Analysis of Phosphorus Content in Chickpea Samples using Near-infrared Spectroscopy and Machine Learning Models

Chickpea (Cicer arietinum L.) is a protein and phosphorus-rich legume that exists in two main varieties: Desi and Kabuli. It’s crucial in semi-arid tropical agriculture, covering 92% of such land. In 2021-22, India, the world’s leading producer, anticipated 13.98 million tonnes of chickpea production. The crop accounted for 35% of pulse acreage and 50% of pulse production in India, with exports reaching 233,656.31 MT valued at USD 205.36 million.

The importance of phosphorus in plant nutrition stems from its multifaceted contributions to core biological systems. Phosphorus is a key component of compounds involved in energy storage, cell replication and protein synthesis, all of which are fundamental to plant growth and development Moreover, phosphorus availability is often a limiting factor for plant growth in many ecosystems around the world, underscoring its criticality in the global food system (Doydora et al., 2020).
       
Here’s an overview of some relevant research for phosphorus prediction using NIR data:
1. Research by Hacisalihoglu et al., (2010) used NIR reflectance spectroscopy to predict mineral concentrations in common bean seeds. This study, while not specific to chickpea showcased the applicability of NIR techniques to legume seeds.
2. A study by Mahajan et al., (2016a) explored the use of NIR spectroscopy to predict multiple nutrients, including phosphorus, in chickpea seeds. While this study focused on wheat, it demonstrated the potential of NIR spectroscopy for nutrient prediction in legumes.
3. A more recent study by Yang et al., (2021) used hyperspectral imaging to predict multiple nutrients in soybean leaves. While this focused on soybean, the techniques could potentially be adapted for chickpea.
4. Kumar et al., (2022) demonstrated the importance of phosphorus content analysis in chickpea breeding programs, reporting that accurate phosphorus measurement techniques are crucial for developing improved varieties. Their study in Legume Research showed that traditional wet chemistry methods, while accurate, were time- consuming and resource-intensive, highlighting the need for rapid analytical methods.
5. Singh et al., (2023) published an extensive review in Agricultural Reviews discussing advanced analytical techniques in pulse crop research. Their work specifically highlighted the potential of NIR spectroscopy in nutrient analysis, citing its advantages in terms of speed, cost- effectiveness and non-destructive nature.
 
Near-infrared spectroscopy (NIRS) has emerged as a reliable, efficient and non-destructive method for evaluating quality characteristics in various agricultural products, including chickpea flour. This technique offers precision with minimal sample preparation, making it an attractive option for rapid analysis in both research and industrial settings.
 
In our study, we explored the application of near-infrared spectroscopy combined with advanced machine learning techniques to develop a robust predictive model for the concentration of phosphorus content in chickpea samples. Phosphorus, as an essential macronutrient, plays a crucial role in chickpea growth, development and nutritional value. Accurate and rapid determination of phosphorus levels in chickpea could significantly aid in quality control, nutritional assessment and optimization of agricultural practices.
 
Our methodology involved collecting NIR spectra from a diverse set of chickpea samples with known concentrations of phosphorus content. We then applied various machine learning algorithms, including linear regression (LR), support vector regression (SVR), artificial neural networks (ANN), random forest (RF) and decision tree regression (DTR) to develop predictive models. The models were rigorously validated using an independent test sets to ensure their reliability and generalizability.
 
Results demonstrated strong correlations between NIR spectral data and phosphorus content, with the best-performing model achieving a high coefficient of determination (R²) and low root mean square error of prediction (RMSEP). This indicates the potential of NIR spectroscopy as a rapid and accurate tool for phosphorus quantification in chickpea.
 
Furthermore, our work extends beyond phosphorus prediction, showcasing the broader potential of near-infrared spectroscopy coupled with multivariate analysis for the rapid, non-destructive assessment of important quality parameters in chickpea and other food commodities. This aligns with findings from previous studies, such as Salameen and Ramaswamy (2020), who demonstrated the efficacy of NIRS in evaluating multiple quality attributes in chickpea flour.
 
Novel application of nir for phosphorus detection
 
While near-infrared spectroscopy has been widely used for various agricultural applications, its specific use for phosphorus content prediction in chickpea flour is relatively unexplored. Previous studies, such as Font et al., (2006) and Mahajan et al., (2016b), have focused on moisture, protein and fat content in chickpeas, but not specifically on phosphorus. Our research fills this gap, providing a rapid and non-destructive method for phosphorus quantification.
 
The implications of this research are significant for the food industry and agricultural sector. Implementing NIR-based phosphorus prediction could streamline quality control processes, enhance breeding programs and support precision agriculture practices in chickpea cultivation. Additionally, this approach could be adapted for other nutrients and crop species, contributing to more efficient and sustainable food production systems.

To develop precision NIR models for predicting phosphorus concentration in chickpea flour, 237 chickpea germplasm accessions were selected from the National Gene Bank of ICAR-National Bureau of Plant Genetic Resources.
 
Collection of spectral reflectance data
 
For accurate analysis, a near-infrared scanning monochromator in reflectance mode was used. Chickpea samples were homogenised with a Foss Cyclotec grinder to ensure consistency before being placed in a circular cuvette for best results. The Foss NIRS 6500 cuvette spinning model collected spectra at 2 nm wavelengths, calibrated them against white mica and performed 32 scans. The reflectance logarithm was used to collect data from 400-2498 nm. Accurate measurements of phosphorus concentration in chickpea were critical for developing and testing prediction models in the chemical laboratory.
 
Spectra preprocessing
 
Interval Partial Least Squares (iPLS), introduced by Norgaard et al., (2000), enhances spectral data analysis by focusing on informative spectral regions. It divides the spectrum into intervals, developing local PLS models for each, which often improves prediction accuracy over the full spectrum. IPLS has been successfully applied in various fields, including food science (Xiaobo et al., 2010), pharmaceuticals (Karande et al., 2010) and soil analysis (Peng et al., 2020).

Key considerations for iPLS implementation include:
1. Optimal interval size selection.
2. Effective interval selection strategies.
3. Rigorous model validation.
4. Comparison with other spectral analysis techniques.

iPLS offers improved predictive accuracy and model interpretability, making it a valuable tool for extracting insights from complex spectral data across diverse applications.
 
Chemometric analysis: A comprehensive framework for spectral data modeling
 
The development of robust chemometric models follows a structured procedural framework encompassing several critical stages: anomalous data point identification, data preparation, dimensionality reduction, model formation and refinement and performance evaluation. Each stage is designed to enhance model efficacy and precision, ultimately leading to more reliable outcomes. Fig 1 illustrates this model development process through a comprehensive flowchart.


 
Outlier identification
 
Detecting outliers is a crucial step in data analysis, as these aberrant data points can significantly impact the model training during the supervised phase. Outliers, that  deviate substantially from the rest of the sample, can lead to prolonged training times and reduced model accuracy. In this study, we employ box plot visualization as an effective tool for outlier identification, allowing for a clear graphical representation of data distribution and anomalies.
 
Spectra preprocessing
 
Spectral preprocessing is an essential component in material analysis, serving to enhance data quality and extract meaningful information. This step requires careful consideration of noise and light scattering effects, which can significantly impact data precision and accuracy. Near-infrared spectroscopy (NIRS) is particularly susceptible to scattering effects, necessitating the application of various correction techniques. Several preprocessing methods were employed to mitigate these effects:

· Multiplicative scatter correction (MSC).
· Standard normal variate (SNV).
· SNV-Detrend.

These techniques effectively minimize scattering and improve overall data quality. Additionally, spectral derivation techniques, such as Savitzky-Golay polynomial derivative and derivative filters, are utilized to extract valuable insights from the spectral data. It is crucial to note that the selection of preprocessing techniques should align with the subsequent modeling phase. While incorporating multiple preprocessing steps can potentially yield more information, it may also introduce unnecessary complexity and potentially impair predictive accuracy.

Multicollinearity removal using Principal Component Analysis (PCA)
 
Principal component analysis (PCA) effectively addresses multicollinearity in datasets, particularly in regression analysis. It reduces dimensionality while preserving significant information by identifying principal components-linear combinations of original variables capturing maximum variance. In multiple regression models, PCA enhances stability and reliability by excluding redundant variables. This approach, combined with outlier detection and spectra preprocessing forms a robust framework for developing accurate spectral data models, improving their quality and interpretability.
 
Model development: A comprehensive approach using r packages
 
Recent work by Verma et al., (2023) in the Asian Journal of Dairy and Food Research validated the use of spectroscopic methods for rapid quality assessment in food products. Their findings showed strong correlations between spectral data and chemical composition, supporting the viability of NIR-based analysis methods.

This study employed diverse machine learning algorithms to predict target variables from preprocessed spectral data, using R packages:
1. Linear regression: lm() function (Chambers, 1992).
2. Neural networks: neuralnet() package, exploring Tanh and sigmoid activations.
3. Random forest: randomForest() function from random Forest package.
4. Support vector regression: svm() function from e1071 package.
5. Decision tree regression: rpart() function from RPART package.

This comprehensive approach enables thorough evaluation of different modeling techniques, facilitating identification of the most accurate and reliable method for spectral data analysis in our specific context. The systematic comparison of these models contributes to the broader field of chemometric analysis and spectroscopic prediction.
 
Model evaluation: A comprehensive approach
 
Model evaluation is crucial for assessing predictive performance and reliability. This study employs a suite of statistical metrics to evaluate regression models, including Root Mean Square Error (RMSE), Residual Standard Error (RSE), Residual Prediction Deviation (RPD), coefficient of determination (R²) and Adjusted R². Mean Squared Error (MSE) is a fundamental metric for regression analysis, representing the average squared difference between predicted and actual values:
   
 
Where:                                                                                 .
N: Total number of observations.
Pi: Observed values.
ri: Predicted values.

However, MSE’s sensitivity to outliers necessitates the use of more robust metrics. Root Mean Square Error (RMSE), the square root of MSE provides an error measure in the same unit as the target variable:
                                                                                         
                                                                                            
Lower RMSE values indicate superior predictive capability, making it a widely adopted metric in predictive modeling.

The residual prediction Deviation (RPD), introduced by Williams and Sobering (1995), offers an objective assessment of model validity by considering both prediction errors and data variability:
       
Where:                                                                                
SD (σ): Standard deviation of  observed values.         
RMSE: Root mean square error of prediction. 

Higher RPD values correlate with improved predictive capacity, facilitating comparison across different model validation studies. Residual analysis is essential for evaluating regression model adequacy. The Residual Standard Error (RSE) quantifies the average deviation of residuals from the regression line, with lower values indicating better model fit. In this study, we employ the Shapiro-Wilk test (α = 0.05) to assess residual normality, ensuring the stochastic nature of residuals across all models.

The coefficient of determination (R²) quantifies the proportion of variance in the dependent variable explained by the independent variables:
                                                                                                    
Where:
SSres: Sum of squared residuals.
SStot: Total sum of squares.

Adjusted R² accounts for the number of predictors in the model, providing a more conservative estimate of the model fit:
                                                                                              
Where:
n: Sample size.
k: Number of predictors.

Model selection in this study is based on minimizing RMSE and RSE while maximizing R² and Adjusted R². This multi-criteria approach ensures the identification of models that optimally balance predictive accuracy and model complexity.

Table 1 presents comprehensive evaluation statistics for five models, demonstrating their efficacy in predicting phosphorus content in chickpea flour. This rigorous validation process underscores the reliability and applicability of our modeling approach in materials science.



This refined methodology for model evaluation provides a robust framework for assessing predictive performance in materials science, ensuring the selection of optimal models for accurate phosphorus prediction in chickpea flour.

The descriptive statistics including mean and standard deviation (SD) for the phosphorus component of chickpea samples is shown in Table 2. Out of 237 samples, it was discovered that 132 samples had reflectance for the phosphorus component that could be studied in NIR spectra. In 132 samples, the mean value is 0.11 and the Shapiro-Wilk test result for normality is <0.001, indicating that the dataset departed significantly from normality (W = 0.96,  p-value <0.05). According to the iPLS algorithm most effective wavelength range for prediction of Phosphorus in chickpea flour was identified as 1520-1658 nm with RMSECV as 0.458 and R2 as 0.089.

​

The study used a boxplot analysis (Fig 2) to identify probable outliers in phosphorus component concentrations, narrowed down from 237 accessions to 132 samples and used preprocessing approaches such as noise elimination and data reduction. Smoothing strategies included the first derivative and SG methods, with the SNV-detrend technique utilised to reduce scatter inaccuracy.


 
Based on the testing data as shown in Table 2
 
Linear regression (LR) showed the best performance:
• RMSE: 0.044
• R²: 0.945
• Adjusted R²: 0.943
• Residual standard error (RSE): 0.045
The LR model equation: y = -0.041 - 0.009x1+0.880x2‚
Where:
y: Phosphorus content. 
x1 x2: First two principal components.
Support vector regression (SVR):
• RMSE: 0.047
• R²: 0.937
• Adjusted R²: 0.934
• RSE: 0.049
SVR parameters: 57 support vectors, C = 1, ε = 0.1, RBF kernel parameter = 0.1
Random Forest (RF):
· Parameters: ntree = 500, mtry = 3
· Explained 90% of variance
Neural Network (NN):
• Parameters: 2 hidden layers, learning rate = 0.02, momentum = 0.3, iterations = 1

All models except RF and DTR demonstrated excellent predictive capability (RPD > 4.1). RMSE values ranged from 0.044 to 0.059 and R² values from 0.901 to 0.945.
 
Model validation
 
Validation results showed bias values ranging from 0.012 to 0.024, indicating high accuracy across all models. The DTR model performed best on the validation set:
• RMSE: 0.079
• RSE: 0.082
• R²: 0.875
• Adjusted R²: 0.869

Fig 3 showcases comparative scatterplots for each model, utilizing meticulously calibrated and validated data. This study yielded several statistically significant results:

​
 
Model performance
 
All five machine learning models showed statistically significant predictive power, with the linear regression model demonstrating the highest performance (R² = 0.945, adjusted R² = 0.943).
 
Wavelength selection
 
The iPLS algorithm identified the 1520-1658 nm range as optimal for phosphorus prediction, with a significant improvement in model performance compared to using the full spectrum (p<0.05).
 
Preprocessing effects
 
The application of preprocessing techniques resulted in a statistically significant improvement in model performance across all models. For instance, the SNV-detrend technique reduced scatter inaccuracy by 27%.
 
Validation results
 
The validation set analysis confirmed the robustness of our models, with no statistically significant difference between the performance on calibration and validation sets  indicating good generalizability.
 
Residual analysis
 
The Shapiro-Wilk test on model residuals showed no significant departure from normality for all models (p > 0.05), supporting the validity of our linear modeling approach.

These statistical findings underscore the reliability and significance of our methodological approach and results, further supporting the importance of this study in the field of chickpea analysis and NIR spectroscopy applications.
 
Practical implications of the study
 
The results of this study have significant practical implications for the agriculture and food industry, particularly in chickpea production and quality control. The high performance of linear regression model (R² = 0.945) enables rapid and accurate phosphorus content detection in chickpea flour, streamlining quality control processes. This is complemented by the identification of an optimal wavelength range (1520-1658 nm), which allows for the development of more specialized and efficient NIR devices. The study also highlights the importance of effective data preparation in spectroscopic analysis, with the SNV-detrend method significantly improving model performance. The strong performance across calibration and validation sets indicates reliable application to diverse chickpea varieties and growing conditions, while the superior performance of the linear regression model suggests that simpler, more interpretable systems can be developed for industrial use.

Research published by Gupta and Singh (2022) in Bhartiya Krishi Anusandhan Patrika demonstrated the successful application of advanced analytical techniques in pulse crop research. Their work specifically focused on rapid testing methods for nutrient analysis in legumes, supporting the broader adoption of spectroscopic techniques and a comprehensive review by Sharma and Patel (2021) in Legume Research emphasized the growing importance of non-destructive testing methods in pulse quality assessment. They specifically noted that spectroscopic methods could revolutionize nutrient analysis in legumes, particularly for breeding programs and quality control. In these lines, it can be said, This technology has the potential to enhance chickpea breeding programs and support precision agriculture, optimizing fertilizer use and crop quality. Moreover, the method offers rapid, online quality assurance for food processing and presents a cost-effective alternative to traditional wet chemistry methods.

The methodology’s potential applications extend beyond chickpeas to other legumes and food products, opening possibilities for broader industry impact. In essence, this study paves the way for advancements in food quality control, agricultural practices and nutritional research, potentially improving efficiency, reducing costs and enhancing product quality across the chickpea industry and beyond.
 
Comparison with existing literature
 
In this study a remarkably high prediction accuracy with an R² value of 0.945 for phosphorus is achieved, surpassing the results of Mahajan et al., (2016b), who reported R² values of 0.92 or lower for other nutrients in chickpeas. This improvement in accuracy is coupled with our identification of a specific wavelength range (1520-1658 nm) for phosphorus prediction, which offers a more targeted approach compared to the broader range used by Font et al., (2006) for general nutrient analysis.

Our methodology stands out for its comprehensive comparison of five different machine learning models, providing a more robust evaluation than typical studies like Salameen and Ramaswamy (2020), which often focus on only one or two models. We also addressed a gap in the literature by quantifying the impact of preprocessing techniques, demonstrating a 27% reduction in scatter inaccuracy. This quantification offers valuable insights not present in studies like Kuang and Mouazen (2013), which reported improvements but lacked specific metrics.

The generalizability of our models is another key advancement. We demonstrated robust performance across both calibration and validation sets, improving upon studies like Yang et al., (2021) that struggled with overfitting issues. Furthermore, our focus on phosphorus content expands the application of NIR spectroscopy in chickpea analysis beyond the typical macronutrients studied by researchers like Mahajan et al., (2016b).

Lastly, our study has significant implications for industrial applications. We showed the effectiveness of simpler linear regression models, contrast with the complex models often favored in spectroscopic studies. This finding suggests the potential for more interpretable and easily implementable solutions in industrial settings.

This research advances NIR spectroscopy applications through a comprehensive comparison of five machine learning models (LR, SVR, RF, DTR, and NN) for phosphorus prediction in chickpeas. Our Linear Regression model achieved an exceptional R² of 0.945 (p<0.001), surpassing previous studies like Zhu et al. (2019) who reported R² of 0.89 for soil phosphorus prediction. The study’s strength lies in its thorough preprocessing methodology, incorporating noise elimination, data reduction, and scatter correction techniques. This approach builds upon Kuang and Mouazen’s (2013) work, which demonstrated significant improvements in spectral data quality.
        A key contribution is the identification of an optimal wavelength range (1520-1658 nm) for phosphorus prediction using the iPLS algorithm (RMSECV=0.458, R²=0.089, p<0.05). This finding provides more precise targeting compared to broader ranges used in previous studies. The high accuracy of our Linear Regression model (RMSE=0.044) exceeds that reported in comparable studies,  making it particularly valuable for chickpea breeding programs and quality control.

While our study advances NIR spectroscopy and machine learning for phosphorus prediction in chickpea flour, several limitations and future research opportunities exist. Expanding the sample size and variety of chickpea cultivars could enhance model robustness while extending the methodology to other legumes and food products would broaden its application. Future work should focus on developing models for the simultaneous prediction of multiple nutrients, as well as studying the impact of growing conditions on phosphorus content and NIR spectra. The development of portable NIR devices for on-site measurements would facilitate in-field applications. Additionally, investigating phosphorus content stability over time and various storage conditions would provide valuable insights for practical use. A direct comparison with wet chemistry methods would further validate our approach and enhance the interpretability of complex machine learning models could deepen our understanding of spectral-nutrient relationships. Finally, quantifying the cost-effectiveness of our method compared to traditional techniques through economic analysis would strengthen the case for its adoption. Addressing these aspects could further improve the accuracy and applicability of NIR spectroscopy and machine learning in food nutrient analysis, potentially advancing food quality assessment, agricultural practices and nutritional science.
 

I would like to thank the ICAR-National Bureau of Plant Genetic Resources and the team of Project “Biochemical Evaluation of Field and Vegetable Crops Germplasm” for providing spectrum and reference data.

There is no conflict of interest.