Introduction

As one of the staple foods for more than half of the world’s population, rice plays an extremely important role in maintaining world food security and social stability1. The yield and quality of rice directly affect the stability and quality of the global food supply, which in turn affects people’s living standards and social development. Therefore, ensuring high yield and quality of rice is one of the most important measures to maintain global food security and social stability2.

In China, about 99 million hectares are affected by salinization, of which about 15 million hectares can be used for agricultural production, representing about 10% of China’s cultivated land3. Saline rice can suffer from poor cooking and flavor quality due to salt stress4. Mixed fertilizer application enables the simultaneous utilization of quick-acting nutrients and total nutrients, which is conducive to improving the quality of rice5. Rice quality includes four aspects: processing quality, appearance quality, cooking flavor quality, and nutritional quality6. Starch accounts for 85–90% of rice and is mainly composed of amylose and branched chain starch. Starch is a key determinant of all aspects of grain quality, with cooking and eating quality being the most affected7,8. Traditional testing methods determine whether organic fertilizers have been added by analyzing the nutrient content of the planting soil, which is time-consuming, labor-intensive, and costly, although the testing accuracy is high, and is not suitable for rapid testing needs. Therefore, it is necessary to develop an accurate and rapid method for rapid detection of whether organic fertilizers are added during rice cultivation. Raman spectroscopy technology has many advantages such as non-destructive, rapid, and low cost9, and has been widely used in various fields such as food10, medicine11, agriculture12, and so on.

Yang et al.13 found that application of chemical fertilizers affected rice starch and protein content. Iqbal et al.14 studied the effect of organic-inorganic integrated amendments on leaf physiology and seed starch viscosity characteristics and soil quality of rice and found that chemical-organic fertilizer blending increased the seed starch content and straight chain starch content of rice. Ichinose et al.15 focused on how to quantitatively predict the digestibility of rice starch using Raman spectroscopy and multivariate calibration analysis, this study provided a new method for rapid and non-destructive assessment of the nutritional quality of rice. Sha et al.16 developed a method to improve the efficiency of Raman spectroscopy in rice variety identification by combining principal component analysis (PCA), window analysis, hierarchical cluster analysis, and support vector machine (SVM). Giang et al.17 used Raman spectroscopy combined with multivariate analysis methods including PCA, hierarchical cluster analysis, and k-nearest neighbor classification (K-NN) to successfully distinguish four different rice varieties in Vietnam. By analyzing the Raman spectra of the rice samples, the researchers identified key factors such as amylose chain, branched amylose, branched amylose, and protein content, which are essential for distinguishing between different varieties of rice. Sliwinska-Bartel et al.18 provided a comprehensive review of fraud in the global rice industry, including species mislabeling, geographic origin falsification, and adulteration, and explored the role of various analytical tools in detecting the authenticity of rice. Shi et al.19 utilized various preprocessing and machine learning algorithms combined with Raman spectroscopy to distinguish the freshness of rice of different years and showed that partial least squares discriminant analysis (PLS-DA) and SVM performed well in identifying the freshness of rice with 100% accuracy. These studies demonstrate that different algorithms such as K-NN, neural networks, and SVM can make good use of the intrinsic information in spectra.

In summary, there are fewer studies on the rapid identification of rice chemical-organic fertilizers, and the accuracy of Raman spectroscopic techniques for determining rice fertilization remains to be studied. In this paper, SVM, feedforward neural network (FFNN), and FFNN complexes are used for modeling and classification to explore the feasibility of using spectroscopic techniques to determine the fertilizer distribution during rice growth, with the ultimate goal of determining the accuracy of various algorithms to make recommendations for rice fertilizer application identification. The research results are intended to provide experimental and data support for smart agriculture and soil improvement.

Materials and methods

Experimental samples

Rice cultivation started in 2023, and the rice samples in this paper were obtained from the data of 2023–2024, and the rice variety was TongJing 612. Rice was planted in the test field of Yanjiatun, Taonan City, Baicheng City, Jilin Province (122° 48′ E, 45° 08′ N, as shown in Fig. 1). The experimental field was a saline soil with total soil nitrogen storage of 450 mg/kg, effective phosphorus of 30 mg/kg, effective potassium of 105 mg/kg, and a pH value of 8. The inorganic compound fertilizer used for planting (N-P2O5-K2O ratio: 20-12-10) was purchased from Jiangsu Huachang Chemical Co. The organic fertilizers used for planting included soil conditioner (organic matter content ≥ 75%) and multi-functional microbial fertilizer (effective bacterial count ≥ 200 million/g, total organic matter ≥ 85.0%, organic matter ≥ 75.0%, easily oxidizable organic matter ≥ 20.0%), which were purchased from Beijing Jiaboqian Biotechnology Co.

Fig. 1
figure 1

Geographic location of the test site, sample-related treatments, and numbering (created using Baidu Maps).

Full size image

There were seven experimental plots of 30 m2 each, one of which was set up as the control plot, and six plots were applied with different chemical fertilizers and organic fertilizers as follows: (1) no fertilizer treatment (CK), without any fertilizer treatment, with all the nutrients supplied by the soil; (2) single application of chemical fertilizer (CF), with all the nutrients supplied by the soil and inorganic composite fertilizer (1.9 kg/plot); (3) single application of organic fertilizer (OF), all nutrients were provided by soil and organic compound fertilizer (24 kg/plot); (4) applying full chemical fertilizer along with 25% organic fertilizer (D25), the amount of chemical fertilizer remained unchanged, while organic compound fertilizer was reduced to 25% (6 kg/plot); (5) applying full chemical fertilizer along with 50% organic fertilizer (D50), the amount of chemical fertilizer remained unchanged, while organic compound fertilizer was reduced to 50% (12 kg/plot); (6) applying full chemical fertilizer and 75% organic fertilizer (D75), the fertilizer dosage remained unchanged, while the organic compound fertilizer was reduced to 75% (18 kg/plot); (7) applying full chemical fertilizer and full organic fertilizer (D100).

Samples were taken at five points from each of the seven treated lands. Rice seeds that were full, well-proportioned, medium-sized, free of insect pests, mold and mechanical damage were manually selected. About 10 kg of each sample was taken after manual selection, and after hulling with a hulling machine (JLGJ-45, Jing’ao, China), deteriorated and defective brown rice was sifted out, and the brown rice seeds with intact and undamaged seeds were selected for the subsequent experiments. The rice was then milled into rice using a rice milling machine (YL-M003, Yungu, China), then crushed with a mortar and pestle, and then sieved through a 100-mesh sieve. The sieved powdered samples were put into different numbered plastic bags and stored in a cool and dry place. Each sample was weighed 2 kg and the amylose content in rice was determined by Jilin Junshi Inspection and Testing Technology Co. with reference to the national standard GB_T 15683-2008.

Raman spectroscopy data collection

The experimental setup was a TriVistaTM555CRS three-stage microzone confocal Raman spectrometer from Princeton. It was equipped with an OLYMPUS 10× magnification objective and a CCD detector. Liquid nitrogen was used as the coolant for the CCD detector. The excitation light sources used included a 532 nm laser (MW-GL type, Changchun LeiMagnet Optoelectronics Technology Co., Ltd.), with laser linewidths of 0.1 nm, and laser power ranging from 0 ~ 500 mW, and the laser power for this experiment was 92.5 mW. Experimental environments: temperature of 20 °C, relative humidity of 30%, reflectance of the laboratory wall < 3%, and a dust concentration of 3 × 105 (≥ 0.5 μm/m3).

The Raman spectral data of 210 rice samples were collected. The spectral scanning wave number range was 200 ~ 2000 cm−1, the resolution was 0.585 cm−1, the integration time was 1s, and the number of integrations was 10 times. The powder sample was taken and dispersed uniformly on a slide and flattened using a coverslip to ensure that the surface of the sample was flat and free of particles or agglomerates. The average spectrum of the data collected using the five-point sampling method was used as the raw Raman spectral data of this sample. Repeat the above measurements by taking the powder from the sample again to obtain 40–50 raw Raman spectral data, 30 spectral data were selected for modeling and analysis after eliminating the invalid data, and each Raman raw data contained a total of 3315 data points.

Experimental data processing

In this paper, multiple methods are used to preprocess the original spectra, and various preprocessing methods are compared and analyzed. Meanwhile, in order to eliminate the effect caused by the difference in spectral intensity between different samples, the maximum–minimum normalization (Norm) process was used to scale the values to the range of [0, 1] before using the various preprocessing methods, and the normalization is shown in Fig. 2. The spectral preprocessing methods used include multiple scattering correction, standard normal variable, and Savitzky–Golay first derivative (SGD, polynomial order of 2 and number of smoothing points of 15). The MSC method is used to separate the information absorbed in the sample from the scattering signals in the spectral data20. The standard normal variable (SNV) method standardizes the spectral data by subtracting the mean value and dividing it by the standard deviation, resulting in a mean value of 0 and a standard deviation of 1 for each wavelength21. Savitzky–Golay filtering algorithm is based on polynomial fitting and is commonly used to eliminate spectral noise22.

Fig. 2
figure 2

Normalized Raman spectra of rice for seven different fertilization treatments.

Full size image

Machine learning model

Different algorithms are used for the recognition of rice fertilization, including SVM, FFNN, and K-NN. SVM is an efficient classifier whose core idea is to construct an optimal classification decision surface, that is, the classification hyperplane23, so as to achieve accurate classification of the data. SVM has a low computational complexity and a high classification accuracy and shows good results on small sample data24. FFNN is a type of artificial neural network25 that offers significant advantages in classification tasks, including the ability to handle complex data, structural flexibility, and efficiency in training and prediction. These features enable them to effectively capture and represent complex patterns and relationships in data, thus improving classification accuracy and efficiency26. K-NN is one of the most classical and simplest methods in classification. The K-NN rule is based on the idea that instances that are close in the feature space may belong to the same category. To classify the query instances, the category with the highest frequency of occurrence among the k nearest neighbors is selected27. This classifier has the advantages of simplicity, good performance, low sensitivity to local noise patterns, and describability28.

Modeling methods and evaluation

The 210 Raman data were randomly divided into 147 training set data and 63 validation set data in the ratio of 7:3. Different algorithms were used to identify the fertilization treatments of rice, including FFNN, SVM, and K-NN. The evaluation parameters of the models included the area under the ROC curve (AUC) of the receiver operating characteristic curve (ROC curve) and the confusion matrix for accuracy, precision, sensitivity, and F1 score.

The AUC quantifies the overall effectiveness of the classifier. the closer the AUC value is to 1, the closer the performance of the classifier is to the ideal, while an AUC value of 0.5 is equivalent to a random guess. Since AUC exists for each category, this paper ultimately uses the average AUC as the final metric. Accuracy measures the model’s overall ability to correctly predict on all categories; precision and sensitivity assess the accuracy and coverage of the model’s judgments on the positive categories, respectively; these metrics collectively reflect the model’s performance in different aspects, which helps us to identify the model’s strengths and weaknesses, and thus perform targeted optimization.

As the model is a multi categorical model, assuming that each category contributes equally to the model, the macro-averaged accuracy, macro-averaged sensitivity, and macro-averaged F1 scores of the model are calculated using the macro method29, and the evaluation parameters of the model are calculated as follows:

$${\text{Accuracy}} = \frac{{{\text{TP}} + {\text{TN}}}}{{{\text{TP}} + {\text{TN}} + {\text{FP}} + {\text{FN}}}}$$
(1)
$${\text{Precision}}_{{{\text{macro}}}} = \frac{1}{{\text{q}}}\mathop \sum \limits_{{{\text{c}} = 0}}^{{{\text{q}} - 1}} { }\frac{{{\text{TP}}_{{\text{c}}} }}{{{\text{TP}}_{{\text{c}}} + {\text{FP}}_{{\text{c}}} }}$$
(2)
$${\text{Sensitivity}}_{{{\text{macro}}}} = \frac{1}{{\text{q}}}\mathop \sum \limits_{{{\text{c}} = 0}}^{{{\text{q}} - 1}} { }\frac{{{\text{TP}}_{{\text{c}}} }}{{{\text{TP}}_{{\text{c}}} + {\text{FN}}_{{\text{c}}} }}$$
(3)
$${\text{F}}1_{{{\text{macro}}}} {\text{score}} = \frac{1}{{\text{q}}}\mathop \sum \limits_{{{\text{c}} = 0}}^{{{\text{q}} - 1}} { }\frac{{2{\text{TP}}_{{\text{c}}} }}{{2{\text{TP}}_{{\text{c}}} + {\text{FN}}_{{\text{c}}} + {\text{FP}}_{{\text{c}}} }}$$
(4)

where TP is true positive, TN is true negative, FP is false positive, FN is false negative, TPc, FPc, and FNc are the number of TP, FP, and FN predictions made by the classifier for class C, respectively, and q is the number of categories.

Results and discussion

Rice consists of a variety of components and the main characteristic bands of the different component groups are present in the spectrum. As shown in Fig. 3. Each characteristic spectral feature of rice shows the vibrational or rotational modes (stretching, swaying, bending, shearing, and torsional fundamental vibrations) of the different functional groups, as well as the ring components30. The Raman characteristic peaks of rice are listed and attributed in Table 115,17,30,31,32.

Fig. 3
figure 3

Raman spectra (de-baseline) of rice from seven different fertilizer treatments.

Full size image
Table 1 Vibrational band attribution in rice spectra.
Full size table

In this study, the performance differences of three models, SVM, K-NN, and FFNN, were compared after applying four different preprocessing techniques, as shown in Fig. 4 and Table 2. These preprocessing techniques include basis normalization (Norm), multivariate scattering correction (Norm + MSC), standard normal distribution of variables (Norm + SNV) and SG filtering (Norm + SGD). It is found that the multivariate scatter correction improves the performance of all the models most significantly, especially the K-NN model, which shows outstanding performance with this preprocessing technique, with significant improvement in all the evaluation metrics such as accuracy, precision, sensitivity, F1 score, and AUC value. Comparatively, Norm + SNV and Norm + SGD preprocessing provide more limited performance enhancement for FFNN and SVM models, suggesting differences in the sensitivity of different models to the preprocessing methods. Basis normalization provides an important benchmark for comparison, showing the baseline performance of all models without advanced preprocessing. Therefore, an appropriate preprocessing strategy is decisive for improving the classification accuracy and generalization ability of a particular model, and the selection of the best preprocessing method needs to be based on the model characteristics and the specific requirements of the dataset.

Table 2 Statistics on the evaluation indicators of the three models.
Full size table
Fig. 4
figure 4

Comparison of radargrams of the three algorithms.

Full size image

In this study, the performance differences of three models, SVM, K-NN, and FFNN, were compared after applying four different preprocessing techniques, as shown in Fig. 4 and Table 2 (see Supplementary Fig. S1 to Supplementary Fig. S6 online). These preprocessing techniques include basis normalization (Norm), multivariate scattering correction (Norm + MSC), standard normal distribution of variables (Norm + SNV) and SG filtering (Norm + SGD). It is found that the multivariate scatter correction improves the performance of all the models most significantly, especially the K-NN model, which shows outstanding performance with this preprocessing technique, with significant improvement in all the evaluation metrics such as accuracy, precision, sensitivity, F1 score, and AUC value. Comparatively, Norm + SNV and Norm + SGD preprocessing provide more limited performance enhancement for FFNN and SVM models, which suggests that there are differences in the sensitivity of different models to the preprocessing methods. Basis normalization provides an important benchmark for comparison, showing the baseline performance of all models without advanced preprocessing. Therefore, an appropriate preprocessing strategy is decisive for improving the classification accuracy and generalization ability of a particular model, and the selection of the best preprocessing method needs to be based on the model characteristics and the specific requirements of the dataset.

Fig. 5
figure 5

Scatter plot of actual versus predicted values of amylose content in rice.

Full size image

Figure 5 shows the distribution of the predicted and actual values of straight-chain amylose content in rice by the SVM regression model established after MSC preprocessing. In the figure, the predicted and actual values of straight-chain amylose content are distributed near the 1:1 line, which proves the validity of the Raman technique for the detection of straight-chain amylose content after SVM combined with MSC processing. At the same time, this regression model shows excellent performance, where the R2 value of 0.93 indicates that the model successfully captures most of the variability in the data, and the low values of MSE and RMSE (0.04 and 0.2, respectively) demonstrate the high accuracy and low error of its prediction.

Conclusion

The objective of this study was to investigate the characteristics of rice starch changes under different fertilization conditions with different preprocessing methods (Norm, Norm + MSC, Norm + SNV, and Norm + SGD) and machine learning methods (SVM, FCNN, and K-NN) to distinguish between types of rice fertilization and different ratios of chemical and organic fertilizers applied. The results showed that the prediction coefficient of determination, the root mean square error of prediction, and the average relative error of prediction of the starch content detection model built by the SVM regression model after Norm + MSC preprocessing were 0.93, 0.04%, and 0.20%, respectively, which indicated that its prediction had high accuracy and low error. Meanwhile, the accuracy of Raman data after Norm + MSC preprocessing is close to 1 in all three classification models, and the performance of Norm + SNV and Norm + SGD preprocessing is much lower than that of Norm + MSC, although they also improve the performance. Meanwhile, the overall performance of SVM is higher than that of FCNN and K-NN in the three classification performances. Therefore, the three models processed by MSC can be used for rice fertilization discrimination, while SVM has an advantage in rice fertilization discrimination. The method in this study provides a quick way to detect rice starch changes and improves the accuracy of fertilizer discrimination and helps optimize agricultural fertilization, which is important for promoting sustainable agriculture. Future research could further explore the changes in the spectral characteristics of rice under different fertilization conditions and how these techniques can be applied to other crops for accurate fertilization management. In addition, researchers can explore more kinds of machine learning algorithms and consider combining these techniques with IoT and big data analytics to further enhance the efficiency and effectiveness of smart agriculture.