Deep Learning Model Predicts and Classifies Pea Protein Content Using Visible and NIR Spectroscopy

In a recent study, a team of researchers from China Agricultural University and The Soil-Machine-Plant Key Laboratory of the Ministry of Agriculture of China examined a new way to detect protein content in peas. This study, which was published in the journal Food Chemistry, presented a new deep learning model called PeaNet. The PeaNet model utilizes visible and near-infrared (vis-NIR) spectroscopy data to accurately predict pea protein content (1).

Why is pea protein important?

Pea protein is generally made from yellow split pea, and it comes in three forms: pea protein isolate; pea protein concentrate; and textured pea protein (2). It is an essential component in plant-based diets and food product development because of its high nutritional value and sustainable production (1,2). In the past, pea protein was often analyzed using Kjeldahl or Dumas techniques. However, these traditional methods of determining protein content are time-consuming, costly, and destructive (1). Therefore, despite these methods’ accuracies, researchers have examined new, improved methods in analyzing pea protein content.

What is PeaNet?

PeaNet is a deep learning model that combines advanced spectroscopy with artificial intelligence (AI) to provide a universal, non-destructive, and accurate solution. The researchers built this model by using spectral data sets from 52 distinct pea varieties cultivated under varying environmental conditions, including differences in climate, soil type, and geographic region (1). In total, 156 visible and NIR spectra were analyzed. The data underwent preprocessing using Savitzky-Golay smoothing and multiplicative scatter correction to enhance spectral clarity and reduce noise (1). These preprocessing steps were vital in preparing the data for the deep learning model, ensuring accurate feature extraction and interpretation.

The researchers used a modified convolutional neural network (CNN) framework to build PeaNet. The goal of this study was for the researchers to train PeaNet to perform two specific functions. First, to predict the protein content of pea samples, and second, to classify the samples into discrete protein content categories (1). This multifunctional approach addresses a critical gap in current research, where models often specialize in either prediction or classification, but rarely both (1).

How did PeaNet perform in performance testing?

The researchers ran their model through a series of performance tests to evaluate whether it can effectively perform the two specific functions mentioned above. During performance testing, PeaNet achieved a coefficient of determination (R²) of 0.84 and a root mean square error (RMSE) of 0.87 on the primary test set (1). When tested for its classification accuracy for protein content levels, the model reached 85.33% (1). When tested on an independent validation set composed of different pea varieties not used in training, the model maintained its robustness, securing an R² above 0.80, an RMSE of 0.94, and a classification accuracy of 83.33% (1).

Other important results from the performance test include comparing PeaNet to other ML or deep learning models. The researchers found that when compared against partial least squares regression (PLSR), support vector regression (SVR), multilayer perceptron (MLP), and the AlexNet architecture, their model performed better. They theorized that this was because PeaNet was able to automatically interpret multidimensional spectral inputs without additional processing (1).

The study highlights several critical wavelengths, specifically at 550, 650, 950, 1200, 2000, and 2300 nm, that contributed significantly to the model’s predictive power (1). These wavelengths correlate with the physical and chemical characteristics of pea proteins, making them particularly valuable in spectral analysis.

What are the key takeaways from this study?

This study is important because it highlights the growing use of deep learning models for spectral analysis. The PeaNet model presented in this study shows that it can serve as a template for the exploration of future models, including ones that can analyze other crops (1). The authors advocate for the development of desktop or handheld vis–NIR devices tailored to specific operational environments, enabling on-site and real-time protein monitoring in breeding stations, food factories, and supply chains (1).

References

1. Xiao, T.; Xie, C.; Yang, L.; et al. A General Deep Learning Model for Predicting and Classifying Pea Protein Content via Visible and Near-infrared Spectroscopy. Food Chem. 2025, 478, 143617. DOI: 10.1016/j.foodchem.2025.143617
2. Cleveland Clinic, Everything You Should Know About Pea Protein. Cleveland Clinic. Available at: https://health.clevelandclinic.org/pea-protein (accessed 2025-07-15).