Prediction of Fruit Traits of Camellia oleifera after Grafting Based on Deep Neural Networks

doi:10.21203/rs.3.rs-3563775/v1

Download PDF

Research Article

Prediction of Fruit Traits of Camellia oleifera after Grafting Based on Deep Neural Networks

https://doi.org/10.21203/rs.3.rs-3563775/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 04 Feb, 2024

Read the published version in Plant Methods →

You are reading this latest preprint version

Background: Camellia oleifera, an essential woody oil tree in China, propagates through grafting. However, in production, it has been found that the interaction between rootstocks and scions may affect fruit characteristics. Therefore, it is necessary to predict fruit characteristics after grafting to identify suitable rootstock types.

Methods: This study used Deep Neural Network (DNN) methods to analyze the impact of 108 6-year-old grafting combinations on the characteristics of C.oleifera, including fruit and seed characteristics, and fatty acids. The prediction of characteristics changes after grafting was explored to provide technical support for the cultivation and screening of specialized rootstocks. After determining the unsaturated fat acids, palmitoleic acid C16:1, cis-11 eicosenoic acid C20:1, oleic acid C18:1, linoleic acid C18:2, linolenic acid C18:3, oil content, fruit height, fruit diameter, fresh fruit weight, peel thickness, fresh seed weight, and the number of fresh seeds, the DNN method was used to calculate and analyze the model. The model was screened using the comprehensive evaluation index of Mean Absolute Error (MAE) and time consumption.

Results: The results show that when using 32 neurons in two hidden layer, the artificial neural network model had a Mean Absolute Percentage Error (MAPE) of less than or equal to 16.30% on the verification set and less than or equal to 17.69% on the test set. Compared with traditional machine learning methods such as support vector machines and random forests, the DNN method demonstrated more accurate predictions for fruit phenotypic characteristics, with MAPE increase rates ranging from 0% to 28.06% and 0% to 9.23%. In conclusion, the DNN method developed in this study can effectively predict the oil content and fruit phenotypic characteristics of C. oleifera, providing a valuable tool for predicting the impact of grafting combinations on the fruit of C. oleifera.

Camellia oleifera

Grafting

Artificial neural network

Fruit characteristics

C.oleifera Abel, an evergreen shrub or small tree in the Camellia of the Theaceae, is an essential woody edible oil tree in southern China with a cultivation history of over 2000 years[1]. It is an edible oil obtained by pressing the mature seeds of C.oleifera. With a unique woody oil in China, it has been known as a treasure in oil since ancient times[2]. The unsaturated fatty acids, including oleic acid, linoleic acid, and linolenic acid inseed oil of C.oleifera, are up to 83%~95%, ranking first among all edible vegetable oils [3-5]. In the high-quality and high-yield cultivation of C.oleifera, grafting is the primary means of breeding and improving C.oleifera varieties. Adopting grafting seedling cultivation technology can not only fully leverage the advantages of rootstock varieties, improve the quality of scion varieties, and increase the yield of C.oleifera trees, but also expand the planting range of C.oleifera, reduce planting costs, and increase economic benefits [6-8]. The research of artificial intelligence in agriculture has made many significant achievements. It has made great progress in the Internet of Things control, pest control, variety identification, and yield prediction, especially in wheat, corn, and rice[9-10]. However, in the complex planting environment of forestry, the diversity of terrain and climate environment makes the research and application of forestry still in the field of forest resource investigation. At the same time, as China's most promising woody oil tree, the vast market demand for C.oleifera and the supply-demand contradiction between production urgently require rapid improvement in quantity and quality. Predicting growth and yield is currently one of the challenging issues in precision agriculture, and many models have been proposed and validated to date. This issue requires multiple datasets, as growth and yield depend on many factors, such as climate, weather, soil, fertilizer use, and seed variety [11]. This indicates that predicting growth and yield is a challenging task. On the contrary, it consists of several complex steps. Nowadays, prediction models can reasonably estimate growth and yield, but better yield prediction performance is still needed. The rational and efficient use of artificial intelligence to promote the development of the C.oleifera industry is of great significance at both the economic and strategic levels.

Traditional machine learning (ML) methods, such as decision trees, naive Bayesian algorithms, fuzzy logic, support vector machines, and gradient enhancement algorithms, typically require manual participation in feature extraction and preprocessing steps before model use [12]. The handcrafted feature extraction and non-standard preprocessing measures limit the model's scalability and make the analysis work time-consuming and challenging. Experts with sufficient knowledge are always necessary and considered critical [13]. Support Vector Machine (SVM) is widely used in research fields such as data classification and prediction [14-18] due to its effectiveness in machine learning and reliance on structural risk minimization. In addition to classification and prediction, Support Vector Regression (SVR) is another application of SVM that specifically addresses regression problems. For example, Guo et al. introduced an active multi-classification method based on SVM [19]. Furthermore, a study by Yuan et al. successfully distinguished the level of moldiness in corn granules using SPA and SVM [20]. Random forests are built on the concept of Bagging. Random forest adds a new feature to Bagging, which is randomly choosing a random number of features and constructing a tree with them, repeating this procedure many times, and continuously averaging all the predictions made by all trees [21]. Therefore, random forests have processed the bias and variance components of errors and have been proven to be robust [22]. At present, the RF algorithm has been applied to agricultural research. For example, it was estimated sugarcane yield by RF algorithm according to different prediction ranges and achieved good results [23]. It was also used to establish a rice yield estimation model and conduct model precision evaluation [24].

In the past decade, neural networks have covered almost all scientific fields and become an essential component of various real-world applications. Deep neural networks (DNN) have multiple characteristics of nonlinear mapping, which can fit highly complex functions. It can learn and find the relevant information between them autonomously and continuously improve the optimization results. It has become an indispensable tool in various applications such as image classification, speech recognition, or natural language processing. These techniques have achieved highly high prediction accuracy, and in many cases, they are comparable to human performance. Due to the continuous development of modern computational methods, data-based prediction methods are increasingly being applied in various fields [25-26]. This method provides fast and accurate results for agricultural applications, such as predicting the internal temperature or transpiration rate of greenhouses [27-28]. It was used deep learning techniques such as convolutional neural networks and recurrent neural networks to predict soybean yield in the United States based on a sequence of remotely sensed images taken before the harvest [29-30]. Their model outperformed traditional remote-sensing-based methods by 15% regarding Mean Absolute Percentage Error (MAPE). The convolutional neural networks were used to predict crop yield based on satellite images [31]. Their model used 3-dimensional convolution to include spatiotemporal features and outperformed other machine learning methods. Khaki and Wang et al. designed a deep neural network model (Fig. 1) for predicting corn yields at 2247 locations from 2008 to 2016 [32]. Wang et al. designed a deep learning framework to predict soybean crop yields in Argentina, and they also achieved satisfactory results by using transfer learning methods to predict soybean yields in Brazil with less data [33]. The key to a deep neural network model is that it does not require specifying appropriate functions to fit the relationship between data. It can learn and find the relevant information between them autonomously and continuously improve the optimization results. The DNN also provides a general approximation framework, which means that no matter what function we want to learn, the deep neural network can be used to represent such a function [34-35]. Deep neural networks belong to the class of representation learning models that can find the underlying representation of data without handcrafted input of features and have multiple stacked non-linear layers that transform the raw input data into higher and more abstract representations at each stacked layer [36]. As the network grows deeper, more complex features are extracted, which contribute to the higher accuracy of results. Given the right parameters, DNNs are known to be universal approximation functions, meaning they can approximate almost any function, although it may be very challenging to find the right parameters [37-38]. Compared with the aforementioned neural network models in the literature, which were shallow networks with a single hidden layer, deep neural networks with multiple hidden layers are more powerful in revealing the fundamental nonlinear relationship between input and response variables [36], but they also require more advanced hardware and optimization techniques to train. For example, the neural network's depth (number of hidden layers) has a significant impact on its performance. Increasing the number of hidden layers may reduce the classification or regression errors, but it may also cause the vanishing/exploding gradients problem that prevents the convergence of the neural networks [32, 39-42].

To sum up, the success of DNNs in solving problems relies on several factors, including the size, type, quality, and preprocessing steps of the training data. What sets DNNs apart is their ability to autonomously learn and find correlations between data without the need to specify appropriate functions to fit the relationships. Deep neural networks can represent any function we want to know, providing an accurate framework for predicting various indicators in C.oleifera grafting combinations. This will enable practical data analysis and technical guidance for C.oleifera grafting. In this study, we hope to obtain a robust artificial neural network algorithm prediction model to quickly and accurately predict the oil content and phenotypic characteristics of C.oleifera fruit to help us predict the impact of grafting combinations on the fruit characteristics in advance. Based on these, we can establish accurate network models to predict various indicators after the grafting of C.oleifera, thus providing practical data analysis and technical guidance for C.oleifera grafting.

Materials

We conducted a study in Guanshang Town, Zhangshu City, Jiangxi Province, China, collecting fruits from over seven years old trees of twelve varieties, including CL3, CL4, CL18, CL23, CL27, CL40, CL53, CL116, and so on. 108 grafting combinations were obtained by scions and half-sibling seed rootstocks of 12 varieties (Table S1 and Table S2). The fruits of these tree combinations were collected for characteristics determination. The varieties were chosen 5 trees, each with similar age, good growth, and no pests or diseases, for each combination. We randomly selected and measured 30 fruits from each combination. After peeling and drying the seeds, we conducted the next measurement. Each test consisted of three biological replicates. The collection of all the samples complies with institutional, national, or international guidelines and legislation. The local forestry management department authorizes the collection of all samples for this research.

Determiation of Fruit characteristics

To measure the characteristics of fruit, including the weight (g), height (mm), and diameter (mm) of fresh fruits, the weight of dried seeds (g), dried kernels (g), and kernel oil content (g), we can use a vernier caliper with a sensitivity of 0.01 mm and a 0.01 g electronic balance. Additionally, we can calculate the fruit shape index, kernel ratio of dried seeds, oil content of dried kernels ratio, and dry seed oil content using specific formulas. The fruit shape index is calculated by dividing the fresh fruit height by the fresh fruit diameter and multiplying the result by 100%. The oil content of dried kernels can be calculated by dividing the weight of kernel oil by the weight of dried kernels, multiplied by 100%. The dry kernel oil ratio can be calculated by dividing the weight of kernel oil by the weight of dry seeds, multiplied by 100%. The dry seed oil ratio is calculated by multiplying the kernel oil content by the kernel-fruit ratio of dry seeds and multiplying the result by 100%.

The oil extraction from seeds by Soxhlet Extraction (SE)

All samples of C. oleifera seeds were powdered by a laboratory plant grinder. Approximately 10 g of ground sample were weighed and recorded as w0(g), then transferred to a Soxhlet extractor filled with 180 mL petroleum ether (60–90 ℃), and extracted at 88 ℃ for 6 h. Finally, the solvent was evaporated under vacuum. The residual was dried at 60 C in vacuum to a constant weight of w1(g). The oil content was calculated and expressed according to the formula: w = w1/w0 100%. Experiments were conducted in triplicate.

Fat analysis of the extracted C. oleifera oil by GC

As FA of C. oleifera oil presented in the form of fatty acid triglycerides in general, it must be transformed to be methyl esters of fatty acids means of sodium hydroxide. 0.2 ml of the extracted C. oleifera oil were put in 10 mL tube. Two millilitres of 0.5 mol/L sodium hydroxide–methanol was added into the tube, shook, and then placed at 60 ℃ in water-bath for 30 min, 5 mL nhexane were added. The supernatant was taken for injection to a gas chromatography spectrometer (HP6890 series, Agilent Techologies Inc.), equipped with a Hp-5 capillary column (30 m 0.25 mm 0.25 lm). The injector and detector temperature were set at 280 ℃. The oven temperature was programmed from 100 ℃ to 270 ℃ with a speed of 5 ℃/min and a final hold of 5 min. The signals from the detector were integrated as normalised percentages from the calibration curve by the HP software, and the main four individual fatty acid (oleic, linoleic, palmitic, stearic acid) were expressed as % of the total fatty acids. The unsaturated acids were considered as the sum of the oleic acid and linoleic acid.

Deep Neural Network (DNN)

This study used the combination of rootstock and scion varieties as input. It measured the parameters of C.oleifera outputs, including palmitoleic acid C16:1 (y1), cis-11 eicosanoic acid C20:1 (y2), unsaturated fatty acid (y3), oleic acid C18:1 (y4), linoleic acid C18:2 (y5), linolenic acid C18:3 (y6), oil rate (y7), fruit height (y8), fruit diameter (y9), and peel thickness (y10). Since there were significant variations in fruit phenotypic characteristics among different varieties, five parameters, namely fruit height, fruit diameter, fresh fruit weight, fresh seed weight, and number of fresh seeds, were selected instead of the variety number.

The hidden layer between the input and output layers can consist of one or more layers. The number of layers and neurons depends on the number of samples and the complexity of the task. Generally, a deeper and more layered model can improve accuracy by providing better nonlinear expression ability. This enables the model to learn complex transformations and fit more complex feature inputs. However, more network parameters also require more time and samples for training.

Out of 2108 samples, 1300 were used for training and 400 were used for validation, with the rest used for testing. Using the Relu activation function in the hidden layer is necessary for the network to learn nonlinear functions. The output layer uses the linear transfer function directly, and each hidden layer is connected with a dropout function (with a dropout rate of 0.1) to temporarily discard network information and reduce overfitting. The DNN uses the Stochastic Gradient Descent (SGD) optimizer for training (200 epochs), with a learning rate of 0.01. The loss function selects the Mean Squared Error (MSE), while the evaluation index selects the Mean Absolute Percentage Error (MAPE) to measure the performance of the model. The model that performed the best was selected based on the MAPE values of the validation set, and comprised 5 fully connected layers (1-5) and 8 neurons (2, 4, 8, 16, 32, 64, 128, 256) with different numbers.

This article constructs a double loop that combines the number of hidden layer layers and the number of neurons to form a network model. The training and validation sets are inputted to obtain the minimum MAPE value in each epoch and store the corresponding model information.

Data analysis

The experimental data were organized and analyzed using PyCharm 2020, Anaconda 3, and Tensorflow 2.1.

Characteristics of C.oleifera under different grafting conditions

C.oleifera is a valuable oil crop with a wide range of applications. To study the effect of different rootstock and scion combinations on the quality of C.oleifera, we conducted a comprehensive evaluation by taking into account the following parameters: palmitoleic acid C16:1, cis-11 eicosanoic acid C20:1, Unsaturated fat acid, oleic acid C18:1, linoleic acid C18:2, linolenic acid C18:3, oil content, fruit height, fruit diameter, and peel thickness.

We found that the fruit of varieties in C.oleifera significantly difference under different rootstock and scion combinations (Table S1 and Table S2). Different combinations of rootstocks and scions can significantly affect the fruit characteristicss of C.oleifera. When CL18, CL40, and CL53 are scions, there are differences in fruit characteristics among different grafting combinations. Grafting CL18 onto half-sib rootstocks of CL22, CL21, and CL27 significantly increased fruit height and diameter, and there was a significant difference compared to the original rootstock. Grafting with CL26 and CL40 half-sib rootstocks resulted in lower peel thickness while grafting with CL27 resulted in the highest thickness. CL40 had the highest grafting consequence on the stock, while the fruit diameter obtained the highest after grafting on CL21 half-sib rootstocks. It had the highest grafting consequence and the minor fruit diameter compared to CL4 half-sib stock and had significant differences from other combinations. The thickness of the fruit peel is lower after grafting with CL59 and CL4 half-sib rootstocks. CL53 has the highest fruit height and diameter after grafting with CL3 half-sib rootstock during scion, while it has the lowest after grafting with CL21 half-sib rootstock, which is entirely different from the performance of CL26 as rootstocks. When CL59 and CL40 are rootstocks, the fruit peel thickness after grafting is the lowest, while it is the highest after grafting with CL21 half-sibling rootstock.

As an essential characteristics in the production value of C.oleifera, content of oil and fatty composition were found to differ among grafted varieties in this study, and rootstocks may impact content of oil and fatty composition (Table S2). CL18 reaches its highest value at CL26 half-sibling rootstock, followed by CL59 half-sibling rootstock, while this stock has the lowest value, showing significant differences after grafting with multiple stock varieties. When CL40 is grafted onto CL27 half-sibling rootstock, it reaches the highest, followed by CL59, and CL21 and CL40 half-sibling rootstock are the lowest. Within the CL53 half-sibling rootstock combination, the CL21 and CL27 half-sibling rootstock are also lower, while the CL59 half-sibling rootstock is the highest. CL59 half-sibling rootstock may impact improving oil content. In addition, the content of oleic acid is an essential indicator for evaluating the quality of camellia oil, and the grafted oleic acid exhibits differentiated performance. When CL59 is used as the rootstock, there is a high proportion of oleic acid content among various combinations. While CL21 half-sibling rootstock is used as the rootstock, there is a decreasing trend in the oleic acid content of CL18 and CL40, while CL53 significantly increases. From these two indicators, CL59 half-sibling rootstock may have a significant regulatory effect on oil content and oleic acid content. Furthermore, it can also be seen that some varieties of rootstocks have different effects on the oil composition and fruit characteristics.

The MAPE values of fruit characteristics for different grafting combinations

After data segmentation and processing, 40 different model combinations were constructed from two neurons in the one hidden layer to 256 neurons in the five hidden layers. These combinations were used to construct different levels of DNN, train the network, and predict ten phenotypic characteristics to obtain the minimum MAPE table for different combinations in the DNN model under ten characteristics (Table S3).When the MAPE value is greater than 25, the difference between the predicted value and the actual value is too significant, and the predicted results are not accurate, so the data are discarded.Although it is impossible to achieve the optimal performance of the model on every characteristic, observing the performance of the model on the ten characteristics by ranking the average of the ten characteristics helps to evaluate the comprehensive performance of the model.

Prediction results of fruit performance characteristics under different model combinations

The MAPE values of 10 fruit characteristics obtained were averaged and calculated. It was found that the average values obtained from 2 × 32 and 2X64 were the smallest (Table S3). Although the MAPE values obtained from the characteristic parameters of 2 × 32 and 2 × 64 were not the minimum values, the data of the two models had a small gap from the minimum value, indicating the good comprehensive evaluation effect of the two models (Table S4 and Table S5). In addition, due to the small gap between the MAPE values of 2 × 32 and 2 × 64 and the fact that the model combination of 2 × 32 requires fewer network parameters for training, which is only half of the number of parameters in 2 × 64, fewer operational network parameters can improve the efficiency and quality of data operation. Therefore, the two-layer fully connected layer and 32 neurons were finally selected as the optimal model. After determining the model combination, the training and testing loss graphs of the DNN model under ten characteristics were obtained by training the ten characteristics again (Fig. 2). The MAPE of the test set data on the 2 × 32 combination was controlled within 17.69. It indicates that the model can obtain accurate prediction results for ten characteristics of the fruit.

Comparison results with support regression vector machine (SVR) and random forest (RF) models, DNN

The same dataset was predicted on SVR and RF models to obtain the MAPEs (Table S6 and Table S7), which were compared with the prediction results of DNN model.Among the various characteristicss of the fruit, the improvement rates of DNN compared to SVR on the validation set and test set were 0-31.37% and 0-28.61%, respectively, and the improvement rates of DNN compared to RF on the validation set and test set were 0-9.06% and 0-9.23%, respectively (Table 1). Overall, DNN can reduce errors compared to SVR and RF models, especially for y2, which has a significant improvement on the validation set, with an improvement rate of 23.85% and 2.85% (Table 1). In addition, the DNN model does not require manual specification of an appropriate function to fit the relationship between data and can autonomously learn and search for relevant information between the two to continuously improve and optimize the results, which helps to simplify the process and improve analysis efficiency.

Regulatory effect by grafting on the fruit characteristics and oil content of C.oleifera

Grafting is an ancient plant reproduction technique where the scion and rootstock are grafted and healed to form a new plant [43-44]. Rootstock plays an essential role in affecting scion growth [45-46], growth, development, yield, and potential flowering and fruit quality by releasing or improving the absorption and transportation of mineral nutrients, hormones, and carbohydrates, thereby affecting the increase in fruit yield, quality, and quality [47-49]. Therefore, selecting appropriate rootstocks is an important determining factor for achieving high and stable fruit yield performance. This study observed differences in fruit characteristics, oil content, and fatty acid composition among scions of the same variety under the action of differentiated rootstocks after grafting. Of particular concern is that the half-sib rootstocks of CL21 and CL59 significantly impact the characteristics of each variety. This indicates that using only species as the selection criteria for oil tea rootstocks may pose a risk to yield quality after oil tea grafting. At the same time, it also indicates that the interaction between rootstocks and scions in Camellia oleifera is not only the regulation of root growth by scions [50-51], but also reflected in the influence of fruit characteristics. Due to the significant impact of rootstocks on the physiological characteristics and other aspects of the growth and development of grafted plants, the reduction of scion growth is one of the most interesting phenomena. Therefore, it is necessary to explore the critical mechanisms of regulating fruit characteristics between rootstocks and scions and understand the role of hormones and other substances in the development of the entire post-grafting period.

The quality of fruit through data prediction after grafting achieves early selection of rootstock

To improve the yield and quality of C.oleifera, we collected data on rootstock and scion combinations of different varieties. By combining SVR, RF, and DNN, we successfully predicted ten phenotypic characteristics of the grafted fruit. Previous research in this field has mainly focused on grafting techniques without deep learning involved. The DNN model used in this study achieved a Mean Absolute Percentage Error (MAPE) of 17.69 on the test set, indicating the potential for deep learning models to predict the phenotypic characteristics of grafted oil tea fruits accurately. In addition to predicting fruit phenotypic characteristics, we also studied the performance differences between traditional machine learning methods and DNN models. The SVR model achieved a MAPE of 24.78 on the test set, while the RF model achieved a MAPE of 16.73. The DNN model exhibited an average improvement rate of 9.20% compared to the SVR model and 2.79% compared to the RF model across the ten fruit phenotypic characteristics. Notably, the DNN model exhibited a remarkable enhancement of 23.85% compared to SVR and 2.85% compared to RF in predicting cis 11 eicosenoic acid C20:1. This study demonstrates the advantages of DNN in predicting the phenotypic characteristics of grafted C.oleifera fruits, outperforming the SVR and RF models. Furthermore, DNN has the advantage of automatically extracting feature information from samples without manually selecting functions, which makes it superior to other machine learning algorithms. Using DNN models effectively simplifies the experimental process and prevents it from becoming too specialized. Nevertheless, choosing suitable network parameters to balance model accuracy with the number of parameters is crucial.

Our study used brute force enumeration to find appropriate network parameters and considered 40 combination methods with 1-5 hidden layers and 2¹-2⁸ neurons. The training result show the minimum MAPE for different combinations in the DNN model under ten characteristics (Table S3). However, it is observed that there needs to be more data for four characteristics, namely Unsaturated fatty acid, oleic acid C18:1, fruit height, and fruit diameter. This phenomenon mainly occurs when the number of hidden layers reaches four, as the increased depth leads to gradient explosion and prevents model training. We found that an ideal state without excessive network complexity can be achieved using 32 neurons in two hidden layer, resulting in a desirable MAPE for all ten characteristics characteristics. Overall, the effectiveness of DNN in predicting the phenotypic characteristics of grafted C.oleifera fruits has been demonstrated in our study. The automation and simplification provided by DNN models make them superior to SVR and RF models. However, careful consideration of suitable network parameters is necessary to strike a balance between accuracy and complexity. In conclusion, DNN is a valuable tool for predicting plant phenotypes, and its potential for application in agriculture is enormous.

This study focused on C.oleifera and found differences in fruit characteristics between the same variety and different rootstocks after grafting. Therefore, the research uses pre and post-grafting fruit phenotype data to establish a model for predicting fruit characteristics using deep neural networks.By setting different levels of hidden layers and the number of neurons, it was found that when using 2 hidden layers and 32 neurons, the overall performance achieved the best. The MAPE values of this model on the test set are 0-17.69. Compared with traditional support regression vector machines and random forests, deep neural networks have MAPE improvement rates of 0-28.61% and 0-9.23% for the ten characteristics on the test set, which have the advantages of high accuracy and avoid traditional machine learning model selection. They can predict fruit phenotypic characteristics data after C.oleifera grafting. This achievement can provide effective technical support for improving C.oleifera cultivation. In addition, an accurate prediction and evaluation system can be developed by adding more C.oleifera varieties, enriching fruit prediction parameters and input characteristics, improving model accuracy, and other means to assist in determining the impact of variety grafting on fruit characteristics, thereby reducing the time and labor costs of related experiments.

ACKNOWLEDGMENTS

Data summarized in this paper have been generated through work of several authors and we would like to thank them for their continuous efforts which contribute to the study of propagation of the Camellia oleifea.

AUTHOR CONTRINBUTIONS

Y.Z designed the experiment, performed data processing and drafted the manuscript. F.Y helped in analysis and data interpretation. L.L and K.W prepared the materials and performed the experiments. W.L and J.D participated in the design to the study, helped in data processing, and revision of the manuscript. Z.D and F.Y assisted in results interpretation and manuscript preparation. W.L conceived the study and revised the manuscript. All authors read and approved final version of the manuscript.

FUNDING

This research was funded by Pioneer and Leading Goose R&D Program of Zhejiang (2021C02038); Zhejiang Science and Technology Major Program on Agricultural New Variety Breeding (2021C02070-2) .

Availability of Supporting Data

The datasets used and/or analyzed in the current study are available from the corresponding author upon reasonable request.

Ethics approval and consent to participate

Not applicable

Consent for publication

Not applicable

Competing interests

The authors declare no conflict of interest

Permission to use varieties

The varieties owner has approved us to use 12 varieties of Camellia oleifera in current research. The local forestry management department authorizes the collection of all samples for this research.

Zhuang Ruilin. Chinese Camellia oleifera. Beijing: China Forestry Press. 2012.
Nie HY. Comprehensive utilization of Camellia oleifera seeds. Grain Oil Process Food Mach. 2004;10(06):39-41. (in chinese)
Staneley J. What are the effects of linear acid oxidation products on cardiovascular health. Lipid Tech. 2002;34(5): 59-61.
Becker N. Effects of saturated, monosaturated, and n-6 poly unsaturated, monosaturated fatty acids on plasma lipids, liproteins, and aporoteins in humans. Am J Clin Nutr. 2003;37 (11): 355.
Zhou SM, Wang Q. Development, utilization and prospect analysis of tea seed resources in China. Resour Prod. 2004;15(23):17-21.
Guo XM, Li KP, Zhang WY, Zou KS, Hu DN, Niu DK. Analysis on bottlenecks and counter measures of Camellia oleifera industry development in Jiangxi. Nonwood Forest Res. 2013;31,(2):1-6. (in chinese)
Lei XL, Huang JJ, He XS, Xu LC, Zhou WC, Liu S, et al. A study on key cultivation technique of super strong of high-yielding Camellia oleifera. J Central South Univ Forest & Tech. 2011; 31(6):39-43.
Cao JH, Lin WF, Chen JM. Studies of affinity between rootstock and scion. Trop Agri Sci. 2005;25(4):64-69. (in chinese)
Latif MS, Kazmi R, Khan N, Majeed R, Ikram S, Ali-Shahid MM. Pest prediction in rice using IoT and feed forward neural network. KSII Trans Internet Inf Syst (TIIS). 2022;16(1), 133-152.
Orchi H, Sadik M, Khaldoun M. On using artificial intelligence and the internet of things for crop disease detection: A contemporary survey. Agric. 2021;12(1), 9.
Shruthi U, Nagaveni V, & Raghavendra BK. A review on machine learning classification techniques for plant disease detection. In 2019 5th In conference on advanced computing & communication systems (ICACCS) ( pp. 281-284). IEEE .2019.
Yi W, Dai S, Jiang Y, Yuan C, & Yang L. Computer-aided Visual Modeling of Rice Leaf Growth Based on Machine Learning. In 2020 XXIII International Conference on Soft Computing and Measurements (SCM) ( pp. 226-229). IEEE. 2020.
Paturkar A, Gupta GS, Bailey D. Plant trait segmentation for plant growth monitoring. In 2020 35th International Conference on Image and Vision Computing New Zealand (IVCNZ) (pp. 1-6). IEEE, 2020.
Wang W, Xu Z, Lu W, Zhang X. Determination of the spread parameter in the Gaussian kernel for classification and regression. Neurocomputing. 2003;55(3):643-663.
LEE YJ, Mangasarian OL. SSVM: A smooth support vector machine for classification. Comput Optim Appl. 2001;20(1):5-22．
Wang WJ, Liang Z,Guo HS. A Multi-classification SVM algorithm based on data relationship. J Shanxi Univ Nat Sci Ed. 2012;35(2):224-230．
Zhang X, Ding SF. Mahalanobis distance-based twin multi-class classification Support Vector Machine. Comput Sci. 2016; 43(3):49-53．
Palaniappan S, Awang R. Intelligent heart disease prediction system using data mining techniques. IEEE/ACS international conference on computer systems and applications (pp. 108-115). IEEE, 2008.
Guo H, Wang W. An active learning-based SVM multi-class classification model. Pattern Recognit. 2015;48(5):1577-1597.
Yuan Y, Wang WJ, Chu X, Xie MJ. Selection of characteristic wavelengths using SPA and qualitative discrimination of mildew degree of corn kernels based on SVM. Spectrosc and Spectr Anal. 2016;36 (01):226-230. (in chinese)
Pintelas P, Livieris IE. Special issue on ensemble learning and applications. Algorithms. 2020;13(6):140.
Cutler DR, Edwards Jr TC, Beard KH, Cutler A, Hess KT, Gibson J, et al. Random forests for classification in ecology. Ecol. 2007;88(11):2783-2792.
Everingham Y, Sexton J, Skocaj D, Inman-Bamber G. Accurate prediction of sugarcane yield using a random forest algorithm. Agron sustain dev. 2016;36:1-9.
Wang F, Yao X, Xie L, Zheng J, Xu T. Rice yield estimation based on vegetation index and florescence spectral information from UAV hyperspectral remote sensing. Remote Sens. 2021; 13(17):3390.
Yin S, L, X, Gao H, Kaynak O. Data-based techniques focused on modern industry: An overview. IEEE Trans Ind Electron. 2014;62(1):657-667.
Chen Z, Pang M, Zhao Z, Li S, Miao R, Zhang Y, et al. Feature selection may improve deep neural networks for the bioinformatics problems. Bioinformatics. 2020; 36(5):1542-1552.
Nauta A, Han J, Tasnim SH, Lubitz WD. A new greenhouse energy model for predicting the year-round interior microclimate of a commercial greenhouse in Ontario, Canada. Inf Process Agric. 2023. 1-20
Jung DH, Lee TS, Kim K, & Park SH. A deep learning model to predict evapotranspiration and relative humidity for moisture control in tomato greenhouses. Agro. 2022; 12(9), 2169.
Kim N, Ha KJ, Park NW, Cho J, Hong S, Lee YW. A comparison between major artificial intelligence models for crop yield prediction: Case study of the midwestern United States, 2006-2015. ISPRS Int Geo-Inf. 2019; 8(5): 240.
You J, Li X, Low M, Lobell, D, Ermon S. Deep gaussian process for crop yield prediction based on remote sensing data. Proc AAAI Con on Artif Intell. 2017; 31(1): .
Nevavuori P, Narra N, Lipping T. Crop yield prediction with deep convolutional neural networks. Comput Electron Agric. 2019;163, 104859.
Khaki S, Wang L. Crop yield prediction using deep neural networks. Front Plant Sci. 2019;10:621.
Wang AX, Tran C, Desai N, Lobell D, Ermon S. Deep transfer learning for crop yield prediction with remote sensing data. In Proceedings of the 1st ACM SIGCAS Conference on Computing and Sustainable Societies (pp. 1-5). 2018.
Bau D, Zhu JY, Strobelt H, Lapedriza A, Zhou B, Torralba A. Understanding the role of individual units in a deep neural network. Proc Nat Acad Sci. 2020;117(48): 30071-30078.
Du Y, Zaki TA. Evolutional deep neural network. Phys Rev E. 2021;104(4): 045303.
LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015; 521(7553): 436-444.
Hornik K, Stinchcombe M, White H. Universal approximation of an unknown mapping and its derivatives using multilayer feedforward networks. Neural Netw. 1990;3(5):551-560.
Goodfellow I, Pouget-Abadie J, Mirza M, Xu B, Warde-Farley D, Ozair S, et al. Generative adversarial networks. Commun ACM. 2020; 63(11):139-144.
Modarres C, Astorga N, Droguett EL, Meruane V. Convolutional neural networks for automated damage recognition and damage type identification. Struct Control Health Monit. 2018; 25(10):e2230.
Rosa JP, Guerra DJ, Horta NC, Martins RM, Lourenço NC, Rosa JP, et al. Overview of artificial neural networks. In: Using Artificial Neural Networks for Analog Integrated Circuit Design Automation. SpringerBriefs in Applied Sciences and Technology. Springer, Cham. 2020; pp.21-44.
Narkhede MV, Bartakke PP, Sutaone MS. A review on weight initialization strategies for neural networks. Artif Intell Rev. 2022;55(1):291-322.
Manoharan A, Begam KM, Aparow VR, Sooriamoorthy D. Artificial Neural Networks, Gradient Boosting and Support Vector Machines for electric vehicle battery state estimation: A review. J Energy Stor. 2022;55,105384.
Goldschmidt EE. Plant grafting: new mechanisms, evolutionary implications. Front Plant Sci. 2014;5:727.
Rasool A, Mansoor S, Bhat KM, Hassan GI, Baba TR, Alyemeni MN, et al. Mechanisms underlying graft union formation and rootstock scion interaction in horticultural plants. Front Plant Sci. 2020;11,590847.
Martínez-Ballesta MC, Alcaraz-López C, Muries B, Mota-Cadenas C, Carvajal M. Physiological aspects of rootstock-scion interactions. Sci Hortic. 2010;127(2):112-118.
Albacete A, Martínez-Andújar C, Martínez-Pérez A, Thompson AJ, Dodd IC, Pérez- Alfocea F. Unravelling rootstock×scion interactions to improve food security. J Exp Bot. 2015;66(8): 2211-2226.
Turhan A, Ozmen N, Serbeci MS, Seniz V. Effects of grafting on different rootstocks on tomato fruit yield and quality. Horticultural Science. 2011; 38(4):142-149.
Rouphael Y, Schwarz D, Krumbein A, Colla G. Impact of grafting on product quality of fruit vegetables. Sci Hortic. 2010;127(2):172-179.
Grieneisen ML, Aegerter BJ, Scott Stoddard C, Zhang M. Yield and fruit quality of grafted tomatoes, and their potential for soil fumigant use reduction. A meta-analysis. Agron Sustain Dev 2018;38:1-16.
Long, W., Yao, X. H., Wang, K. L., Wang, Y. The analysis of the affinity of rootstock and scion in the grafting seedlings of Camellia oleifera. Acta Agric. Univ. Jiangxiensis (Nat Sci Ed) .2013;35 (2):346–351.
Zeng J, Liu J, Lian L, Xu A, Guo X, Zhang L, et al. Effects of scion variety on the phosphorus efficiency of grafted Camellia oleifera seedlings. Forests. 2022;13(2): 203.

Table 1 Comparison of MAPE value improvement rates among DNN models and SVR and RF models under ten characteristics. Positive values as an increase, and negative values (-) a decline. y1: palmitoleic acid C16:1 , y2: cis-11 eicosanoic acid C20:1, y3: unsaturated fatty acid, y4: leic acid C18:1, y5: linoleic acid C18:2, y6: linolenic acid C18:3, y7: oil rate, y8: fruit height, y9: fruit diameter, y10: pericarp thickness.
Increase	Data	y1	y2	y3	y4	y5	y6	y7	y8	y9	y10
SVR	val	31.37	23.85	-1.85	7.91	20.45	5.5	5.39	-6.82	5.23	0.96
SVR	test	28..61	20.47	-5.66	4.71	13.99	3.33	4.79	-11.72	1.28	-2.47
RF	val	0.97	2.85	0	-1.87	9.06	7.08	5.49	-3.69	4.5	3.47
RF	test	-5.74	4.27	1.75	-7.69	-1.49	9.23	5.09	-18.01	1.64	3.28

Additional file 8 (Table S8) is not available with this version

No competing interests reported.

TableS1AnalysisofvarianceinfruittraitsunderdifferentgraftingcombinationsinCamelliaoleifera.xlsx
Additional file 1: Table S1. Analysis of variance in fruit traits under different grafting combinations in Camellia oleifera.
TableS2AnalysisofvarianceinfatandfattyacidcompositionoffruitunderdifferentgraftingcombinationsinCamelliaoleifera.xlsx
Additional file 2: Table S2. Analysis of variance in fat and fatty acid composition of fruit under different grafting combinations in Camellia oleifera.
TableS3TheMAPEvaluesoftencharacteristicsoffruitsunderdifferentcombinations.xlsx
Additional file 3: Table S3. The MAPE values of ten traits of fruits under different combinations.
TableS4TheMAPEvalueoftenfruitcharacteristicsinDNNmodlewith2hiddenlayersand32neurons.xlsx
Additional file 4: Table S4. MAPE value of 2×32 combination under ten features in DNN Model.
TableS5TheMAPEvalueoftenfruitcharacteristicsinDNNmodlewith2hiddenlayersand64neurons.xlsx
Additional file 5: Table S5. MAPE value of 2×64 combination under ten features in DNN Model.
TableS6TheMAPEvaluesbasedonthevalidationandtestingsetsofthetencharacteristicsintheSVRmodel.xlsx
Additional file 6: Table S6. MAPE value of SVR Model under ten characteristics.
TableS7TheMAPEvaluesbasedonthevalidationandtestingsetsofthetencharacteristicsintheRFmodel.xlsx
Additional file 7: Table S7. MAPE value of RF model under ten characteristics.

Download PDF

Journal Publication

published 04 Feb, 2024

Read the published version in Plant Methods →

Editorial decision: Revision requested
01 Dec, 2023
Reviews received at journal
08 Nov, 2023
Reviewers agreed at journal
06 Nov, 2023
Reviewers invited by journal
06 Nov, 2023
Editor assigned by journal
06 Nov, 2023
Submission checks completed at journal
06 Nov, 2023
First submitted to journal
05 Nov, 2023

You are reading this latest preprint version

Prediction of Fruit Traits of Camellia oleifera after Grafting Based on Deep Neural Networks

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Material and Methods

Results

Discussion

Conclusion

Declarations

References

Table

Additional file

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1