wavelength selection affects intact canes, particularly between 800 and 950 nm. Similar to the peeled cane, the results indicate further improvements, except for the temperature models. Specifically, the 800–950 model reflects the Rp2 values of 0.79 for Brix and 0.78 for Pol, with RMSEP values of 1.3 % and 1.4 %, respectively. The temperature model for the peeled cane did not show better performance owing to the same issues previously discussed.
The optimal model for intact cane within wavelengths of 800 to 950 nm, achieving Rp2 values of 0.40 for Brix, 0.65 for Pol, and 0.68 for Temp., along with RMSEP values of 2.2 %, 1.7 %, and 0.7 °C, respectively. All models reported RPD values that did not exceed 1.9. According to Williams’ criteria, these models are considered inadequate and require further enhancement.
In contrast, the optimal model for the peeled sample is also within wavelengths of 800 to 950 nm, displaying Rp2 values of 0.79 for Brix, 0.78 for Pol, and 0.33 for Temp., with corresponding RMSEP values of 1.3 %, 1.4 %, and 0.8 °C for Brix, Pol, and Temp., respectively. These values are within acceptable limits. Furthermore, the RPD values are 2.2 for Brix, 2.2 for Pol, and 1.2 for Temp.; the values for Brix and Pol, exceeding Williams’ rough screening criteria, fall within acceptable limits. However, the temperature model failed to meet the expected standards, as previously stated.
| Model | LVs | Rc2 | RMSEC (%), (°C) | Rcv2 | RMSECV (%), (°C) | Rp2 | RMSEP (%), (°C) | Bias (%), (°C) | RPD |
|---|---|---|---|---|---|---|---|---|---|
| Intact | |||||||||
| 600–1,000 nm | |||||||||
| Brix | 3 | 0.19 | 3.1 | 0.12 | 3.3 | 0.15 | 2.6 | −0.13 | 1.1 |
| Pol | 3 | 0.28 | 3.0 | 0.20 | 3.2 | 0.17 | 2.7 | −0.24 | 1.1 |
| Temp. | 4 | 0.69 | 0.7 | 0.62 | 0.8 | 0.34 | 1.0 | −0.12 | 1.2 |
| 800–1,000 nm | |||||||||
| Brix | 6 | 0.74 | 1.8 | 0.53 | 2.5 | 0.42 | 2.2 | −0.09 | 1.3 |
| Pol | 8 | 0.80 | 1.6 | 0.59 | 2.3 | 0.51 | 2.1 | −0.07 | 1.4 |
| Temp. | 4 | 0.73 | 0.7 | 0.61 | 0.8 | 0.64 | 0.8 | −0.05 | 1.7 |
| 800–950 nm | |||||||||
| Brix | 5 | 0.69 | 1.9 | 0.64 | 2.1 | 0.40 | 2.2 | −0.34 | 1.3 |
| Pol | 5 | 0.69 | 2.0 | 0.66 | 2.1 | 0.65 | 1.7 | −0.29 | 1.8 |
| Temp. | 4 | 0.72 | 0.7 | 0.65 | 0.8 | 0.68 | 0.7 | −0.13 | 1.7 |
| Peeled | |||||||||
| 600–1,000 nm | |||||||||
| Brix | 7 | 0.57 | 2.3 | 0.37 | 2.8 | 0.16 | 2.9 | −0.18 | 1.0 |
| Pol | 3 | 0.24 | 3.1 | 0.11 | 3.4 | 0.26 | 2.4 | −0.32 | 1.2 |
| Temp. | 2 | 0.26 | 0.93 | 0.13 | 1.0 | 0.27 | 0.8 | 0.16 | 1.2 |
| 800–1,000 nm | |||||||||
| Brix | 6 | 0.78 | 1.6 | 0.67 | 2.0 | 0.59 | 2.0 | 0.17 | 1.4 |
| Pol | 9 | 0.87 | 1.3 | 0.73 | 1.9 | 0.74 | 1.6 | −0.21 | 1.8 |
| Temp. | 2 | 0.28 | 0.9 | 0.25 | 0.9 | 0.24 | 0.8 | 0.17 | 1.2 |
| 800–950 nm | |||||||||
| Brix | 6 | 0.83 | 1.4 | 0.78 | 1.6 | 0.79 | 1.3 | −0.17 | 2.2 |
| Pol | 6 | 0.84 | 1.4 | 0.78 | 1.7 | 0.78 | 1.4 | −0.28 | 2.2 |
| Temp. | 2 | 0.30 | 0.9 | 0.28 | 0.9 | 0.33 | 0.8 | 0.14 | 1.2 |
LV: latent variable.
This research was conducted during the late maturity phase of Japanese sugarcane to examine sugarcane stalks from various fields on the island and utilized a custom portable Vis-NIR spectrometer with a wavelength range of 600–1,000 nm. The findings show that applying the peeling technique enhances the quality of the absorbance spectra, while identifying a significant and reliable wavelength range of 800–950 nm further improves model prediction performance, allowing for precise measurements of sugarcane sugar content in the field. This procedure enables the use of new commercial portable Vis-NIR spectrometers with fewer light sources to develop calibration models. Building on these benefits, the study presents a non-destructive method for rapid, on-site sugar quality assessment, allowing breeders to estimate sugar quality instantly in the field and enabling farmers to judge their crop maturity and detect anomalies. Further research could focus on combining portable Vis-NIR spectrometers with locally adapted predictive models and user training, integrated with drone or satellite imagery to enable large-scale, spatially resolved sugar quality analysis, advancing precision agriculture and informed crop management.
The study was funded by an ‘On-farm Demonstration Trials of Smart Agriculture’ Grant (H06) administered by Ministry of Agriculture, Forestry and Fisheries of Japan.
The authors declare no conflicts of interest.
The supplementary data, Tables and Figures in this article are published in J-STAGE Data.
