Study on the Identification of Mildew Disease of Cuttings at the Base of Mulberry Cuttings by Aeroponics Rapid Propagation Based on a BP Neural Network

Abstract

Accurate detection of cutting diseases in the process of aeroponic rapid propagation is very important for improving the rooting rate and survival rate of cuttings. This paper proposes to use image processing, with a dataset of the growth of mulberry cuttings and a backward propagation (BP) neural network, to identify mildew on the roots of mulberry branches in the process of rapid propagation, before extracting texture and color features. An intelligent control aeroponics system was designed to control the ambient temperature and humidity of the entire rapid propagation incubator according to the mildew rate, thereby improving the rapid propagation time of aeroponics, as well as the rooting and survival rates. In order to distinguish the extracted features, they were classified and identified using a constructed BP neural network model. The results indicated that the performance of the neutral network showed the lowest mean square error in the validation set after three rounds of training; therefore, the model of the third round was chosen as the best model. Furthermore, the training effect of the model revealed that the BP neural network model had good stability and could accurately identify diseases in the root zone of mulberry cuttings. After using MATLAB for neural network training, the regression results revealed correlation coefficients R of 0.98 for the fitting curve of the training dataset, 0.98 for the fitting curve of the test set, and 0.99 for the fitting curve of the validation set, indicating that the prediction results aligned well with the actual results. It can be concluded that research method described in this paper had excellent performance in identifying the health status of mulberry cuttings during the aeroponics rapid propagation process, and it was able to quickly and accurately identify mulberry cuttings affected by mildew disease with an accuracy rate of 80%. This research provides a technical reference for aeroponics rapid propagation factories and intelligent nurseries.

1. Introduction

The first major agricultural challenge is to produce enough food for a continuously growing population in the complex context of population growth and urbanization, poverty, increasing demand for food, growing competition for water and land, climate change, climate uncertainty, and drought [1]. Soils and crops are particularly vulnerable to climate change and environmental stress. In many agricultural systems, the soil biodiversity and ecosystem services provided by soils are threatened by a range of natural and human-driven factors. Agricultural soils are often subject to agronomic practices that disrupt soil nutrient networks and reduce the long-term productivity of the soil [2]. Therefore, future studies will focus on more efficiently producing food from natural resources, including soil, water, and air. However, approximately one-quarter of arable land has been declared unproductive, uncultivable, and unsuitable for agricultural activities [3]. Researchers have begun to address this challenge using currently available technologies. Conventional soil cultivation has resistance and constraints from the soil, with continuous cropping obstacles [4]. In contrast, soilless culture (hydroponics, substrate, and aeroponics) effectively alleviates the problem of continuous soil cropping and the problem of heavy-metal residues caused by soil pollution, while reducing water and fertilizer use; thus, it has emerged as a new technology for rapid development [5].

Aeroponics is a soilless cultivation system in which a fresh nutrient solution is intermittently or continuously misted to the roots of the plants to maintain plant growth over a long period while maintaining a constantly updated nutrient solution [6]. Aeroponics frees plants from soil and water and provides a good root water, air, and fertilizer environment for plant growth with an atomized nutrient solution, which fundamentally solves the contradiction between plant roots absorbing nutrients from soil or solution and oxygen supply, improves the problem of plant roots lacking oxygen, and increases the utilization rate of land [7].

Aeroponics rapid propagation technology has quickly developed in recent years, becoming a new means of seedling breeding [8]. Aeroponics rapid propagation is a system in which plant organs or cuttings are placed in contact with a nutrient solution through evenly distributed nozzles and held to deliver nutrients to propagate plants. Aeroponics rapid propagation incorporates asexual reproduction, in which the good traits of the mother tree are passed to intact offspring. This process makes seedling breeding no longer dependent only on high-quality seeds, which enables the plants to grow and rapidly develop within a short period and greatly shortens the breeding cycle of seedlings. Aeroponics rapid propagation technology regulates plant growth by controlling the root zone environment; thus, the automated control and management of the root zone environment are particularly important [9]. In the process of aeroponics rapid propagation, the temperature and humidity in the root zone of plants are the most influential factors in the rapid growth of plants. Excessive temperature and humidity in the root zone may cause mildew or necrosis of the plant roots. Therefore, identifying a rapid and accurate method to detect plant root health during aeroponics rapid propagation of seedlings has become a critical problem to be solved.

Many previous studies considered image recognition and employed a specific classifier to classify images into healthy or diseased images [10]. The authors of [11] presented different techniques for the early identification and classification of plant diseases; the main techniques were support vector machine (SVM), k-means clustering (K-MC), deep learning (DL), and k-nearest neighbor (K-NN). The authors of [12] developed a convolutional neural network model for plant disease detection and diagnosis using images of healthy and diseased tomato leaves via a deep learning approach. A study by [13] explored the application of deep convolutional neural networks in classifying rice health status based on leaf images. Three classifiers representing normal, unhealthy, and snail-infested plants were designed using transfer learning of the deep network, which obtained 91.23% accuracy using stochastic gradient descent. The authors of [14] employed a fully localized binary model for apple fruit disease detection; the proposed method included a k-means clustering algorithm for feature extraction and a multiclass support vector machine to classify the images with a classification accuracy of 93%. The authors of [15] proposed a system for automatically diagnosing cotton leaf diseases using a simple image processing method, selecting appropriate color and image texture features, and using an SVM classifier to classify cotton leaf diseases. Due to the complexity of image backgrounds, research on the automatic recognition of plant disease images still needs a breakthrough. Recently, neural networks, especially BP neural networks, have been widely employed and have good learning abilities. The authors of [16] designed an image feature recognition system for cucumber foliage and used BP neural network for image feature extraction and pattern recognition of cucumber fungus foliage, which achieved a correct recognition rate greater than 85%. The authors of [17] applied digital image processing and a BP neural network to classify and detect four diseases of pomegranate plants with a maximum accuracy of 90%. The authors of [18] designed and developed a Yali pear quality inspection system based on image features and a BP neural network. This system mainly classifies Yali pears as a function of four image features: size, shape, color, and surface defects. After neural network training, the detection results obtained by this detection method were compared with the human detection results, and the matching rate reached 90%.

With economic and social development, it has been recognized that mulberry is a tree species with economic and ecological benefits, in addition to the traditional fodder for sericulture [19]. Due to the good adaptability and robustness of BP neural networks [20], it is commonly used in the fields of pattern recognition, data classification, and predictive analysis [21]. It has been shown that BP neural networks can provide support for mulberry cuttings disease identification. However, improper feature selection and unstable network structure can affect the model performance. Therefore, on the basis of the above research background, this paper extracts the texture and color features of mulberry cuttings have and proposes the application of image processing and a BP neural network to identify mildew disease in the root zone of mulberry cuttings according to visual appearance. The model was trained and tested using a dataset including all stages of cuttings growth. On the basis of the recognition results, the mold rate was calculated, and then the spray frequency and the environmental humidity during the aeroponics rapid propagation process were controlled to reduce mold and root rot.

The remainder of the paper is divided into four main sections. Section 2 presents the design of an automatic aeroponics rapid propagation system for controlling the temperature and humidity, the image processing methods, and the establishment of the BP neutral network recognition model. Section 3 includes experimental results and their analysis, while the results are discussed in Section 4. Lastly, Section 5 presents the conclusion of the experimental study and some useful recommendations for future study.

2. Materials and Methods

2.1. Construction of a Temperature and Humidity Control System for Aeroponics Rapid Propagation

2.1.1. General System Design

The temperature and humidity system in this paper consisted of six parts: data acquisition module, data processing module, human–machine interaction module, data storage module, actuator module, and data transmission module. The data acquisition module consisted of a No. 1 lower computer and a temperature and humidity sensor. Every 0.5 s, the No. 1 lower computer sent a signal to the temperature and humidity sensor to read the temperature and humidity data, and then sent the temperature and humidity data packets to STM32 using serial communication. The data processing was performed by Synchronous Transport Mode-32 (STM32). When the mildew rate was within the normal range, the temperature and humidity data sent from the serial port by the No. 1 lower computer were sent to the message queue in uCOS-III for calculation and data processing, and the Emergency Managers Weather Information Network (EMWIN) task displayed the temperature and humidity in real time by reading the data in the message queue and drawing the temperature and humidity image. When the mildew rate was higher than the threshold, STM32 sent control command packets to the actuator module to control the spraying frequency of the atomizing sheet and fresh water from the liquid supply pump. The human–machine interaction module consisted of a thin-film transistor liquid crystal display (TFT-LCD) capacitive screen and a touch acquisition task in uCOS-III. The screen captured the user’s touch operation and executed the corresponding command. The data storage module consisted of STM32 FLASH, a Secure Digital (SD) card, and the FatFs file management system. After starting the SD card storage, the system stored the collected temperature and humidity values in the dataTH.txt file on the SD card. Note that the last stored data are eliminated each time the SD card is started. The system’s text library, temperature and humidity thresholds, mildew rate thresholds, spray time, and spray interval data were stored in STM32 FLASH. The actuator module consisted of a No. 2 lower computer relay, cooling fan, liquid supply pump, fogging tablet, alarm, light-emitting diode (LED) light board, and power supply. STC89C52 received the actuator control data packet sent by STM32 through a serial port and executed the corresponding action according to the content of the data packet. The data transmission module adopted the ZigBee serial transmission module, through which serial communication could be upgraded to wireless communication. The control system architecture is shown in Figure 1.

The above system used one STM32 as the upper computer and two STC89C52s as the lower computers. One STC89C52 was used to receive sensor data and send them to STM32 through the serial port, while the other STC89C52 was used to receive data sent from STM32 through the serial port to control the actuator. In the next section, the STC89C52 used to collect temperature and humidity is referred to as the No. 1 lower computer, and the STC89C52 used to receive data and control the relay is referred to as the No. 2 lower computer. The hardware system built in this paper is shown in Figure 2.

2.1.2. System Control Function

This control system was realized in automatic and manual control modes. The device was applied in the greenhouse; accordingly, there was no heater in the system, and the minimum temperature was ensured through the greenhouse. The functions of the control system of the aeroponics rapid propagation are described below.

Automatic control mode. When the mildew rate was within the normal range, the control system compared the current collected temperature and humidity values with the set temperature and humidity threshold. The control system automatically controlled the switch of the cooler and the atomizing sheet to adjust the temperature and humidity so that the temperature and humidity in the device were within the set range. The LED fill light was always on in the automatic control mode. The cooling fan was turned on when the temperature was higher than the upper limit of the set temperature to blow cold air into the device for the cooling effect and turned off when the temperature reached the middle value of the set threshold. The alarm was turned on when the temperature fell below the set temperature lower limit to alert the staff in the greenhouse that the temperature was too low. When the humidity was lower than the set lower limit of humidity, the fogging piece was opened until the humidity reached the set upper limit before the atomizing sheet was closed. When the mildew rate was higher than the threshold, the control system automatically controlled the switch of the cooler and the atomizing sheet to reduce the temperature, humidity, and spraying frequency and to control the liquid supply pump to supply fresh water.

The control system can also be manually controlled. The operator can manually control the switch of each actuator of the system. In manual mode, all actuators are turned off first. The actuators that can be controlled include the liquid supply pump, cooling fan, fill light, and atomizing sheet. On the basis of the mildew rate, manual control of the supply pump switch can replenish the nutrient tank with nutrient solution. Manual control of the cooling fan can cool the system. Manual control of the fill light can provide light to the plants. Manual control of the atomizing sheet requires inputting two parameters, namely, spray time and a spray interval of 1 min. For example, for a spray time of 2 and a spray interval of 1, the system sprays for 2 min, stops spraying for 1 min, and then continues this cycle.

2.2. Aeroponics Rapid Propagation Experiment and Data Collection

The aeroponics rapid propagation experiment in this study was conducted in the laboratory of Jiangsu University, Zhenjiang, Jiangsu, China, as shown in Figure 3. Initially, 1–2 year old healthy, disease-free mulberry cuttings with strong metabolism, vitality, and fast rooting were selected as experimental samples. The lower and middle parts of the cuttings were selected for cutting. Each cutting was kept with 2–4 buds, and the cutting was 15 cm to 20 cm long. When cut, the upper cut was cut flat, 1 cm to 1.5 cm from the bud, and the lower cut was cut diagonally. The cut surface needed to be smooth. The cuttings were disinfected with potassium permanganate solution for 30 min, retrieved, and dried by rinsing with water. Then, rooting powder was selected for a soaking treatment to promote rooting. The solution was dipped to approximately 3 cm above the base and soaked for 2 to 3 h. After the cuttings were soaked, they were taken out to remove the epidermal moisture and rooted. The cuttings were moved into the atomizing sheet incubator, composed of four layers of structure, namely, the fill light layer, plant layer, atomizing sheet layer, and nutrient solution layer, each of which had corresponding functions. The fill light layer was used to install the fill light, the plant layer was used to fix the plants, the atomizing sheet layer was used to install the atomizing sheet, and the nutrient solution layer was utilized to store the nutrient solution. After 10 to 15 days of the experiment, large amounts of callus and adventitious roots formed on the roots of the cuttings.

Image data were collected on mulberry cuttings’ root growth during the rapid propagation process of aeroponics. In this study, the researchers used cell phone devices to take photos of the root growth of the mulberry cuttings at regular intervals to collect images of the cuttings’ health and disease at various root growth stages to establish a database, as shown in Figure 4.

The database included eight healthy and eight unhealthy images, for 16 images of mulberry roots. These images were taken under background conditions of uneven illumination intensity. All cropped images collected in this paper were classified into disease and health categories and saved in png format. The diseases mainly consisted of mildew and decay.

2.3. Image Preprocessing and Mildew Characteristics

The main purpose of image preprocessing was to eliminate irrelevant information in the images, recover useful and true information, enhance the detectability of relevant information, and maximize the simplification of data, thus improving the reliability of feature extraction, matching, and recognition. The images in the database were first cropped; then, the mulberry root photos were cropped to a uniform size, and the images were enhanced to reduce the background interference information of the images, as shown in Figure 5.

Mildew on mulberry roots was characterized by white mildew on the surface of the plant roots, as shown in Figure 6. If the mildewed cuttings were not treated in time, the mildew would spread and affect the healthy cuttings adjacent to them; thus, the health condition of the plants should be quickly and accurately detected and treated.

2.4. Feature Extraction

The accuracy of recognition results is difficult to guarantee if the features are not properly selected or do not contain enough information about the attributes of the object to be recognized. Therefore, before designing a neural network classifier, quickly and accurately extracting the effective information of the object to be recognized is the key to pattern recognition. This study focused on feature extraction from two aspects—texture features and color features—for root samples of mulberry cuttings in aeroponics rapid propagation trials.

2.4.1. Texture Feature Extraction Based on the Gray-Level Co-Occurrence Matrix

Texture features reflect the object’s properties and help distinguish two objects (or two images). A texture is some variation in the gray level or color of an image pixel point, recurring texture primitives and their arrangement rules, and this variation is spatially statistically relevant.

The gray-level co-occurrence matrix reflects comprehensive information about the image grayscale about the direction, adjacent interval, change magnitude, etc. The matrix is the basis for analyzing the image’s local characteristics and arrangement pattern. The gray-level co-occurrence matrix reflects comprehensive information about the image grayscale about the direction, adjacent interval, change magnitude, etc. The matrix is the basis for analyzing the local characteristics and arrangement law of the image. In essence, starting from the pixel with grayscale i in the image (whose position is (x, y)), the number of times P(i, j, d, θ) simultaneously appears with its pixel (x + Dx, y + Dy) with grayscale j at a distance d is counted. The mathematical expression is presented as follows:

$$\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right)=\left\{\right[\left(\mathrm{x},\mathrm{y}\right),\left(\mathrm{x}+\mathrm{Dx},\mathrm{y}+\mathrm{Dy}|\mathrm{f}\left(\mathrm{x},\mathrm{y}\right)=\mathrm{i},\mathrm{f}\left(\mathrm{x}+\mathrm{Dx},\mathrm{y}+\mathrm{Dy}\right)=\mathrm{j}\right]\},$$

(1)

where x and y are the pixel coordinates in the image, i and j are the gray levels of the pixels in the image, Dx and Dy are the position offsets, d is the step size for generating the gray cogeneration matrix, and θ is the generation direction.

In this study, MATLAB software was used to conduct texture feature extraction on the basis of a gray-level co-occurrence matrix. The processing included, first, converting each color component into grayscale; then, to reduce the computation, the grayscale level of the original image was compressed. The grayscale value was quantized to 16 levels. Second, four grayscale cooccurrence matrices P were calculated. In this paper, the grayscale cooccurrence matrices in the four directions of θ = 0°, 45°, 90°, and 135° were calculated, the step size d = 1 was set, and the calculated grayscale cooccurrence matrices P were normalized.

The delineation of image texture features is expressed by a series of feature quantities calculated from the gray-level co-occurrence matrix; therefore, this paper used the matrix to calculate five texture parameters—energy, entropy, inertia, correlation, and inverse difference—which are calculated as follows [22]:

$$\mathrm{ASM}={\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{L}}{\displaystyle \sum }_{\mathrm{j}=1}^{\mathrm{L}}{[\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right)]}^2,$$

(2)

$$\mathrm{ENT}=-{\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{L}}{\displaystyle \sum }_{\mathrm{j}=1}^{\mathrm{L}}\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right)·{\log }_2\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right),$$

(3)

$$\mathrm{CON}={\displaystyle \sum }_{\mathrm{i}=1}^{\mathrm{L}}{\displaystyle \sum }_{\mathrm{j}=1}^{\mathrm{L}}{\left(\mathrm{i}-\mathrm{j}\right)}^2\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right),$$

(4)

$$\mathrm{CORR}=\frac{\left[{\sum }_{\mathrm{i}=1}^{\mathrm{L}}{\sum }_{\mathrm{j}=1}^{\mathrm{L}}\left(\mathrm{i}\mathrm{j}\right)\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right)-{\mathsf{\mu }}_{\mathrm{x}}{\mathsf{\mu }}_{\mathrm{y}}\right]}{{\mathsf{\sigma }}_{\mathrm{x}}{\mathsf{\sigma }}_{\mathrm{y}}},$$

(5)

$$\mathrm{IDM}=\frac{{\sum }_{\mathrm{i}=1}^{\mathrm{L}}{\sum }_{\mathrm{j}=1}^{\mathrm{L}}\mathrm{P}\left(\mathrm{i},\mathrm{j},\mathrm{d},\mathsf{\theta }\right)}{\left[1+{\left(\mathrm{i}-\mathrm{j}\right)}^2\right]},$$

(6)

where P (i, j, d, θ) is the cooccurrence matrix, L is the width of the cooccurrence matrix, $\mathsf{\mu }$ is the mean of the cooccurrence matrix, ${\mathsf{\mu }}_{\mathrm{x}},{\mathsf{\mu }}_{\mathrm{y}}$ are the means of P(i) and P(j), respectively, and ${\mathsf{\sigma }}_{\mathrm{x}}$ and ${\mathsf{\sigma }}_{\mathrm{x}}$ are the variances of P(i) and P(j), respectively.

2.4.2. Color Feature Extraction

One of the main characteristics of mildew lesions on the roots of mulberry cuttings during the aeroponics rapid propagation process was its color change; the mildew was mainly white, while the roots of healthy mulberry cuttings were mainly dark green. To scientifically measure, study, and describe color, several color models have been established. In this study, the red (R), green (G), and blue (B) and hue saturation luminance (HSL) systems were mainly utilized to describe the color characteristics of the images of the diseased parts of mulberry roots.

RGB color space: The color variation presented by different diseased parts of the roots of mulberry cuttings can greatly vary. Therefore, color was analyzed as a major feature in distinguishing root diseases of mulberry. In a normal-color image, light intensity at the red, green, and blue wavelengths was recorded at each pixel point. R, G, and B were used to represent the information of their colors, known as RGB model representation. In this study, all images of mulberry roots collected by mobile devices were color images composed of different R, G, and B components. The images of mulberry root molds were analyzed by MATLAB software. The histograms of each component of R, G, and B are shown in Figure 7.

HSL color space: In the RGB chromaticity system, a color must be represented by the three RGB components, and its spatial dimension involves three dimensions. In contrast, in the HSL system, H (hue) represents the range of colors. By converting the color information from the RGB system to the HSL system and the grayscale values of the RGB triads to hue values, the color characteristics are reduced from three dimensions to one dimension, without losing the color information [23].

In the aeroponics rapid propagation box, the photos taken by the researchers were mainly disturbed by the light and background information in the laboratory. Since the three components R, G and B in the RGB color space are greatly influenced by light, the image noise was large, thus affecting the accuracy of root disease feature extraction of mulberry cuttings. To improve the accuracy of color feature extraction, the RGB color space was converted to the HSL color space, where the influence of the luminance component could be eliminated from the color information carried by the color images. The HSL system directly used three quantities in the sense of color characteristics—luminance, hue, and saturation—to describe color. The conversion formulas from the RGB coordinate system to the HSL coordinate system are expressed as follows [24]:

$$\mathsf{\theta }=\arccos \left\{\frac{\frac{1}{2\left[\left(\mathrm{R}-\mathrm{G}\right)+\left(\mathrm{R}-\mathrm{B}\right)\right]}}{{[{\left(\mathrm{R}-\mathrm{G}\right)}^2+\left(\mathrm{R}-\mathrm{G}\right)\left(\mathrm{G}-\mathrm{B}\right)]}^{\frac{1}{2}}}\right\},$$

(7)

$$\mathrm{S}=1-\frac{3}{\mathrm{R}+\mathrm{G}+\mathrm{B}}\min \left(\mathrm{R},\mathrm{G},\mathrm{B}\right),$$

(8)

$$\mathrm{L}=\frac{\left(\mathrm{R}+\mathrm{G}+\mathrm{B}\right)}{3},$$

(9)

$$\mathrm{H}=\{\begin{array}{ll}\mathsf{\theta }& \mathrm{G}\ge \mathrm{B}\\ 360-\mathsf{\theta }& \mathrm{G}<\mathrm{B}\end{array},$$

(10)

where L is the luminance, H is the hue, and S is the saturation.

After the color space transformation, the HSL component map of the disease image was as shown in Figure 8. It is obvious that the image mold spots, and background environment had the greatest difference, and that the L and S components effectively suppressed the effect of uneven light intensity on the image. The H component image was affected by noise and the light environment; thus, the L and S components were selected as a feature value for image classification.

After observation and comparison, it was determined that the R, S, and L components of image lesion site features were more obvious, that the background area was uniform and less affected by light and noise, and that the boundary between the mold and the background was distinct. Therefore, this study chose R, S, and L as the initial color feature quantities.

2.5. BP Neural Network Recognition Modeling

The BP neural network is a typical supervised neural network classifier that achieves a nonlinear mapping function from input to output to extract reasonable solution rules via learning automatically; thus, it has a certain generalization ability. Considering the characteristics of the BP neural network, this paper selected the BP neural network as a classifier and a three-layer network structure, which included an input layer, an output layer, and a hidden layer.

Backward Propagation (BP) Neural Network Structure Design

Input layer: The color and texture features of the mulberry cuttings in the image were extracted as the main basis for detection, and Tansig was employed as the input layer transfer function. Because of the large number of dimensions of the input feature matrix, the computation was large, which generated a long training time for the model. A dimensionality reduction process was carried out, whereby the feature matrix was reduced using principal component analysis (PCA). The features obtained after dimensionality reduction significantly improved the training speed of the neural network recognition model.

Output layer: According to the growth conditions of the cuttings, they were classified into two categories: healthy and moldy. Therefore, the output layer of the neural network was designed with two output layer nodes, in which the healthy samples were set to 1 and the moldy samples were set to 2. Purelin was applied as the output layer transfer function.

Hidden layer: The hidden layer was designed between the input layer and the output layer. If the number of neurons in the hidden layer is too high, the overall network may be overfitted, but if the number of hidden layers is too low, the training result of the network may not reach the desired goal [24]. The sigmoid-type function served as the function of the hidden layer function. After evaluating different numbers of hidden layer neurons, the stability and training speed of the model were compared, and the number of hidden layer neurons was set to eight in this study, thus forming a neural network with an 8–8–2 structure, as shown in Figure 9.

3. Results

In this experiment, identification tests and model validation were conducted for mulberry root mold disease identification using MATLAB R2022a software.

3.1. Identification of Diseases in the Root Zone of Mulberry by a BP Neural Network

The dataset was divided into a training set and a test set at a ratio of 11:5. Sixteen images of healthy or diseased mulberry roots, including eight healthy images and eight diseased images, were collected at each growth stage in the aeroponics rapid propagation experiment. According to the division ratio, 11 mulberry sample images were obtained for training the BP neural network and five images were used for testing the neural network. To improve the recognition accuracy of healthy and diseased cuttings, the texture and color feature values were combined because of their different orders of magnitude and range of values. Their data were normalized and employed as the BP neural network input feature variables. The BP neural network was trained on the input samples. During the training process, the weights and thresholds of the network were continuously adjusted and modified by backpropagating the network until the total error was less than a given value. Then, the training was considered complete, the final weights and thresholds were determined, the final test set was used to test the neural network, and the test results were obtained. The flowchart of neural network recognition is shown in Figure 10. The BP network parameters were configured as shown in

Table 1. BP network parameters.

Network Parameters Configuration	Parameters
Number of training sessions	1000
Learning rate	0.01
Minimum error of training target	0.0001
Display frequency	Every 25 times
Momentum factor	0.01
Minimum performance gradient	0.000001
Maximum number of failures	6

.

The performance of the neural network is shown in Figure 11A; one forward propagation of the neural network and one backward propagation of the neural network comprised one round. The model showed the lowest mean square error in the validation set after three training rounds; hence, the model of the third round was chosen as the best model. Figure 11B shows the classification images of the predicted and actual kinds of the model, Figure 11C shows the classification error graph of the test set, and Figure 11D shows the histogram of the error distribution. The training effect of the model reveals that the BP neural network model in this study had good stability and could accurately identify diseases in the root zone of mulberry cuttings.

3.2. Evaluation and Validation of the Model

As shown in Figure 12A, for the regression results after using MATLAB for neural network training, we obtained the correlation between the original target value and the predicted value, as well as whether the model was overfitted. The correlation coefficient R was selected to analyze the results. An R-value closer to 1 means that the degree of linearization is higher, and that the fitting effect is better, indicating better results. As shown in Figure 12A, R reached 0.98 for the fitting curve of the training dataset, 0.98 for the fitting curve of the test set, and 0.99 for the fitting curve of the validation set, indicating that the prediction results aligned well with the actual results. Figure 12B represents the training state diagram of the BP neural network. The vertical coordinate was the gradient in order; a gradient of 0 denotes the lowest point of the image, which is the optimal position. Mu is a hyperparameter, which is utilized to modulate the neural network weights to avoid falling into a local minimum during the training of the BP network. Mu ranges from 0 to 1. The value of mu decreased with time, which means that the model’s performance improved with time. The error value represents the lowest mean square error by the third round of training. As the network was trained with the training samples, it was confirmed that the error curve of the samples no longer decreased for three consecutive iterations. If the training error did not decrease after three consecutive training sessions, it was assumed that the effect would not improve if the training were to be continued; thus, the training was stopped.

The performance evaluation metrics of a model can also usually be expressed in terms of a confusion matrix, where the parameters of the confusion matrix represent the model’s performance. A confusion matrix is a matrix that categorizes classification problems according to two dimensions—the true case and the discriminant case—and comprises a summary of the prediction results for the classification problem, using count values to summarize the number of correct and incorrect predictions which can be broken down by each class. A 2 × 2 confusion matrix for the neural network model of this study is shown in Figure 12C. Five samples were randomly selected from the database for class prediction, which contained two healthy samples and three samples with mildew disease. All samples were correctly predicted except one sample with mildew disease that was misclassified, and the prediction results showed that the class prediction accuracy of all mulberry disease images was 80%. Thus, the BP neural network had a high accuracy for the identification of mulberry mildew disease, which shows the effectiveness of the proposed method. The model performance was validated using four metrics—recall, accuracy, precision, and the value of F1—calculated as follows:

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}\times 100\%.$$

(11)

$$\mathrm{Accuracy}=\frac{\mathrm{TP}+\mathrm{TN}}{\mathrm{TP}+\mathrm{TN}+\mathrm{FP}+\mathrm{FN}}\times 100\%.$$

(12)

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}\times 100\%.$$

(13)

$$\mathrm{F}1-\mathrm{Score}=\frac{2\mathrm{Precision}·\mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}.$$

(14)

The recall is the proportion of correct model predictions among all outcomes for which the true value is positive. Accuracy is the proportion of all correct judgments of the classification model to the total number of observations. Precision is expressed as the proportion of correct model predictions among all outcomes for which the model prediction is positive. The F1-score integrates the results of precision and recall outputs. The F1-score values range from 0 to 1, with 1 representing the best output of the model and 0 representing the worst output of the model, where TP, FP, TN, and FN represent the number of true positives, false positives, true negatives, and false negatives, respectively. The corresponding evaluation metrics of the confusion matrix are presented in

Table 2. Results of the detection of mulberry diseases by the described method.

Evaluation Metrics	Recall	Accuracy	Precision	F1
Test Results	100%	80%	66.67%	0.8

, where the accuracy of the model identification reached 80%, the precision reached 66.67%, the recall was 100%, and the F1 value was 0.8.

4. Discussion

The different features selected for different diseases can affect the accuracy of the model identification. On the basis of the literature reviewed, very few researchers have conducted research on automatic aeroponics rapid propagation to identify the mildew disease of cuttings at the base of mulberry cuttings based on a BP neural network. Our results are in line with those of Zhou et al. [25], who proposed a multi-feature-based machine learning method to identify and evaluate the black spot disease incidence by selecting texture and shape features of black spot regions and healthy regions in order to construct a feature matrix, which finally achieved the correct classification of each pixel in the test sample with an accuracy of 85%. Another study [26] concluded that a cucumber region detection method based on a multipath convolutional neural network combined with color component selection and support vector machine could convert cucumber images into color space and extract three color component features using a multipath convolutional neural network structure with 90% recognition accuracy. The authors of [23] studied six cucumber diseases for recognition, and the processed disease images were extracted with features from the grayscale statistics, color, and geometry, and the average accuracy of recognition reached 93.3%. The above comparative analysis showed that, for different disease features, the selection has a great influence on the accuracy of the identification model.

Many researchers have studied crop identification and classification using BP neural network models. The results of the researchers are more or less similar to our results. Zhang et al. [27] proposed a cotton disease identification method based on rough sets and BP neural networks under natural environmental conditions, and the average cotton disease identification rate reached 92.72% accuracy. Zhang et al. [28] developed a wheat water requirement prediction model using an optimization algorithm BP neural network. The above study showed that BP neural networks have high accuracy in plant disease identification and prediction.

In this study, we extracted texture and color features from mulberry cuttings; in order to distinguish the extracted features, they were identified for classification using the constructed BP neural network model. The results showed that this study’s BP neural network classification method could accurately identify mulberry cutting diseases during aeroponics rapid propagation.

5. Conclusions

This study demostrated the use of image processing and a BP neural network to identify mildew diseases on the roots of mulberry cuttings in the aeroponics rapid propagation process, to control the ambient temperature and humidity of the whole aeroponics rapid propagation incubator according to the mold rate, and then to improve the rooting rate and survival rate in the aeroponics rapid propagation process. It can be determined from the results that the neural network model had good performance in identifying the health condition of mulberry cuttings in the aeroponics rapid propagation process and could accurately identify mulberry cuttings suffering from mildew disease with an accuracy of 80%. In future research, we will use the above research methods to simulate human decision making to accurately identify various types of diseases that occur in aeroponics rapid propagation to monitor and identify plant diseases automatically, as well as treat them on time. With the increase in the number of aeroponics rapid propagation seedlings, it is becoming time-consuming and labor-intensive to take photos and record each cutting; therefore, this study did not consider the problem of automating the photos and performing recognition, which becomes a key problem to be addressed in future research. The feature extraction method in this study can be further improved and optimized, and it can be used in combination with other algorithms. In the future, we need to continue to enlarge our database to further identify plant diseases. The system’s accuracy will increase as more data become available.