CNN Architecture for Classifying Types of Mango Based on Leaf Images

ABSTRACT


INTRODUCTION
Processing the digital image, vision techniques computer, machine learning algorithms, and deep learning are increasingly being developed due to handling complex data and good precision results (Chouhan et al., 2019). The stages in developing a computer-based automation system (Yamparala et al., 2020) based on images need to be processed and segmented images (Razi, 2012), then learning to find out the pattern of an image (Chouhan et al., 2019) (Prasad et al., 2019(Prasad et al., ) al., 2016. The goal of developing an automation system in agriculture is to help the agricultural team, as well as maximum agrarian output, and facilitate more efficient work (Yamparala et al., 2020) (Ranjan et al., 2015) (Arivazhagan et al., 2013 ( Samajpati & Degadwala, 2016) (Mishra et al., 2021) (Zarrin & Islam, 2019). Not everyone knows the ins and outs of agriculture because not everyone studies in that field. Therefore, people in agriculture are trying to develop a system that can classify types of plants, identify plant diseases. The hope is that they can help others who do not study in agriculture to take good care of plants and help facilitate the work of caring for plants (Yamparala et al., 2020) (Aakif & Khan, 2015 (Prasetyo, 2016) (Arivazhagan et al., 2013 ) (Samajpati & Degadwala, 2016) (Mishra et al., 2021). Farmers who work in the agricultural sector also 113 http://dx.doi.org/10.35671/telematika.v14i2.1262 sometimes have difficulty identifying the type of plant or knowing the kind of plant disease because it does not have experience. In such conditions, a system that can automatically classify plant types or identify types of plant diseases is needed using machine learning (Yamparala et al., 2020) (Ranjan et al., 2015) (Prasetyo, 2016) (Arivazhagan et al., 2013 ) (Rumpf et al., 2010) (Dutta & Basu, 2013) (Samajpati & Degadwala, 2016) (Mishra et al., 2021) (Zarrin & Islam, 2019) (Mia et al., 2020) (Srunitha & Bharathi, 2018) (Prasad et al., 2016) (Madiwalar & Wyawahare, 2017) (Kaur et al., 2019)  (Nafi'iyah, 2020). The purpose of this study is to propose a CNN architectural model in classifying mango species based on leaf imagery. Table 1 describes the distribution of the dataset for training and testing. Table 1 describes the types of mangoes divided into 5, namely Alphanso, Amarpali, Ambika, Austin, Kent. The dataset is taken from Kaggle, and only five types of mango are used. Figure 1 is an example of a leaf image.  The dataset used is an image of colored mango tree leaves with a size of 224 x 224, as shown in Figure   1.

Proposed Research
This study proposes a CNN architectural model to classify the types of mangoes based on leaf images.

Convolution Neural Network
The CNN architectural model proposed in the study is 5; the model is shown in Table 2. Table 2 shows the CNN model made by explaining the number of parameters.  Each architecture from the first model to the fourth model in Table 2 is described in Figure 3, Figure 4,

RESULTS AND DISCUSSION
This study proposes 4 CNN architectural models and uses InceptionV3 transfer learning. Each architectural model differs in the number of layers and nodes in the convolution layer, the max-pooling layer. The number of parameters of each CNN architectural model is different. The most significant number of parameters in the InceptionV3 learning transfer model. Simultaneously, the most considerable number of proposed model parameters is the first model in Figure 3. Each proposed architectural model is trained and evaluated, as many as 50 epochs of training are carried out. We understand that the number of datasets used is minimal, namely 24 of each type of mango shown in Table 1, so we also augment the training.
When we did our training, we used error testing with categorical loss cross-entropy. Equation 1 Table 3. The purpose of calculating the error value for each iteration in training is to improve each node's weight value in the convolution and max-pooling layers. The explanation of Equation 1 is that y is the actual data, while the algorithm's output data ŷ . Table 3 Figure 9 represents the performance of the loss value and the accuracy of the third CNN architectural model. Figure 10 illustrates the implementation of the loss value and the accuracy of the fourth CNN architectural model. Figure 11 describes the performance of the loss value and accuracy of the It can conclude that the third architectural model has the best performance compared to the other four models.

ACKNOWLEDGEMENT
Thank you to all those who helped correct and publish this article, especially to the Lamongan Islamic University.