Grapes are a popular fruit in both developed and developing countries. The grape is also essential for the wine industry [1]. However, in producing grapes, preventing and controlling diseases is crucial. Controlling the disease of grapes is necessary for healthy grape cultivation. It minimises losses and reduces the use of pesticides. The worldwide population is increasing, and ensuring a sustainable food supply is essential. However, due to grape diseases, grape cultivation can suffer significantly in quality, quantity, and value [1]. Early identification of grape leaf diseases is crucial to prevent their spread within vineyards. Healthy plants can be protected from diseases by promptly removing infected leaves and treating them as soon as possible.

A precision agricultural solution can improve grape production and reduce cultivation costs. Thus, precision agriculture offers efficient protection through remote monitoring and automation. Manual diagnosis is time-consuming and expensive. Therefore, artificial intelligence (AI) has recently positioned itself in precision agriculture, incorporating machine learning with machine vision and image processing. The outcome is quick, real-time, and more accurate disease detection. The AI-based tools generate an alert once leaf disease symptoms are present and formally confirm crop disease’s presence [2]. For example, Lu et al. [3], Li et al. [4], and Lu et al. [5] introduced novel AI models for grape leaf disease detection (GLDD). However, there are a few significant gaps in precision agriculture. Scholars are more inclined to present novel AI and the convolutional neural network (CNN) models than implement them on farms. This strength of precision agriculture is not extended to the users. Moreover, scientists are interested in increasing the model’s accuracy but do not examine how the model can be fitted into the device since the CNN models require substantial computational resources. Hence, end users are deprived of the advantages of precision agriculture. Lastly, some efficient algorithms, such as You Only Look Once (YOLO), are applied in a limited number of works, specifically in grape disease detection.

Realising the gap in this study, a mobile-based application, Grape Gaurd, has been developed and applied to grape farms. The application uses YOLO as an AI model to detect and classify grape leaf disease modalities. This study tests two different YOLO models, YOLOv5 and YOLOv8, to find the best solution for mobile-based devices. Lastly, YOLOv8 was customised because of its inherent capabilities in real-time object detection, making it suitable for implementation on mobile devices. In the field of machine learning, this is a significant contribution.

Realising the effectiveness of precision agriculture, significant AI-based techniques have been developed to address grape disease monitoring. For example, Lu et al. [3] proposed a transformer-based model that classifies grape leaf diseases named the ghost convolutional enlightened transformer. The network contains a ghost network as its backbone, which helps it generate intermediate feature maps and perform cheap linear operations. By analysing five hyperparameters, the model achieved an accuracy of 98.14% in classifying grape leaf diseases. In classifying grape leaf diseases, Li et al. [4] proposed a dense convolutional transformer (DCT) model, which introduces densely connected modules; the compact convolutional transformer serves as the backbone of this model, which improves the original model’s convolutional module. Lu et al. [5] developed Swin-T-YOLOv5 by architecturally integrating YOLOv5 and Swin-transformer detectors in real-time to detect wine grape bunches in natural vineyards. Two different Chardonnay and Merlot varietals were used for the experiment. The model achieves a mean average precision (mAP) of up to 97% and an F1 score of 89%. To improve GLDD, Praveen et al. [6] introduced You only look once-X (YOLO-X) with attention mechanisms. The authors applied attention techniques such as convolutional block attention modules (CBAM), squeeze-and-excitation networks (SE), and efficient channel attention (ECA) to focus on important features and reduce irrelevant ones. The YOLO-X model with SE, ECA, and CBAM attention achieved 89.77% precision, 86.97% recall, 85.91% F1 score, and 88.96% mAP. Xie et al. [7] developed the faster deep region-based inception and attention convolutional neural network (DR-IACNN) to enhance the Faster R-CNN model for detecting grape leaf diseases. The researchers first created the GLDD dataset using image processing techniques. Then, they enhanced the Faster R-CNN model with Inception-v1, Inception-ResNet-v2 modules, and SE blocks to improve feature extraction. The Faster DR-IACNN model achieved an 81.1% mAP score. For the efficient localisation of crop seedlings in complex environments, Kong et al. [8] proposed a target detection network where the YOLOv5 network and transformer module were used to detect targets. The whole crop labelling strategy (strategy A) and the single leaf labelling strategy (strategy B) were proposed as two labelling strategies to improve model accuracy and efficiency. With whole crop labelling, mAP@0.5 can be increased from 83.1% to 84.3%, and for radishes, from 77.3% to 81.9%. Shaheed et al. [9] proposed an efficient residual multiscale transformer network (RMT-Net) model to classify potato leaf diseases. With efficient RMT-Net, distinct features are extracted using the CNN model, and computational demands are reduced by depth-wise convolution. On a general image dataset, Efficient RMT-Net achieved an accuracy of 97.65%, and on a potato leaf dataset, it achieved an accuracy of 99.12%.

An inflorescence detection model based on transformers, named multiple-transformers-enabled YOLO (MTYOLOX), is presented by Xia et al. [10]. To explore the potential global context information and extract more distinguished features for inflorescence detection, the spatial-temporal path aggregation feature pyramid network (ST-PAFPN) module and dual attention transformer Darknet (DAT-Darknet) module were designed and rationally embedded into the backbone and neck of the network, respectively, based on multiple self-attention mechanisms. When faced with an orchard’s uncontrolled and challenging environment, MTYOLOX can adapt to varying illumination directions. Based on the modelled parameters, floating point operations (FLOPs), average precision (AP) and detection speed, MTYOLOX achieves the highest average precision (AP@0.5) of 83.4% and average recall (AR50) of 93.3%. Leng et al. [11] proposed a YOLOv5-based lightweight maise leaf blight disease detection model. Their model introduces the feature restructuring and fusion module and the Mobile Bi-Level Transformer, achieving 87.5% mAP@0.5 accuracy on the NLB dataset, a 5.4% improvement over previous models. Lu et al. [12] proposed a combined mixed attention mechanism (CMA-YOLO), a grapefruit detection model based on YOLOv5, which enhances detection accuracy through a dual-stream data loading scheme, grayscale processing, mosaic augmentation, and a novel CMA-cross convolutional cross stage partial (CMA-C3) module combining channel and spatial attention. The authors’ model incorporates a shifted window and global self-attention to improve feature distinction. Tested on the WGISD dataset, CMA-YOLO achieved a precision of 89.6%, an F1 score of 86.5%, and AP of 90.2%.

With YOLOv5s for region detection and a bidirectional cross-modal transformer (BiCMT) classifier for feature fusion, Feng et al. [13] proposed an end-to-end disease identification model. The model achieves 99.23% accuracy, 97.37% precision, 97.54% sensitivity, and 99.54% specificity on a small dataset. In another study, Li et al. [14] proposed YOLOv5s-FP (Fusion and Perception), a multi-scale collaborative perception network for pear detection. The authors introduce a pear dataset emphasizing small and occluded pears, comprising 3680 images captured from the ground tripod and unmanned aerial vehicle (UAV) platforms. YOLOv5s-FP utilizes a modified cross-stage partial (CSP) module with a transformer encoder for global feature extraction and attentional feature fusion. The network achieved AP of 96.12% for pear detection. Jiang et al. [15] proposed a tea leaf blight (TLB) detection method for natural scene images using the lightweight and efficient lightweight convolutional neural network (LC3Net) model. The authors employed the Retinex algorithm to enhance contrast and mitigate lighting variations. The LC3Net model, with image normalization and reduced down-sampling frequency, efficiently detects leaves of varying morphologies. Experimental results demonstrated an AP value of 92.29% for the LC3Net model. Sun et al. [16] proposed the SE-vision transformer (SE-VIT) hybrid network for sugarcane leaf disease identification. Their model utilizes support vector machine (SVM) for lesion extraction and integrates the SE attention module into ResNet-18, achieving 97.26% accuracy on the PlantVillage dataset. Using UAVs, Huang et al. [17] introduced the YOLO-EP algorithm for monitoring pomacea canaliculata eggs in rice fields. Based on YOLOv5s, the model incorporates transposed convolution and attention algorithms, achieving AP@0.5 of 88.6%, precision of 85.1%, and recall of 82.6%. A transformer-based model for leaf disease detection named Former Leaf was introduced by Thai et al. [18], addressing the increasing prevalence of leaf diseases due to climate change and pollution. The authors introduced the Least Important Attention Pruning algorithm to optimize the model size and evaluation speed while enhancing accuracy by 3%. This approach reduced the model size by 28% and improved the evaluation speed by 15%, utilizing sparse matrix-matrix multiplication for efficient computation. Chen et al. [19] introduced the ESP-YOLO model for accurately detecting mature table grapes. The proposed method enhances YOLO by incorporating efficient layer shuffle aggregation networks (ELSAN), Partial Convolution (PConv), SE, and soft non-maximum suppression (Soft_NMS) to improve feature extraction and detection efficiency. When tested on embedded platforms, the ESP-YOLO model achieved impressive mAP of 98.3%. Moreover, Liu et al. [20] introduced Fusion Transformer YOLO, a real-time and lightweight model for detecting four grape diseases using RGB images from North China. The authors utilized a lightweight high-performance VoVnet (LH-VoVNet) backbone enhanced with squeeze and excitation blocks, an improved dual-flow path aggregation network (PAN) + feature pyramid network (FPN) structure with a real-time transformer, and a decoupled head for balancing accuracy and speed. The model achieved mAP of 90.67 in disease detection. Table 1 shows a research matrix concerning previous studies.

Prior studies suggest that scholars have focused on developing AI models to improve accuracy but need more practical demonstrations. To fill this gap, this study demonstrates how a CNN model can be efficiently integrated into a mobile application to detect grape leaf diseases.

The experiment used Google Collaboratory and Android Studio Integrated Development Environment (IDE). The collaboration was used for YOLO model experiments, and IDE was used to develop the mobile app.

The experiment adopted in this study is presented in Fig. 1. The details of the experiment are described below:

Precision (P), Recall (R), and mAP as performance evaluation metrics were used in the research to evaluate the detection accuracy of the models. The corresponding formulas are given below:

$ \mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}} $ (1)

$ \mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}=\frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}} $ (2)

$ \mathrm{A}\mathrm{P}=\frac{1}{n}\displaystyle\sum _{i=1}^{n}\mathrm{A}{\mathrm{P}}_{i} $ (3)

Among all detected objects, Precision represents the ratio of accurately detected objects, while recall represents the ratio of accurately detected objects among all actual objects. True positive (TP) refers to how often a model correctly identifies positive instances. In a false positive (FP), the model incorrectly identifies a negative example as positive. When the model incorrectly identifies a positive instance as unfavorable, it is known as a false negative (FN). mAP is an evaluation metric used in various categories to evaluate object detection algorithms. A confidence score and an intersection over union (IOU) threshold are used to calculate mAP for each object category. Precision-recall curves at different IOU thresholds are expressed as the area under the curve or AP. mAP is calculated by averaging the AP scores of all classes. Higher mAP values mean better classification results among different courses.

This research collects a grape leaf disease dataset from Roboflow, which is publicly available [21]. The dataset contains 1598 annotated images of 4 classes: Black measle, black rot, healthy, and blight fungus. However, the performance of a classification task can be improved by applying various image processing techniques [22]. All the images are resized to 640×640 pixels. The dataset is split into 70% for training, 20% for validation, and 10% for testing. Fig. 2 illustrates the images from each labelled class. Table 2 details the different grape leaf diseases, outlining their key characteristics.

YOLO was chosen as a machine learning model since YOLO has proven to be effective in localizing and detecting objects. Objects here refer to the pattern of the disease leaf. YOLO treats object detection as a single regression task, which reduces computational complexity. It processes the entire image simultaneously, directly predicting bounding boxes and class probabilities. This method significantly increases detection speed over the faster region-convolutional neural network (Faster R-CNN) [27]. Two variants of YOLO, YOLOv5 and YOLOv8, were utilized in the experiment. A thorough description of YOLOv5 is presented in subsection 3.3.1, while YOLOv8 is addressed in subsection 3.3.2.

The YOLOv5 architecture has three main components: the backbone, neck, and head. The backbone employs a cross-stage partial Darknet (CSPDarknet), which incorporates cross-stage partial networks (CSPNets) into the Darknet to extract essential elements from the input image [28]. The system comprises Convolutional Layer + Batch Normalization + Sigmoid Linear Unit (CBS) modules and C3 modules, with a spatial pyramid polling fast (SPPF) module at the end to improve feature expression. SPPF eliminates redundant processes by max pooling pooled features, in contrast to SPPNet. Anchor boxes improve object detection accuracy, while n-maximum suppression (NMS) eliminates duplicate detections of the same object.

The image is processed through an input layer and then undergoes feature extraction by the backbone. The backbone generates feature maps of different sizes, combined using a feature fusion network to create three ultimate feature maps [29]. The maps are transmitted to the prediction head for confidence computation and bounding-box regression ((4)―(7)) for each pixel, using specified prior anchors. Irrelevant information in the array is removed by establishing certain thresholds and using the NMS method to set the final detection results.

$ {\mathrm{ga}}=2\sigma \left({s}_{a}\right)-0.5 +{r}_{a} $ (4)

$ \mathrm{g}\mathrm{b}=2\sigma \left({s}_{b}\right)-0.5 +{r}_{b} $ (5)

$ \mathrm{g}\mathrm{c}=\mathrm{p}\mathrm{c}{\left(2\sigma \left({s}_{c}\right)\right)}^{2} $ (6)

$ \mathrm{g}\mathrm{d}=\mathrm{p}\mathrm{d}{\left(2\sigma \left(s_d\right)\right)}^{2} $ (7)

The coordinate value of the upper left corner of the feature map is defined as (0, 0) in this context. The raw coordinates of the anticipated center point are denoted as $ {r}_{a} $ and $ {r}_{b} $. Also, the ga, gb, gc, and gd contain information about the updated prediction box, while pc and pd describe specifics about the preceding anchor. The model’s computed offsets are represented by $ {s}_{a} $ and $ {s}_{b} $. s_c represents the predicted offset for the x-coordinate of the center of the bounding box, while s_d represents the predicted offset for the y-coordinate of the center of the bounding box.This technique entails modifying the centre coordinate and dimensions of the initial anchor preset to match the ones on the final predicted box. The model uses the Adam optimizer, which combines the advantages of adaptive gradient algorithm (AdaGrad) and root mean square propagation (RMSProp) optimizers by integrating momentum and RMSProp. This optimizer modifies model parameters by utilizing moment estimates that include both first and second-moment estimates. The architecture uses the Mish activation function. Mish lessens the disappearing gradient problem in deep neural networks because it is not monotonic. The Mish activation function is shown

$ \mathrm{M}\mathrm{i}\mathrm{s}\mathrm{h}\left(x\right)=x\mathrm{t}\mathrm{a}\mathrm{n}h\left(\mathrm{s}\mathrm{o}\mathrm{f}\mathrm{t} \; \mathrm{p}\mathrm{l}\mathrm{u}\mathrm{s}\left(x\right)\right) $ (8)

where tanh is the hyperbolic tangent function, and soft plus (x) = $ \mathrm{l}\mathrm{n}\left(1 +{{\mathrm{e}}}^{x}\right) $ is a smooth approximation of the ReLU function.

YOLOv8 represents a significant advancement over its predecessor, aiming to enhance performance, speed, accuracy, and user-friendliness. The backbone network of YOLOv8 maintains the architecture of the CSP module from YOLOv5. The backbone network and neck module are inspired by the YOLOv7 efficient layer aggregation network (ELAN) design, choosing to substitute the C3 module of YOLOv5 with the more effective effective coordinates-to-features (C2f) module [30]. The Head module in YOLOv8 has been updated with a decoupled structure, separating the classification and detecting heads. The approach transitioned from an anchor-based to an anchor-free approach, increasing flexibility and adaptability.

YOLOv8 calculates loss using the task aligned assigner from task-aligned one-stage object detection (TOOD) and incorporates the distribution focal loss into its regression loss. The task aligned assigner employs a matching approach that selects positive samples according to the weighted scores of classifications and regression. The alignment metric for each anchor is determined by multiplying the predicted classification score of the corresponding class by IOU between the predicted bounding box and the Ground Truth bounding box.

$ T={q}^{\alpha }\times {p}^{\beta } $ (9)

where q represents the prediction score associated with the Ground Truth category, p denotes IOU between the prediction bounding box [31] and the Ground Truth bounding box. T represents the alignment metric for each anchor. α is a parameter that controls the influence of the prediction score q on the alignment metric, and β is a parameter that controls the influence of the IOU score p on the alignment metric.

The task-aligned assigner computes the alignment metric for each anchor for each Ground Truth by combining two values: The predicted classification score of the corresponding class and IOU between the predicted and Ground Truth bounding boxes. The alignment metric is then weighted for each anchor. The top-k samples with the and most significant alignment metric values are chosen as positive directly for each ground truth, as shown in Fig. 3.

The parameters used in the experiment are presented in Table 3.

The experiments were conducted on different image sizes on both YOLOv8 and YOLOv5 models. These image sizes are 640×640, 320×320, 256×256, and 128×128. The classification report of each experiment is represented in Table 4. The result suggests that YOLOv8 performs best when using an image size 640×640 than others image sizes.

Figs. 4 and 5 represent the precision confidence curve and the recall confidence curve of the YOLOv8 model with image size 640×640.

Figs. 6 and 7 represent the precision confidence curve and the recall confidence curve of the YOLOv5 model with image size 640×640.

Although the experiments were conducted using different image sizes, it was observed that both models (YOLOv8 and YOLOv5) performed quite similarly in each image size setting. However, YOLOv8 and YOLOv5 performed slightly better using an image size 640$ \times $640 configuration. YOLOv8 also outperforms YOLOv5 in each class detection and overall detection. Although YOLOv8 requires more training time than YOLOv5, the results of YOLOv8 exhibit significant improvements compared to YOLOv5. As shown in Table 4, YOLOv8 achieved 99.9% precision, 100% recall, 99.5% mAP, and 88% mAP50-95 scores in all class detection. Under the same configuration, YOLOv5 achieved only 99.6% precision, 99.5% recall, and 99.5% and 86% mAP50-95 scores, respectively.

The Grape Guard development process starts with the TFLite model generation from the YOLOv8 model.

The customized YOLOv8 model was trained using the grape leaf disease dataset to generate the TFLite model. The hyperparameters were Image size of 640× 640, Batch size of 16, and Epochs of 100. The you look only once v8 large model weight (YOLOv8l.pt) was selected to save weight as the model is designed for instance segmentation and classification [32]. During training, the best model (best.pt) was saved. The best model was ensured by testing on the validation set. Then, the best training process model was converted to TFLite format (YOLOv8Grape. TFLite) using the TFLite converter. Finally, the performance of the generated TFLite model will be evaluated using test images from the grape leaf disease dataset. Fig. 8 illustrates the comprehensive process of developing a TFLite model (YOLOv8Grape. TFLite) from the YOLOv8 model.

Before integrating the TFLite model into a mobile development framework, it is essential to understand the input tensor’s shape and data type, which helps prepare input data before feeding it into the model. Additionally, knowing the shape and type of the output tensor helps us interpret the results of our model correctly. A TFLite interpreter checks the TFLite model’s input and output tensor shape and data type. The details about the input and output shapes of the TFLite model are provided in Table 5.

The performance of the generated TFLite model is evaluated using test images from the grape leaf disease dataset. The testing parameters were the image size (640×640 pixels) and the confidence threshold (0.25). Setting the image size helps to maintain consistency with the image size used during training and conversion to the TFLite format. Additionally, the confidence threshold filters out detected objects with scores below the specified threshold during object detection. This evaluation verifies that the TFLite model produces accurate and reliable predictions, demonstrating its effectiveness in detecting grape leaf diseases on unseen data. Fig. 9 illustrates the testing results of the generated TFLite model, showcasing the model’s ability to classify grape leaf diseases.

The YOLOv8Grape. TFLite and the grapelabel.txt (contains the class name) were included in the assets directory of the created project (see Fig. 10). The application consists of two main user interfaces (UI) activities: Splash Screen and Main Activity, each with its corresponding extensible markup language (XML) layout file and Java class. Additionally, there are two other Java classes: Recognition.java, which serves as a data class to determine which labels will be displayed after predicting UI (label ID, label name, and label score), and Detector.java, which handles the detection process. In the Detector.java class, the input image size (640×640 pixels) and output shape of the TFLite model {1, 25200, 9} are initialized. The TFLite model and text file containing the box position and required condition are also initialized. In the MainActivity.java, permission is requested to access the camera and gallery, among others. In the application, the image captured by the camera or picked from the gallery in bitmap format is then passed for detection.

After completing the development process, the Grape Guard application was tested using test data from the grape leaf disease dataset. It was also tested in real-time scenarios by capturing images of grape leaves. The application can provide accurate detection in both cases, as shown in Fig. 11.

There is a call from researchers for portable systems applied to plant monitoring and precision agriculture [33–36]. This research attempts to provide a solution for grape disease detection and classification. The application, Grape Guard, is a real-time and portable system applicable to grape farms. The applied customized YOLOv8 outperforms AI studies by Chen et al. [19], Liu et al. [20], and Kaur et al. [37], Kaushik et al. [38] on detecting grape leaf diseases. Moreover, the result of YOLOv8 applied in this study provided better results than the studies that applied YOLO to grape disease detection.

This research makes a useful contribution in terms of its practical usefulness and user accessibility, compared to previous studies. The research is unique in that it directly integrates disease detection models into a mobile application called Grape Guard. Unlike other studies that mostly concentrate on constructing complex models utilizing different architectures, such as transformers, YOLO variations, and CNNs. The integration with a mobile platform enables end-users to identify grape leaf disease effortlessly and accessibly in real-world scenarios. The results indicate that the YOLOv8 model performed better than YOLOv5 in detecting grape leaf disease. This showcases a careful approach to selecting and optimizing models to ensure that the chosen model delivers optimal performance for the given task.

The main focus of this study was to introduce how a classification model can be integrated with a portable grape disease detection system to provide a solution for farmers. However, this study also has some limitations. One can criticize that the study used secondary data; however, in defense, we highlight that grape disease data was unavailable when this research was conducted. Among other limitations, this study used a limited class of diseases. however, we aim to include more diseases and primary datasets to validate our experiments in the future. It is also not confirmed if the application will provide the same result on other grape leaf disease datasets. The application is expected to be tested in other countries grape disease datasets. There is a need for ongoing validation and field testing to ensure the model remains effective over time. YOLOv8, while potentially more accurate, may require more computational resources. Consequently, prolonged use of the application can lead to increased battery drain.

This study utilized the YOLO model and integrated it with an Android application named Grape Guard. Various YOLO models, including YOLOv5 and YOLOv8, were utilized in the experiment to demonstrate their performance. Four categories of grape conditions were used for model training: Black Measles, Black Rot, Blight Fungus, and Healthy. The experiment results showed that YOLOv8 outperformed YOLOv5 in detecting each class, achieving 99.9% precision, 100% recall, 99.5% mAP, and 88% mAP50-95. The YOLOv8 model was selected to generate the TFLite file, which serves as the core of the Grape Guard application. The Grape Guard application’s user-friendly graphical user interface allows even those with limited knowledge to navigate and use its features easily. The results of this study demonstrate that YOLO models are particularly well-suited for mobile-based detection of grape leaf diseases, providing highly accurate results. This finding highlights the potential of YOLO models to enhance grape cultivation practices in the future significantly.

The authors declare no conflicts of interest.