Introduction
Sweet potato grows as a strong plant all over the world and is a product compatible with
drought, temperature, and low fertile soils. Potatoes are high in starch, vitamins, minerals,
and non-mineral salts such as calcium, phosphorus, iron and low in calories. This product is widely consumed fresh, boiled, etc. due to its functions for various reasons, such as
improving immunity and preventing cancer, and its consumption is due to the abundance of nutrients such as carbohydrates, dietary fiber, minerals and other health-promoting
compounds such as beta-carotene, vitamin C, phenolic acids, etc. are on the rise. Conventional evaluation methods for the internal quality of potatoes are mostly destructive and inefficient.
In the practical production of potatoes, the quality evaluation system must have good
accuracy, high speed, and low cost. Such goals can be achieved using modern techniques
such as spectroscopy and electronic nose, as they do not require sample preparation, are non-destructive, efficient, fast, accurate, pollution-free, and inexpensive. Organic acids (OAs) are organic acidic compounds containing carboxyl groups that are widely present in organisms. Organic acids in fruits mainly include citric acid, malic acid, tartaric acid, and succinic acid. The traditional method for detecting OA concentrations is ion chromatography in the laboratory. Ion chromatographic testing requires standard solutions as a reference, also requires the use of chemical reagents, and organic acids must be measured separately. This is a tedious operation that wastes a lot of time. Therefore, a rapid detection technology is needed and preferred as an alternative.Near-infrared spectroscopy is a type of rapid detection technology that extracts spectral information from a sample through the difference between radiated light and reflected light. NIR technology has the advantages of fast performance, no use of chemical reagents and is also able to detect multiple components simultaneously.
Spectral signals can be further amplified by the combined use of the Fourier transform
technique. Fourier transform near-infrared spectroscopy has been widely used in the fields of food science, agricultural informatics, environmental monitoring, biomedicine, and pharmacy.Based on the simplicity of PLS regression, nonlinear methods are investigated to improve the PLS algorithm by embedding nonlinear core functions. This method plots the data before PLS scoring in a high-dimensional feature space, and the data converted in the new space characterize the samples. In this study, a neural network as a core function is designed to optimize PLS in the quantitative NIR analysis of OA concentrations in potato samples. A three-layer lattice with an adjustable number of neural nodes is designed to extract spectral feature variables to optimize the PLS core model.
Methodology
Potato samples were harvested and 248 of healthy size and almost the same size were
selected. The samples were transferred to the laboratory 24 hours after picking and stored at room temperature for 2 days. In the next 5 days, about 50 glands per day were selected and their OA concentration and FT-NIR spectrum were identified. Each potato sample was divided into two parts, half of which were used to detect the OA concentration and the other half to measure the NIR spectrum. The FT-NIR spectrum was measured using a PS-100 spectroradiometer (Apogee Instruments, INC., Logan, UT, USA) made in the USA.
Temperature and humidity were kept constant at 25 ° C and 47% during the spectrum
study.PLS kernel is an improved PLS method to deal with the nonlinear problem of spectral data. Raw data is mapped by a special nonlinear core function in high-resolution image space, so the original PLS linear algorithm can be used to discover the relationship between feature data and sample analysis. In short, this method can be done in two consecutive steps of mapping and regression. In modern studies, a neural network is a good tool for operating dynamic data, as it is flexibly taught by automatically fitting its link weights to the data-based model. A three-layer neural network was constructed in this study as a new nucleus for PLS output in the quantitative NIR analysis of potato OA concentrations.All 248 potato samples were divided into three parts for calibration, validation, and testing. The calibration section is used to create models and teach the model structure as well as the main algorithmic parameters. The validation section is used to check the model and optimize the parameter values. And the test section to evaluate the model. All 248 potato samples were divided into three parts for calibration, validation, and testing. The calibration section is used to create models and teach the model structure as well as the main algorithmic parameters. The validation section is used to check the model and optimize the parameter values. And the test section to evaluate the model.
Conclusion
Core PLS regression was applied to create FT-NIR calibration models to quantify OA
concentrations in potato samples. The proposed network architecture was used as a new
kernel conversion function to select attribute variables. The network was created connected with an input layer, a hidden layer, and an output layer. All 3114 wave number variables were transferred to the input layer. The same number of input nodes were generated to accept the data, and then perceptron units were applied, converting the data into a hidden layer. In the case of using a data-driven learning mechanism, the number of hidden nodes varies from 10 to 200 with step 10. Each Nh value was tested to screen for the best latent structure.
Perceptron calculations converted the hidden data into an output layer, and a total of 20
output neurons were generated in the output layer to reduce the dimensions. These output variables were mostly used for PLS regression.In general, neural perceptron units were adjusted with their link weights, which automatically matched the data. 20 output variables were delivered to the softmax MLR predictor. Predictive errors were used for 50 rounds of error-feedback repetition optimization on link weights. Figure 3 shows that the RMSEV gradually shrinks with more repetitions and gradually decreases for each Nh number. This phenomenon means that the initial feedback and error replication mechanism can optimize machine learning for the network kernel. Duplicate optimized network link weights were used to serve the network architecture as a core evaluation function to optimize PLS regression. The most optimal network structure was constructed with 130 hidden nodes and 20 output nodes.Then, the optimal network structure constructed with 130 hidden nodes and 20 output nodes is used as the core function for PLS regression. Hidden PLS variables were selected by network search mode. We tested PLS regression models with f = 1, 2… 20 based on the optimal network core. The results of model training for validation samples are shown in Figure 4. The optimal number of latent variables was determined as f = 8. The results of the network core model prediction and common cores are listed in Table 2. According to the principle of sample division introduced, PLS core models were quantified for FT-NIR analysis of potato OA concentration based on calibration samples and optimized by validation samples. The PLS model of the selected optimal network core should then be evaluated by 64 experimental samples that were unique to the model training process.
Spectral data of the experimental samples were entered into the core of the optimal network with 130 hidden nodes and 20 output nodes. Table 2 shows that for PLS kernel regression, the proposed network kernel performs better than conventional kernels, regardless of the model training process or the model evaluation process. Therefore, using neural network architecture to optimize the PLS regression kernel is a practical idea. FT-NIR calibration models have clearly improved compatibility by the adjustable network core.