Development of an artificial neural network-based model for estimating the active iron content in grape leaves

Document Type : Research Paper

Authors

• Shabnam Firuzi ¹
• Ebrahim Sepehr ²
• Aydin Imani ¹
• soleiman hossein pour ³

¹ Dept. of Soil Science, Faculty of Agriculture, Urmia University, Urmia, Iran

² Department of Soil Science, Urmia University, Urmia, Iran

³ Assistant Professor, Department of Agricultural Machinery Engineering, Faculty of Agricultural Engineering and Technology, College of Agriculture and Natural Resources, University of Tehran, Karaj: Iran

Abstract

Keywords

• Iron estimation
• Grape leaves
• Artificial neural network
• Image processing

Main Subjects

• ..

EXTENDED ABSTRACT

Introduction

Iron is an essential micronutrient that plays a crucial role in the physiological and biochemical processes of plants. Recent advances in digital imaging have encouraged various industries to use this technology to determine the relationship between color and product content. While there are several methods for determining iron concentration in plants, the use of imaging offers distinct advantages over traditional laboratory methods.

Methods

The leaf samples were taken from the young and middle leaves of the grape branches. In June 2018, 55 randomly selected samples with varying degrees of iron chlorosis were taken from several vineyards. These samples were then transported to the laboratory for imaging and analysis of active and total iron content. After the solutions were prepared for each method, the concentrations of total and active iron were determined using an atomic absorption spectrophotometer. The samples were then placed in an environment with controlled lighting for the imaging procedures. After preprocessing, the images were converted from RGB to HSV color space and statistical features were extracted from the R, G, B, H, S and V color channels. Since leaf colors vary significantly, a powerful network is needed to handle these variations. Therefore, a multilayer perceptron neural network (MLP NN) was developed to model the experimental iron data and the corresponding images. The data were split into training (70%), validation (15%) and test (15%) to prevent overfitting and reduce the dependence of the model on the training data. The color features of the leaves and the amounts of iron measured from each leaf were considered as inputs and outputs of the model, respectively. The optimization of the network structure was trained with different numbers of neurons in the layers, ranging from one to 25, using the root mean square error (R²), root mean square error (RMSE) and mean absolute percentage error (MAPE) as evaluation criteria.

Results

The results of correlating color components with active and total iron showed that active iron correlates with the R, G, H, and S color channels with coefficients of 0.64, 0.58, 0.54, and 0.45, respectively, and that there is no correlation between changes in leaf color and total iron content. A multilayer perceptron network consisting of a hidden layer and an output layer, was used to estimate the active iron concentration using MATLAB 2018 software. The data for training, validation, and testing were selected using a random function. Tangent Sigmoid (tangsig) and linear (purelin) activation functions were applied to the hidden and output layers, respectively. The correlation between the predicted and laboratory data resulted in a coefficient of 0.84, an RMSE of 2.04, and a MAPE of 32.36. The high R² value confirms the ability of the model to estimate the available iron in plant leaves, while the low error values underline the generalization ability of the neural network model.

Conclusion

The algorithm presented in this study is an effective tool for estimating plant iron. For active iron, the neural network model has shown that it is possible to estimate the amount of iron in grape leaves based on the features extracted from the image (mean and standard deviation). The neural network model developed here, based on the output of the image processing system, proved to be successful in predicting the actual active iron amounts with a detection accuracy of the algorithm of 84%. It is proposed to use databases with more samples in future studies and also to develop other machine learning and plant models for estimating active iron content.

Author Contributions

All authors contributed equally to the conceptualization of the article and writing of the original and subsequent drafts.

Data Availability Statement

If the study did not report any data, you might add “Not applicable” here.

Acknowledgements

The authors would like to thank all participants of the present study.

Ethical considerations

The authors avoided data fabrication, falsification, plagiarism, and misconduct.

Conflict of interest

The author declares no conflict of interest.