A Tailored Convolutional Neural Network Architecture with Hyperparameter Tuning for Grapevine Variety Identification

Digital Agriculture (DA) in recent years exponential rise in the total of data created by agricultural go per hectare. In precision viticulture and site-specific management in viticulture have expanded the information available to winegrowers [1]. The industry of grape economic and social importance has fostered the advancement of cutting-edge technologies for the live tracking of vineyards, aiming to enhance the quality of the fruit. As an effect, there is need to do accurate techniques for find and verifying grape varieties, specifically when it comes to selling vine cuttings to vineyards, detecting forbidden species or varieties in specific regions, or estimating the market value of different grape varieties in various production areas [2].

The development of an automatic algorithm for identifying grape varieties based on leaves faces many challenges. Ampelography is the specific discipline of botany that deals with identifying and differentiating different grape varieties [3, 4]. The factors that have been briefly outlined share a common requirement: the presence of a skilled specialist at the grape picking location, who must directly deal with the actual fruit. The agricultural regions are predicted to be substantial, but providing a constant presence of experts is impossible. Hence, it is desirable to automate the process of grape detection by utilizing drones equipped with cameras, for instance. Nevertheless, drones lack the intuitive abilities of professionals and not equipped with organoleptic sensors. Performing a DNA test is the most dependable method for identifying a grape variety. However, it is also challenging to envision integrating a DNA test kit into agricultural equipment [5]. Nonetheless, the utilization of specialist software necessitates the participation of a highly trained expert [6]. The efficacy of this approach is heavily contingent upon the proficiency and expertise of the specialist. A more objective approach, such as DNA analysis [7], is available. However, this process, along with other wet chemical techniques, is damaging, time-consuming, labor-intensive, and necessitates the expertise of a specialist. Recent advances in computer vision algorithms so the availability of new imaging sensors that led to the automated identification of grape varieties. The processing of data from a spectrometer [8, 9] or a high-spectral camera is automated non- invasive and rapid. Accurate identification of grapevine types can be achieved by measuring the interaction between electro- magnetic radiation and matter across various spectral bands. An inherent drawback of this method is the substantial cost of acquiring a spectrometer nor hyperspectral camera, that be significantly greater by many orders of magnitude when compared to a standard camera. Incorporating these sensors into a specialized harvester would substantially raise its cost. Human sensory capabilities constrain conventional approaches for identifying grapevine varietals. For instance, amperometry uses visual perception to identify different grapevine varieties. The classification of grapevines based on their kinds using images is a challenging operation that necessitates the extraction of numerous distinguishing traits. The wide range of elements in an out- side environment adds to the difficulty of extracting specific features [10]. In this paper the dataset that use has eight folders include 8000 images in eight categories (1000 images per category) each representing a specific grape variety, with every file containing 1000 images in JPG format. All the pictures in the collection have a 6000 × 4000 pixels resolution Figure 1 [11].

1.png

Figure 1. Examples of each grape classes [11]

Grapes have various physical, chemical, and geographical properties that affect their health. Moreover, different varieties of the same plants in the same region can also have different yields depending on growth stages or grape quality. Consequently, various stages of growth require different monitoring. Therefore, it is essential to distinguish between wine varieties produced in the same region, and the relationship between varieties and the development of certain substances is a crucial factor to be recognized. This paper aims to detect and analyze several varieties of vineyards by their spectral properties. The results of this correlation will lead to the creation of a vineyard map. Knowing the specific locations of each variety helps farmers avoid errors and ensures the quality of each variety. The objective is to implement the classification methods of varieties with common characteristics to ensure efficient management of different varieties in the same region. In addition to visible (RGB), multispectral images include red edges and near- infrared regions of the spectrum, which are not always visible to the naked eye. The scope of this methodology is to inform farmers about the specific characteristics of grape varieties; this information can be invaluable to vineyard management when planning, cutting, fertilizing, and harvesting during its lifetime. For this task, we proposed the CNN , and the hyperparameters were tuned for every classification algorithm. A change was made to the hyperparameter to verify its accuracy CNN consists of convolutional, pooling, and fully connected layers. The purpose of these layers in the CNN is to generate complex features that enhance classification effectiveness. The convolutional layers in this study comprise a collection of feature maps. Contributions of this research are as follows: Grape varieties with similar characteristics can be clustered using a combination of vegetation on the data set. Random splitting dataset in to training, testing and validation to development of robust and valid model. Automatic recognition of grapevine varieties using of deep learning by design the best model for the grapevine varieties automatic identification.

The residue of this paper as follows: Section 2 involves related work that handles the same problem and explain how other works solved it. In Section 3, we present CNN model, and the hyperparameters were tuned for every classification algorithm. Section 4 explains results and discuss the result. Lastly, Section 5 gives the conclusion and future models.

Our research’s primary objective is to accurately identify grape varieties using only the image of the grape being displayed. For this task we proposed the CNN and the hyperparameters were tuned for every classification algorithm. A change was made to the hyperparameter to verify its accuracy. Figure 2 shows the flowchart of this proposed method.

2.png

Figure 2. Flowchart of the proposed methodology

3.1 Dataset

The dataset has eight folders include 8000 images in eight categories (1000 images per category) each representing a specific grape variety, with each folder containing 1000 images in JPG format. All the pictures in the collection have a 6000 × 4000 pixels resolution. Due to the photographs’ high resolution, their size reached 49.8 GB, which is impractical for downloading and uploading the dataset from the Internet. To address this issue, the images were resized using a Windows application called (resize pictures) to the proportions of 1620 × 1080 pixels. The initial data size is 1.95 gigabytes. Compression using a zip application reduced the data size to 1.83 gigabytes. Eight folders represent different grape variety types (Deas Al- Annz, Kamali, Halawani, Thompson Seedless, aswud balad, riasi, frinsi, and shdah).

1. Preprocessing of image: Images are divided into three sets: train, test, and validation sets. Hyperparameter values are selected according to the accuracy obtained from the train set. The image is standardized by split all image into 255 steps. All channel is 8 bits; when divided, the value is 0-1. It is, therefore precise and efficient in computation based on the following Eq. (1):

$X_{ {normalized }}=\frac{X-X_{min }}{X_{max }-X_{min }}$               (1)

2. Training CNN model: The model is trained using a training set as 6400 images, and a high parameter value selection is based on the accuracy achieved by training the model and testing set with 800 images as testing. External validation datasets have been chosen to minimize the result of over-adaptation on the test dataset during validation with 800 images. The utilization of an external validation dataset enhances the reliability of the models’ outcomes for real-world applications, notwithstanding the possibility of lower validation metrics.

3.1 Dataset

a) Batch Size and Selection: Before commencing the training process, batch size is one of the primary hyperparameters that necessitates adjustment. The batch size is a reference to the quantity of images used in the gradient estimation procedure. Several studies have examined the impact of batch size on network performance, specifically in terms of accuracy and convergence time. The objective was to establish whether small or large batches were more advantageous. While a small batch size may converge more quickly than a big batch, the major batch size has the likely to reach best minima that a smaller batch size cannot achieve. Also, a small batch size might induce real regularization due to its elevated variance [21]. In this work on image classification, we empirically utilize a batch size of 32.

b) Convolutional neural networks (CNNs): CNN consists of convolutional, pooling, and fully connected layers. The purpose of these layers in the CNN is to generate complex features that enhance classification effectiveness. The convolutional layers in this study comprise a collection of feature maps. Each feature map has a receptive field that covers only a small part of the input spectra. Each feature map generated results from performing a convolution (dot product) between the weights of the receptive field and all the points in the spectrum. This indicates that adjacent points in the feature map were identified in overlapping and adjoining portions of the input spectra. In this manner, several characteristics are identified across the entire spectra range. One significant benefit of employing convolutional layers is their much- reduced number of trainable parameters compared to a fully connected neural network. The pooling layer is utilized on the feature maps to achieve down-sampling. In this scenario, the feature maps were partitioned into distinct regions that do not overlap, and the highest value within each region was selected. Subsequently, dropout is implemented where training certain elements that contribute to the subsequent layer are either excluded or included with a specific probability. Following the convolution and pooling layers, which consist of one of each in this study. Fully connected neural network analyzes the features derived from the preceding layers. Applies a flattened layer to a Keras model. The flattened layer transforms the input from a multidimensional array to a one-dimensional one. This is frequently performed when converting from a convolutional layer to a fully linked layer.

Pooling: Reducing the size of the map at the final layer of the deep network enhances its suitability for classification tasks we use. Including a pooling layer ensures that slight variations in input images do not affect the output while also decreasing the dimensions of the feature maps, such as width and height. A nonlinear function f() was applied element-wise to c: a for every feature map c. The activations obtained, represented by the letter a, were then sent to the MaxPooling2D layer is standard in convolutional neural networks [22]. It reduces the output’s spatial dimensions, which can enhance the model’s execution by lowering the number of account and parameters [23].

c) Learning rate and Dropout: for regularization the learning rate is a critical parameter in CNN that determines the speed at which a network adjusts its parameters during backpropagation [24]. Reducing the learning rate promotes gradual pool but also hampers the process speed learning. Nevertheless, maintaining a higher learning rate can accelerate the learning process, but it may impede convergence. Dropout is an effective and uncomplicated regularization method for deep learning models [25], and convolutional neural networks (CNNs) often tend to overfit. Co-adaptation is more likely to occur when a fully connected layer contains numerous nodes or neurons [26]. This approach randomly picks neurons and excludes them throughout the training phase eliminating their impact on subsequent processes.

Table 1. The configuration of the model CNN

d) Batch normalization and L2 regularization: Batch normalization is a further regularization technique that standardizes the collection of activations in a layer. Normalization is achieved by subtracting the mean of the batch from each activated and then using the standard deviation and dividing the batch. Normalization, in conjunction with standardization is a commonly used method in preprocessing pixel values. Conventional L2 regularization is applied to all trainable parameters, meaning the regularization factors remain constant during training. Addition is the mathematical operation of combining two or more numbers to find their total sum. Indeed, the selection of those regularization factors must be done manually through hyperparameter optimization [27]. Table 1 shows the configuration of the best model.

Model hyperparameter optimization strategy following the establishment of a CNN backbone architecture and specific hyperparameters learning rate and Dropout rates to achieve the most outstanding performance (accuracy, loss, etc.). These adjustments are necessary to ascertain the optimal model of this CNN to fit the training and validation dataset. Algorithm 1hyperparameter tuning by systematically trying different combinations of learning rates and dropout rates and selecting the best-performing model.

3. Performance metrics

• Confusion matrix: A model confusion matrix derived from a classification model on the train, valid, test data The confusion matrix is contain of four main item where true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN). Gvie in whose the model accurately sign the positive class, corresponding with real positive results express as true positives [28]. Confusion matrix assesses how well a classification model performs when it is applied to a put of test data with known true values [29].
• Accuracy: is a metrical that count the right predicts the outcome of deep learning model. To cast accuracy using dividing the number of correct forecast by the total number of forecast. which is the rate of accurate predictions (TP and TN) to the total number of instances that were investigated, The Eq. (2) can be used to calculate accuracy:

$Accuracy$ $=\frac{T P+T N}{T P+T N+F N+F P}$                (2)

• ROC: is a graphical plot that clarify the show of a binary classifier type (can be used for multi class classification as well) at alter threshold evaluate. The True Positive rate (TP = Sensitivity) is scheme as a function of the False Positive rate (FP = 1 - Specificity) [30], as shown in Figure 3.

3a.png

3b.png

3c.png

Figure 3. ROC classification model

The classification of grape vines using advanced computer vision techniques represents a significant advancement in the grape industry. By automating the identification process and enabling live tracking of vineyards, this technology can revolutionize crop management and improve the quality of grapes. The proposed CNN model, with its impressive training and testing accuracy, demonstrates the feasibility and effectiveness of this approach. As research in this field progresses, we can expect to see even more sophisticated and accurate classification systems that benefit both grape growers and consumers. The present research employs a transfer learning strategy based on the CNN architecture. It was demonstrated to automatically recognize and categorize eight different types of grapes from a provided image dataset. The CNN Classification successfully trained on correctly detected type of grapes, achieving 100% accuracy. With a 60% testing accuracy, the experimental findings showed the suggested classifier's dependability. In the future work, the hyperparameters of the CNN model will be optimized to achieve better classification accuracy and eliminate the high complexity in the models.