Mango Diseases Classification Solutions Using Machine Learning or Deep Learning: A Review

The mango crop suffers from several diseases that reduce both the production and the quality of the mangoes. This also reduces its price on the international market. Diagnosis of these diseases remains difficult in many countries due to poverty and lack of infrastructure. Plant pathologists use several techniques to identify these diseases. But these techniques are time consuming and relatively expensive for mango growers and the solutions proposed are often not very accurate and sometimes biased. In the last decade, researchers have proposed several solutions in the field of automatic diagnosis of mango diseases. Such solutions are based on Machine Learning (ML) and Deep Learning (DL) algorithms. In this paper, we divided these solutions into two groups: solutions based on classical ML algorithms, and those based on DL. In recent years, DL, especially Convolutional Neural Network (CNN) has become the most widely used method by researchers because of its impressive performance. The critical analysis of the proposed solutions has allowed us to identify their limits and potential challenges in mango disease automatic diagnosis.


Introduction
Mango (Magnifera Indica L.) is a lucrative fruit widely grown in tropical countries. This fruit belongs to the anacardiaceous family [1]. It contains significant amounts of vitamin A and vitamin C [2]. It is called the "King of fruits" because of its alluring aroma, flavorful pulp and high nutritional value that attract many mango lovers from around the world [1] [3]. In 2021, mango was the third most traded tropical fruit after pineapple and avocado in terms of quantities exported. Journal of Computer and Communications The Asian continent (912.510 tons exported) is the largest producer of mango in the world in 2021. It is followed by, South America (620.745 tons), Central America and the Caribbean (545.428 tons), Africa (202.010 tons), and Oceania (4.254 tons) [4].
However, mango suffers from several diseases at all stages of its life. Such diseases lead to a considerable reduction in both quality and quantity of mango production. In addition, its lead to a reduction of mango price on the local and international markets. Due to poverty and the lack of infrastructure in many parts of the word (e.g.: developing countries), the rapid diagnosis of these diseases remains difficult [5]. Mango growers and plant pathologists use naked eye observation to diagnose mango diseases and make decisions based on their experiences. Such decisions are often not accurate and biased sometimes because many types of mango diseases appear to be the same since their symptoms are similar at the early stage [6] [7]. This approach leads to unnecessary use of pesticides, which results in higher production costs. In addition, techniques used by plant pathologists are time-consuming and relatively expensive for mango growers. Timely and automatic diagnosis of mango diseases is, therefore very critical [8].
In recent years, with the advances in the field of computer vision, researchers have proposed several automatic diagnostic solutions for grading mango and diagnosing its diseases. Such solutions are based on classical Machine Learning (ML) algorithms and Deep Learning (DL) algorithms.
This paper proposes a review of mango diseases identification and automatic quality grading solutions proposed by researchers during the last decade.
The specific contributions of this paper include: • A critical analysis of recently proposed mango diseases automatic diagnostic and quality grading solutions.
• Identification of limitations of such solutions and potential challenges could help researchers in this area. The paper is organized as follows: Section 2 identifies the most common diseases of mango treated by researchers in recent years, Sections 3 provides a critical analysis of the proposed solutions based on classical ML and DL algorithms, Section 4 shows the potential challenges in automatic diagnosis of mango diseases and the last section concludes the paper.

A Review of the Most Common Mango Diseases
The production of mangoes suffers from several problems. These problems are caused by pests and diseases witch kill about 30% -40% of the crop [2]. Mango diseases are caused by a number of pathogens like bacteria, fungi, virus, algae and insect which attack all the parts of the plant, such as trunk, brunch, leaf, twig, petiole, flower and fruit [1] [2] [8]. The causes of these diseases can also be of climatic origin or unfavorable environmental conditions [9] [10]. Table 1 provides a summary of fourteen mango disease diagnostic solutions proposed by researchers between 2017 and 2022.  (Figure 1(a)) is the most common mango diseases treated by researchers during this period since losses of mango trees are caused up to 39% worldwide by this disease [11]. It is followed by Sooty Mold (Figure 1 Figure 2 shows the number of times these diseases are treated based on fourteen topics in Table 1.

Automatic Diagnosis of Mango Diseases
In recent years, several solutions for automatic diagnosis of mango diseases have been proposed. These solutions can be divided into two categories: those based on classical ML algorithms on the one hand, and those based on DL on the other.

Automatic Diagnosis Based on Machine Learning
In the field of automatic plant disease detection and identification, IP techniques   Table 1.
model suffers from a lack of training data. In [16] authors proposed a novel segmentation approach to segment the diseased part by considering the vein pattern of the mango diseased leaf. Diseases treated in this study are sooty mold and powdery mildew. The leaf's features were extracted on the basis of color and texture using canonical correlation analysis (CCA)-based fusion. They used ten different classifiers to identify leaf diseases but the best results were obtained with cubic SVM classifier (95.5% of accuracy). However, the amount of data used is very small and the time needed to identify diseased leaves is significant.
Moreover, the proposed architecture does not allow for real-time identification of treated diseases. In the study of [14], a method for automatic mango leaf diseases recognition and classification is developed. Authors used k-means clustering for image segmentation, GLCM (gray level color co-occurrence metrics) for feature extraction and SVM for diseases classification. SVM classifier gets accuracy up to 96%. But one of the proposed model's limits is that the presence of several diseases in the same region of a leaf makes image segmentation difficult.
Similarly, the variation in leaf color, texture, and shape made the feature selection phase difficult. Authors of [11]

Number of times
Mango Diseases accuracy of 98% using the proposed MRKT. The performances obtained are very significant but the proposed system is only limited to anthracnose disease. [13] is an improvement of [12] in order to recognize diseases with more accuracy.

Automatic Diagnosis Based on Deep Learning
In the last decade, in the field of automatic plant disease detection and identification, models based on DL algorithms are the most proposed by researchers.
These models have achieved state-of-the-art performance on ImageNet and other benchmark datasets [22]. The latest generation of Convolutional Neural Networks (CNNs) has achieved impressive results in image classification and is con-   [15] proposed a CNN model based on AlexNet architecture for detecting anthracnose mango leaf disease. The system is developed using Tensor Flow framework and a dataset of mango images captured in real condition with a CDD camera. The developed system was more than 70% accurate to isolate the diseased mango leaves. The model uses data taken under real field conditions but is specific to anthracnose.
In the study of [24], authors developed a CNN based on AlexNet architecture and using TL to detect and classify Mango and Grapes leaf diseases. They achieved an accuracy rate of 89% and 99% respectively for mango and grapes leaves. They implemented this model as an Android app named JIT CROPFIX. However, disease identification in real-time condition is a very challenging task compared to laboratory conditions. Another limitation of the proposed model is that the model ignores tiny or invisible defects on the fruit, which reduces the accuracy. In [17] authors proposed a novel framework for mango leaves disease classification namely Anthracnose, Bacterial black spot, and Sooty mold. They used a CNN with crossover-based levy flight distribution for better feature selection, MobileNetV2 model for the learning stage and SVM for diseases classification. The experimental results show classification performances over other state-of-art methods. Authors of [25] implemented a computer vision algorithm for defect detection on the surface of to Indian mango fruits varieties named Chausa and Dashehari. The overall efficiency and accuracy of proposed algorithm were 93.3% and 88.6%, respectively. However, multiple small defects or single large defect around the edge decreases the accuracy. Efficiency and accuracy of the proposed algorithm decrease when the severity of defect (blemishes or disease infected area) on the surface of fruit increased.
Several DL based solutions have been proposed for automatic sorting or grading mangoes. In [3] for example, combined and studied several DL models based on AlexNet, VVGGs, and ResNet to see which one performs best in clas- Authors of [26] designed and implemented a system using non-destructive thermal imaging by deep learning technique for mango sorting and grading accord- ing to their quality. The proposed system is based on SqueezeNet deep CNN.
Mango quality grading is performed by using parameters such as shape, defects, maturity and size. They used TL technique based pretrained SqueezeNet model and achieved 93.33% and 92.27% of accuracy for RGB and thermal images respectively, with the training time of 30.03 and 7.38 minutes for RGB and thermal images. Results show that the training time is lower with the thermal images (by factor of 4X). Table 3 summarizes proposed models based on DL and used for automatic diagnosis of mango diseases and mango automatic quality grading.

Potential Challenges in Automatic Diagnosis of Mango Diseases
Classical ML and DL algorithms have achieved excellent results in automatic detection and classification of mango diseases. However, the following challenges must be addressed: • The models proposed so far suffer from a lack of training data, which increases the time to identify diseases and reduces the performance of these models [8] [10] [11] [16]. • The proposed models do not provide real-time diagnosis of mango diseases. This would be very beneficial for mango growers and plant pathologists [16].
• The presence of several diseases in the same region of a leaf makes segmentation difficult. This reduces the accuracy of the model [14].
• Images taken in real condition majorly suffers from the problem of the variation in leaf color, texture, shape, temperature, shadowing, overlapping of leaves and presence of multiple objects. This makes the feature extraction phase difficult. Compared to laboratory conditions disease identification in real-time condition is a very challenging task [7] [14] [24].
• Detection of tiny (or invisible) defects (disease infected area or blemishes) and single large defect around the mango leaf or mango fruit edge is also a challenge [24] [25].

Conclusion
This paper is a review of mango's most common diseases treated and solutions proposed for automatic diagnosis of such diseases during the last decade. We divided these solutions into to two categories: classical ML based solutions and DL based solutions. During the last five years, DL-based solutions, especially CNNs, have been the most widely used since this technology requires less or no preprocessing of images compared to other techniques and offers excellent performances. In this study, a critical analysis of the proposed solutions allowed us to know their limitations and then to identify potential challenges that could be of interest to researchers in automatic diagnosis of mango diseases. We aim, for future work, to propose a dataset of mango diseases whose images will be captured in mango orchards of a Sahelian country like Senegal. Then we will use the CNN model (e.g. VGG16, Unet, AlexNet, MobileNet, …) which will give the best performances on this dataset and finally deploy this model in a mobile application to allow mango growers to be able to diagnose diseases in their mango orchards without the intervention of plant pathologist. We will focus on anthracnose and peduncular rot diseases which causes significant damage to mango production in Senegal.