MobiRes-Net: A Hybrid Deep Learning Model for Detecting and Classifying Olive Leaf Diseases

Abstract

The Kingdom of Saudi Arabia is considered to be one of the world leaders in olive production accounting for about 6% of the global olive production. Given the fact that 94% of the olive groves are mainly rain-fed using traditional methods of production, the annual olive production is witnessing a noticeable fluctuation which is worse due to infectious diseases and climate change. Thus, early and effective detection of plant diseases is both required and urgent. Most farmers use traditional methods, for example, visual inspection or laboratory examination, to identify plant diseases. Currently, deep learning (DL) techniques have been shown to be useful methods for diagnosing olive leaf diseases and many other fields. In this work, we use a deep feature concatenation (DFC) mechanism to combine features extracted from input images using the two modern pretrained CNN models, i.e., ResNet50 and MobileNet. Hence, we propose MobiRes-Net: A neural network that is a concatenation of the ResNet50 and MobileNet models for overall improvement of prediction capability. To build the dataset used in the study, 5400 olive leaf images were collected from an olive grove using a remote-controlled agricultural unmanned aerial vehicle (UAV) equipped with a camera. The overall performance of the MobiRes-Net model achieved a classification accuracy of 97.08% which showed its superiority over ResNet50 and MobileNet that achieved classification accuracies of 94.86% and 95.63%, respectively.

1. Introduction

Recently, agriculture is undergoing the Fourth Industrial Revolution by integrating new technological advances into traditional farming practices [1]. Smart farming plays a vital role as it offers various innovative solutions to modernize the farming sector. Unfortunately, for farmers and farming in general, the growth of plants is hindered by a variety of plant diseases which result in huge losses [2]. Recent studies by the Food and Agriculture Organization of the United Nations (FAO) have shown that plant diseases result in around 10% loss of global food production. The situation has worsened due to trans-boundary plant diseases that have become increasingly noticeable and which affect agricultural production. To secure the global food supply, there is a need for early detection of diseases, and therefore this has become the ultimate concern of experts around the world [3].

Different studies have shown that agriculture is the leading activity in the economy of many countries. Agricultural products are traded at a large scale. Olives are one of the products that is highly demanded in many countries [4]. The Kingdom of Saudi Arabia is the leading olive-producing country in the South of the Mediterranean, accounting for 6% of the global olive production, which makes the country the second highest producer in the world after the European Union. Half of the country’s agricultural exports come from olive oil which represents around 5.5% of the total exports of the Kingdom of Saudi Arabia. For the Kingdom of Saudi Arabia, the olive oil trade is considered to be the fifth source of foreign currency, which makes it a vital sector for the economy of the country. In fact, significant efforts have been made in the country to restructure and to modernize olive production with the purpose of improving the quality of olive oil and expanding acreage [5].

Recently, olive farmers in the Kingdom of Saudi Arabia have been dealing with many serious diseases, such as olive leaf spot that is caused by fungi. The symptoms may take the form of purple or brown rings enclosing a green or yellow center, or they may be solid purple or brown circles. Olive leaf spot disease affects the vigor of the olive tree by damaging the leaves and causing them to prematurely fall to the ground, which, in turn, decreases productivity dramatically [6]. In addition, Aculus Olearius is one of the species which has been recently recorded in the Kingdom of Saudi Arabia. It dramatically damages olive trees, causing significant harm to the olive oil sector in the country, by distorting the buds and leaves, mainly in younger trees, and deforming the olive fruit. Purple or brown, circular leaf spots appear on affected leaves that turn yellow and fall prematurely [6]. Among olive species, the leaves are frequently affected by olive scab which is a wide-spread fungal disease that is awful looking and damaging, as it affects the well-being of olive trees, resulting in premature falling of leaves leading to an economic loss for the growers of olive trees. Olive leaf spots, scientifically referred to as Aculus Olearius, as well as olive scab, can be visually noticed and recognized. However, this represents only the first step in the process of detecting the disease affecting olive trees, which may not be enough. Thus, there is a need for contacts with experts to go further in the process of detecting the disease. Such a process may be time consuming and also costly. Human errors are also plausible. As a matter of fact, there is a need for an automatic, inexpensive, accurate, and in particular, a fast method to detect this olive disease [7].

Deep learning (DL) is one of the major domains of machine learning that uses neurons similar to mathematical functions for learning tasks [8]. Recently, neural networks (NN) have been commonly used in decision-making strategies; while the use of deep neural networks (DNN) such as a convolutional neural network (CNN) has attracted enormous interest since it is widely utilized for disease detection [9]. Deep CNN is a deep learning mechanism that offers enormous success as compared with other conventional methods in different image complexities. Deep learning has recently solved many complex problems and has shown excellent performance in many computer vision and machine learning tasks such as image classification, object detection, speech recognition, voice recognition, natural language processing, and medical imaging. Thus, one of the deep CNN tasks has been utilized for plant disease detection aimed at replacing traditional approaches of diseases detection [10].

In this paper, we investigate the feasibility of using DNN models for detecting olive leaf diseases. The dataset consisted of images from real cultivation fields in the Al Jouf region in the Kingdom of Saudi Arabia that were collected using an autonomous drone that flew over olive farm zones with over 400 trees; the experimental work was conducted between June 2021 and February 2022. The considered drone was equipped with a camera for taking images of the olive leaves, and then the images were transmitted to the cloud to identify the olive leaf diseases using deep learning techniques.

The proposed solution in this work consists of a hybrid DL model that includes a combination of ResNet50 and MobileNet named “MobiRes-Net” that is used to detect olive leaf diseases. Therefore, our main contributions can be summarized as follows:

❖

Building the first suitable dataset to be used for training and testing similar work;

❖

Conducting a trail work that combines an innovative deep learning framework with a drone to identify three types of olive leaf diseases, and thus, provide accurate diagnoses and helpful recommendations in an efficient and timely manner;

❖

Comparing the findings of the proposed work to state-of-art works.

The remainder of this paper is organized as follows: Related works are investigated in Section 2; the materials and methods are presented in detail in Section 3; in Section 4, we described implementation and testbed; followed by the experimental results in Section 5; finally, the conclusions are stated.

2. Related Work

In this section, we highlight the major applications of using machine learning (ML) approaches in smart farming to deal with olive diseases. A summary is presented in

Table 1. A summary of related works.

Ref.	CNN Networks	Dataset	Accuracy (%)	Strength	Weaknesses
[11]	AlexNet	PlantVillage	99.11	The work is robust to noisy samples	Takes time during the training process
[12]	VGG16	PlantVillage	98.6	The approach requires less training data	Higher computational complexities
	ResNet		97.5
[13]	VGG-16	Olive leaf images	89.12	Higher accuracy ratings	Higher computational complexities
	VGG-19		88.5
	Proposed model		95.63
[14]	Modified ResNet50	Collected	97.12	Higher accuracy and faster detection speed	Takes higher execution time
[15]	AlexNet	Collected	99.06	Reduced number of parameters with better accuracy	Results are reported for a small dataset
	AlexNetOWTBn		99.44
	GoogleNet		97.27
	Overfeat		98.96
	VGG		99.48
[16]	Inception	Collected	95.91	Improved accuracy in olive disease detection	Classification performance needs further improvements
	AlexNet		89.90
	ResNet.50		94.12
	ResNet-v2		91.81
[17]	AlexNet	PlantVillage	87.34	The technique is computationally efficient	Takes higher execution time
	VGG16		92.87
	Inception-v3		94.32
	ResNet		92.56
	Proposed model		97.87
[18]	ResNet101	PlantVillage	96.46	Fast, inexpensive, and deployable model	Requires a large amount of data to train a model
[19]	AlexNet	Collected	90.12	Ability to reflect the similarity of the same type data itself	Higher computational complexities
	Resnet		96.25
	DensNet		98.35

to be used later on to highlight the research gaps and to report our own research motivations.

An enhanced CNN model with the name AlexNet was proposed by Alruwaili et al. [11] with the purpose of diagnosing 14 various diseases that affect olive leaves, olive tree trunks, and olive fruit. The aim was to enhance the accuracy and to minimise errors in the diagnosis of olive diseases. The following is a description of the system overview: (1) Data are obtained from the PlantVillage dataset; further, it is enriched with olive data gathered for the laboratory of Aljouf. (2) A small window median filter is used to preprocess leaf images. (3) Later on, the above proposed AlexNet Model is used to process images. (4) The final phase is meant to diagnose and feedback. The proposed method in this study using transfer learning achieved a high rate of accuracy equal to 99.11%.

“Abstraction-level fusion” is a model developed by Cruz et al. [12] with the purpose of diagnosing a common disease known as the olive quick decline syndrome (OQSD). The disease was first discovered in Italy where it had a devastating effect on olive trees. The model was trained with a PlantVillage dataset, and it was retrained later on with the dataset of researchers. Experiments were conducted including 100 images of healthy leaves in which 99X fastidiosa-positive leaves and 100X fastidiosa-negative leaves had symptoms of other factors of stress. The transfer learning method, healthy leaves, and leaves affected by OQDS were applied in the proposed work and the classification rate was shown to be about 98.6%.

Uğuz et al. [13] proposed a deep convolutional neural network (DCNN) aimed at the classification of aculeus olearius diseases and olive peacock spot. A total of 3400 samples of olive leaves were chosen in which there were healthy leaves, olive peacock spot, and leaves affected by aculeus olearius. To accomplish the experimental study, transfer learning methods on VGG16 and VGG19 architectures were used. Based on the dataset, experiments were conducted in two major sections: with and without data augmentation. Furthermore, a web-based application of this study was developed. The findings showed that olive peacock spot and Aculus Olearius could both be recognized with low error rates even without the presence of an expert.

To help detect and classify olive spot disease, Aditya et al. [14] proposed a method that used an analysis of leaf image textures. In the presented methodology, histogram thresholding was used to isolate the infected area, k-means segmentation, and texture features such as energy, homogeneity, and entropy. The process of testing and detecting these diseases was conducted through the gathering of olive leaf images affected by Spilocaea oleaginea fungus and Neofabrea fungus. However, the infected images were extracted from online sources. A high connection was identified with Neofabraea leaf spot disease, which included energy (r = 0.97) and entropy (r = 0.92) properties of spot texture.

To perform plant disease detection, researchers in [15] managed to develop convolutional neural network models as well as deep learning methodologies, mainly, AlexNet, VGG, GoogleNet, AlexNetOWTBn, and Overfeat. To train the models, the researchers made use of an open database consisting of 87,848 images of both healthy leaves and infected plants. The dataset was composed of images taken in experimental setups and others captured in the field in real cultivation conditions. Various architecture models were trained and one of them reached a high success rate of about 99.53% in identifying corresponding diseases.

To classify olive-fruit varieties, Juan et al. [16] proposed a system using CNN; 2800 fruits that belonged to seven different olive varieties were photographed. For better image quality, mathematical morphology and global thresholding techniques were used to transform the initial captures. Such a transformation helped to obtain a set of images of individual fruits. Furthermore, various architectures of CNN were tried through training them with the considered images. However, the highest accuracy reached (95.91%) was obtained using the Inception-ResnetV2 architecture.

In [17], a model to identify plant leaf diseases that made use of a nine-layered deep CNN was proposed by the authors. To train the model, an open dataset of 39 various classes of plant leaves and background images was used. To do so, the researchers used six different types of data augmentation methods which were image flipping, noise injection, gamma correction, principal component analysis (PCA) colour augmentation, scaling, and rotation. The performance of the model could be enhanced through the use of data augmentation. To train the proposed model, various training epochs, dropouts, and batch sizes were used. Validation data were used to improve the performance of the proposed model. The extensive simulation resulted in a classification accuracy of around 96.46%.

A novel deep learning method with the purpose of detecting and classifying plant diseases was proposed by the researchers in [18] following three major steps. Initially, to obtain the region of interest, annotations were developed. Then, the researchers introduced an improved CenterNet in which there was a proposal of DenseNet-77 to achieve deep extraction of key points. In the last step, CenterNet was used as a one-stage detector for the detection and categorization of the different plant diseases. Thirty-eight types of crop diseases from the PlantVillage dataset were effectively located and classified by the proposed system. The extensive simulation achieved a classification accuracy of 96.46%.

A proposed ResNet DNN architecture for leaf disease classification was shown by the authors of [19]. Its aim was the development of a deep learning-based system to predict and to classify vegetable leaf diseases. To carry out the work, five major vegetable crops were considered. To train the model, a public dataset was used. However, real-time data were gathered from agricultural fields for testing. The design of the system and its testing were carried out making use of various types of a CNN mainly AlexNet, LeNet, VGG16, VGG19, and ResNets. The prediction of the overall performance of the system was a maximum of 98%.

3. Deep Learning Architectures

In this section, we aim to present an overview of the DL concept along with its well-known model, which help to understand the proposed hybrid DL model. As part of a machine learning algorithm, the immense potential of DL has recently been shown with state-of-the-art performances both on computer visions and the processing of images. DL has been largely used in agricultural images in detection, segmentation, and classification tasks and has proven to produce high results.

3.1. ResNet50 Model

Kaimimg He, in 2015, introduced Resnet-50 which was a type of CNN [20]. It consists of 50 layers; 48 of the total number of layers represent the convolution layers, one layer is the max pooling layer, and the remaining layer is the average pooling layer, as presented in Figure 1. The ResNet model is composed of a pile of comparable or “residual” blocks [21]. A block operates as a pile of convolutional layers. An identifying mapping path is used to connect the output of the block to its own input. To reduce the computational load while calculating the 3 × 3 convolution, we made use of a bottleneck block of three layers that used 3 × 3 convolutions with the purpose of lowering and re-establishing the channel depth. A 7 × 7 kernel size and 64 different kernels give us one layer. In the second convolution, 1 × 1 kernel size and 64 kernels following 3 × 3 kernel size and 64 kernels, and finally 1 × 1 kernel size and 256 kernels give nine layers. In the third convolution, 1 × 1 kernel size and 128 kernels, following 3 × 3 kernel size and 128 kernels, and finally 1 × 1 kernel size and 512 kernels give 12 layers. In the fourth convolution, 1 × 1 kernel size and 256 kernels, following 3 × 3 kernel size and 256 kernels, and finally 1 × 1 kernel size and 1024 kernels give 18 layers. The fifth and sixth convolutions give nine layers and one layer, respectively, giving a total of 50 layers [22]. Hence, a new image data matrix which is more compact is generated to be run throughout the network [23].

3.2. MobileNet Model

Howard et al. (2017) proposed the MobileNet [24]. It is a lightweight DL model. The base paper (Howard et al., 2017) presented the comprehensive tests on resource and accuracy trade-offs and demonstrated a higher performance if we compare it to other well-known models. MobileNet makes use of depthwise convolutions which are separable, i.e., a single convolution is performed on each colour channel instead of joining all three and flattering them, as in a standard 2D CNN, which can be clearly seen in Figure 2. This enables the input channels to be filtered [25]. Due to the depthwise convolutions which divide the input features into two layers, the performance of the model increases. Each layer of them is subdivided into the following layer by linking it with output features until the procedure is accomplished. The activation function used between the different layers in the MobileNet architecture is ReLu [26] which can flatten the nonlinear outputs from the previous layer and feed it as an input to a subsequent layer [27].

4. Proposed Hybrid DL Model

In this section, we present the proposed hybrid DL model of the MobiRes-Net model that can be used for olive disease recognition. Furthermore, the model’s aim is a smart system that enables the processing of olive leaf images supported by machine learning algorithms as well as the detection of various symptoms of diseases affecting olive trees. In this section, we detail our approach for classifying olive leaf images into four statuses, namely: “healthy”, “Aculus Olearius disease”, “olive scab”, and “peacock spot disease”. In this proposed work, we perform several stages of image processing techniques for detection and classification of olive leaf diseases. Figure 3 portrays a schematic view of the proposed hybrid DL model.

Firstly, after collecting data from olive groves using a drone, the input data are preprocessed by resizing the input image into the required pixels for training the dataset. Then, two methods, i.e., contrast enhancement and image normalization, are applied to change the value of the pixel intensity in order to acquire a better-enhanced image. The changed pixel intensity reveals the hidden information existing within the low range. Secondly, a data augmentation technique is used with the purpose of increasing the overall performance of this system by adding more diverse data to the existing limited dataset. The used data augmentation techniques are built on rotation and Gaussian blur.

4.1. Dataset Construction

To train and assess the model, olive leaf images were gathered from an olive grove. The area of the study lies in the Al Jouf region in the Kingdom of Saudi Arabia at an altitude of 37°23′57″ N and a longitude of 3°24′47″ W. The location enjoys a Mediterranean climate with mild-wet winters and serious summer droughts. The average annual rainfall is 100 mm and the annual temperature is of 25 °C. The area in which the test is conducted lies in an olive grove having a surface of 30 hectares. It comprises one thousand olive trees which were planted in 2006. A flat rectangle of 560 m × 280 m that contained around 100 olive trees was used as the object of our study. An aerial view or the proposed concept is shown in Figure 4.

For the aerial inspection mission, a drone, on which all hardware modules for the sender were mounted, was used. The drone is a fixed-wing drone which can cover large geographic surfaces with a low consumption of energy. A strip’s approach was used to plan the flight path of the considered drone in order to improve the efficiency of energy use. Figure 5 shows the use of Mission software calibrated by remote control. Clearly, the figure shows a take-off point of the drone to start its aerial inspection journey, as well as an end point where the drone finishes the task of dedicated test area. A rechargeable battery that enabled flying for about half an hour was installed on the considered drone that could fly at an altitude ranging from 25 to 30 m. Thus, the camera that was fixed onboard the drone was used to capture images and to send them to the cloud for storage and analysis. Then, administrators could access the stored images by using the proposed model to evaluate the status of the olive tree leaves and identify any diseases and insects.

To build the dataset and to use it later in this study, 5400 varied images of olive leaves were gathered from the above-mentioned geographical area. The varied images contained a variety of olive diseases. Further, with the help and support of an agricultural engineer, the gathered data were classified into four different classes (Aculus Olearius, olive scab, peacock spot, and healthy), as shown in Figure 6. Then, the dataset was divided into two major groups; 80% of the dataset was used for training and the remaining 20% of the dataset was used for testing.

Table 2. Dataset description.

Class Name	Train Set	Test Set	Total
Aculus Olearius	1040	260	1300
Olive scab	800	200	1000
Peacock spot	1360	340	1700
Healthy	1120	280	1400

presents the image classes and the number of objects per class.

4.2. Data Preprocessing

In this work, a camera was fixed on the drone, which could capture images from an altitude, and then send them to the user to identify the olive leaf diseases. Given the fact that the camera was able to capture leaf images from various occlusions, there was an urgent need to train the model with different images taken from various angles, directions, and altitudes [28]. Furthermore, the gathered images were preprocessed in order to resize the input image to fit the required pixels to train the dataset. The preprocessing methods were adopted for the sake of enhancing the robustness of the experimental dataset to various illumination conditions. They were also valuable for eliminating the undesirable noise which was present in the presented image. The data preprocessing consisted of two steps:

❖

Step 1: All the images that were originally gathered were first checked to determine the minimum height and width. Once the minimum dimension was determined, the dataset images were resized to fit the required dimensions. Since the minimum dimension in our work was set to be 224 × 224, all the dataset images were resized to 224 × 224.

❖

Step 2: To acquire a better-enhanced image in the present work, two methods, i.e., contrast enhancement and image normalization, were applied to change the value of pixel intensity. By changing the pixel intensity, hidden information that existed within the low range could be revealed. The change of pixel intensity revealed the hidden information already existing within the low range.

4.3. Data Augmentation

Data augmentation is a strategy to significantly enhance data [29]. A variety of image augmentation methods are used for the sake of increasing the performance of a system by adding diversified data to an already existing limited dataset. Table 2 shows the difference in the initial number of acquired images in every image class. Such a difference creates a mass class imbalance which may result in different issues, namely overfitting where the model cannot effectively generalize on the hidden dataset. In fact, data augmentation techniques were adopted in this work to deal with the issue of overfitting. Two data augmentation techniques were adopted in this work, i.e., rotation and Gaussian blur:

❖

Rotation: For the purpose of generating a greater number of images, images were rotated at different angles ranging between −15° and 15°. The augmented images were included in the dataset used to train the model.

❖

Gaussian blur: High frequency components were removed from the images to be blurred and smoother making use of a Gaussian filter of kernel size 5 × 5. The augmented images using the Gaussian filter were included in the dataset used to train the model.

4.4. Deep MobiRes-Net Model Training

In earlier studies, researchers have separately used the Resnet-50 and MobileNet models for olive leaf diseases. In this study, the proposed hybrid DL model extends the design of the Resnet-50 model with the MobileNet network to design the MobiRes-Net model. Figure 7 shows the architecture of the proposed deep MobiRes-Net model, which is divided into two parts: Resnet-50 and MobileNet. Firstly, the previously trained weights of each model, as shown in Figure 7, were loaded. Then, the classification half of each model was replaced by an average pooling layer and a fully connected layer with 1024 neurons. The weights of the fully connected layers of each branch were then merged via concatenation and three more fully connected layers of sizes 1024, 512, and 256 neurons. Olive leaf diseases were finally classified making use of the Softmax classifier in the output layer. The function appearing in the output layer is a type of activation function that is used in multi-class classification problems.

5. Implementation and Testbed

In this section, we analyse the effectiveness of the suggested scheme considering the conducted experiments. To perform the experiments, an Intel Core i7 3.5 GHz processor machine was used, which was equipped with 16 GB RAM and a 2 GB inbuilt NVIDIA graphics card. The implementation of the hybrid DL model took place through python programming making use of deep learning frameworks, namely TensorFlow and Keras. The dataset, as described in Section 4.1, was split into two major groups, (80%) for training and (20%) for validation. The first group was used to carry out the training process while the second group was used to test the final evaluation.

The performance of the developed model was analysed using the following metrics: accuracy, precision, recall, and F1-score. Each measure is delineated as follows:

❖

Accuracy: The accuracy is the percentage of properly predicted images out of the total number of predictions [24]. The accuracy is calculated as:

$$Accuracy=\frac{Number of correct predictions}{Total number of predictions}$$

❖

Precision: The ratio of correctly predicted positive results (TP) to the total number of positive results (TP + FP) forecasted by the model is the accuracy metric. The vast number of FPs results in a lower precision [30]. The precision range is between 0 and 1 and is calculated as follows:

$$Precision=\frac{Number of True Positives}{Number of True Positives+Number of False Positives}$$

❖

Recall: The recall is utilized to measure the correct positive forecasts by calculating the ratio between the number of true positives results (TP) to the total number of samples (TP + FN) [30]. The recall is calculated using the following equation:

$$Recall=\frac{Number of True Positives}{Number of True Positives+Number of False Negative}$$

❖

F1-score: The F1-score is one of the metrics used to measure and evaluate the model’s performance. The weighted harmonic between the accuracy and recall is used to compute the F1-score. Its definition is as follows:

$$F1 score=2\times \frac{Recall\times Precision}{Recall+Precision}$$

where TP is true positive, FP is false positive, TN is true negative, and FN is false negative.

At 100 epoch validation, Figure 8, Figure 9 and Figure 10 show the results of two performance indictors (accuracy and loss) of the ResNet50 model, MobileNet model, and the proposed hybrid model, respectively. Figure 8 demonstrates the accuracy and loss graphs of the ResNet50 model, where the highest accuracy achieved is 92.86% and the loss is 0.15. Figure 9 illustrates the accuracy and loss graphs of the MobileNet model, where the highest accuracy achieved is 94.63% and the loss reaches 0.15. Figure 10 presents the accuracy and loss performance indicators of the proposed hybrid model, where it obtains an accuracy of 97.08%, while the loss scores 0.09. It is worthwhile to note that the best result is obtained by concatenating the ResNet50 and MobileNet models, thanks to the added fully connected layer.

The classification task of the dataset was performed via the three deep learning models to obtain the classification results. The average classification results are shown in

Table 3. Average classification results of the original dataset.

Model	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)	Training Time (s)	Testing Time (s)
ResNet50	92.86	92.75	92.08	92.78	890	2.29
MobileNet	94.63	93.26	93.12	94.16	940	1.79
MobiRes-Net	97.08	97.61	97.11	96.86	1138	4.1

. The table also illustrates the training and testing run time in seconds. As shown in Table 3, the ResNet50 and MobileNet models both provide average accuracy scores below 95%. However, the highest accuracy score of 97.08% is achieved via the proposed MobiRes-Net model. The highest F1-score (96.86%) as well as the highest recall score (97.11%) are also achieved by MobiRes-Net. In addition, the training time of our proposed MobiRes-Net model is 1138 s. This is superior to the ResNet50 (890 s) and MobileNet (940 s) models. Since nothing can be perfect, our proposed MobiRes-Net model has some limitations. One limitation is the higher training and testing run time as compared with other models, because of the complex structure of the models inside modules. Another limitation is that high-performance hardware is required in order to process the models.

Table 4. Model classification results for the four-category classification task for the original dataset.

Model	Category	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
ResNet50	Aculus olearius	93.25	91.25	91.55	91.36
	Olive scab	91.45	91.45	91.85	91.57
	Peacock spot	91.75	91.75	91.35	91.03
	Healthy	96.23	93.23	93.77	93.89
MobileNet	Aculus olearius	94.15	94.52	92.25	94.55
	Olive scab	94.52	93.63	93.15	93.85
	Peacock spot	93.01	94.71	93.75	93.75
	Healthy	97.98	93.05	92.95	93.63
MobiRes-Net	Aculus olearius	96.25	97.35	97.27	96.78
	Olive scab	96.45	97.85	97.48	97.05
	Peacock spot	96.75	96.45	96.72	96.75
	Healthy	98.23	98.45	98.15	97.23

shows the classification results for four categories that represent the olive leaf images (Aculusolearius, olive scab, peacock spot, and healthy) obtained from the ResNet50, MobileNet, and MobiRes-Net models. Classification was studied from four perspectives: accuracy, F1-score, precision, and recall. As shown in Table 4, ResNet50 achieved an accuracy of 93.25% for the Aculus olearius category, 91.45% for the olive scab category, 91.75% for the peacock spot category, and 96.23% for the healthy category. The highest recall (93.77%), precision (93.23%) and F1-score (93.89%) are for the healthy category. In addition, MobileNet also attained the highest accuracy (97.98%) for the healthy category, the highest precision (94.71%) for the peacock spot category, the highest recall (93.75%) for the peacock spot category, and the highest F1-score (94.55%) for the Aculus olearius category. However, it can be noticed that MobiRes-Net achieved an accuracy of 96.25% for the Aculus olearius category, 96.45% for the olive scab category, 96.75% for the peacock spot category, and 98.23% for the healthy category. The highest recall (98.15%), precision (98.45%) and F1-score (97.23%) are for the healthy category.

The confusion matrix for the ResNet50, MobileNet, and MobiRes-Net models used in the study include the findings that have been obtained from the training carried out with the test dataset presented in Figure 11, Figure 12 and Figure 13.

6. Comparative Analysis

In this section, we provide more details and verification of the proposed MobiRes-Net model against existing work of the state-of-the-art methods, as shown

Table 5. The results obtained as compared with state-of-the-art methods.

Ref	CNN Networks	Dataset	Accuracy (%)	Precision (%)	Recall (%)	F1-Score (%)
[11]	AlexNet	PlantVillage	99.11	92.75	92.08	92.28
[12]	VGG16	PlantVillage	98.6	92.26	93.12	92.25
	ResNet		97.5	96.61	97.11	96.86
[13]	VGG-16	Olive leaf images	89.12	92.75	92.08	92.28
	VGG-19		88.5	92.26	93.12	92.25
	Proposed model		95.63	96.61	97.11	96.86
[14]	Modified ResNet50	Collected	97.12	92.75	92.08	92.28
[15]	AlexNet	Collected	99.06	92.26	93.12	92.25
	AlexNetOWTBn		99.44	96.61	97.11	96.86
	GoogleNet		97.27	92.75	92.08	92.28
	Overfeat		98.96	92.26	93.12	92.25
	VGG		99.48	96.61	97.11	96.86
[16]	Inception	Collected	95.91	92.75	92.08	92.28
	AlexNet		89.90	92.26	93.12	92.25
	ResNet.50		94.12	96.61	97.11	96.86
	ResNet-v2		91.81	92.75	92.08	92.28
[17]	AlexNet,	PlantVillage	87.34	92.26	93.12	92.25
	VGG16,		92.87	96.61	97.11	96.86
	Inception-v3		94.32	92.75	92.08	92.28
	ResNet.		92.56	92.26	93.12	92.25
	Proposed model		97.87	96.61	97.11	96.86
[18]	ResNet101	PlantVillage	96.46	92.75	92.08	92.28
[19]	AlexNet	Collected	90.12	92.26	93.12	92.25
	Resnet		96.25	96.61	97.11	96.86
	DensNet		98.35	92.75	92.08	92.28
Our proposed	ResNet50	Collected	92.86	92.75	92.08	92.78
	MobileNet		94.63	93.26	93.12	94.16
	MobiRes-Net		97.08	97.61	97.11	96.86

. The obtained results reveal the superiority of our proposed model in the matter of accuracy for the four-category classification task. In addition, the performance of the MobiRes-Net model is better than the ResNet50 and MobileNet models. These results should only be used for reference because the methodologies and metrics used to calculate each system are different, and therefore, an accurate comparison cannot be made.

7. Conclusions and Future Work

The growing of olive trees is an essential economic activity in various countries, including The Kingdom of Saudi Arabia. However, the presence of plant diseases has affected regular production in a very negative manner. Recently, with the advantages of automatic DL and feature extraction, these techniques have received increased interest from academic and industrial circles. Simultaneously, it has become a hotspot for further research in the field of agricultural plant protection, namely the recognition of plant diseases. This paper is based on experimental work using a remote-controlled agricultural UAV, where it moves autonomously in an olive farm. The UAV is equipped with a camera and coupled with the hybrid DL technique it moves autonomously in olive farms for detecting and classifying olive leaf diseases. Then, the captured images are transmitted to the Cloud for identification of olive leaf diseases using deep learning techniques. To help detect olive leaf diseases, a hybrid DL model is proposed in this work. It is an amalgamation of the ResNet50 and MobileNet models, and is given the name MobiRes-Net. The proposed model achieved a classification accuracy of 97.08%, which was superior rate as compared with the ResNet50 and MobileNet models whose classification accuracies were, respectively, 92.86% and 94.63%. Since nothing can be perfect, our proposed MobiRes-Net model has some limitations. One limitation is the higher training and testing run time as compared with other models because of the complex structure of the models inside modules. Another limitation is that high-performance hardware is required in order to process the models. In the future, the MobiRes-Net model could be installed in a Field-Programmable Gate Array (FPGA) processor to make a standalone device, which would be easier to integrate with a drone monitoring system.