Round 70% of the world’s date fruit comes from Saudi Arabia, thanks to its more than thirty-one million palm trees (The Ministry of Environment and Water and Agriculture Saudi Arabia, 2020). In 2023, the more than 26,000 date farms located in the Madinah area of western Saudi Arabia produced about $253 million. Roots, trunks, leaves, and fruits of palm trees are all vulnerable to a host of infectious illnesses caused by bacteria and fungi (The Saudi Press Agency (SPA), 2024).

The inventory of palm trees is essential for assessing diversity and health; however, data regarding their numbers and distribution is limited, outdated, and inconsistent. Estimations of date palm trees in plantations rely on assessments that exclude non-agricultural areas and natural populations. Data on canary palm populations is restricted to areas of significant public interest, with estimates derived from date palm production rather than geospatial databases (Jaradat, 2015; Sharma et al., 2021; Zaid and de Wet, 2002).

Agriculture, and date palm trees in particular, are very vulnerable to diseases and climate change. Date palm trees suffer greatly from serious diseases, including Brittle Leaf disease, Brown Leaf Spot, Bayoud disease, Black Scorch, SDS, and Brown Leaf Spot, which drastically reduce fruit quality and productivity. This study examines the detection of the SDS disease, which disseminates in a hazardous manner, complicating control efforts and resulting in substantial losses in fruit output. This disease poses a significant threat to Date palm farming worldwide and hinders fresh planting efforts (Khamparia et al., 2020).

Palm leaves are integral to numerous ecosystems, economies, and civilizations. They play a crucial role in various aspects, supporting individuals in earning a livelihood and staying on their chosen course in life. Nevertheless, the Dubas bug presents a significant risk to date palm trees and their leaves, especially in areas where date palm farming is prevalent. Identifying and classifying plant diseases is crucial for precision agriculture; however, farmers face challenges in diagnosing these diseases and assessing the extent of infestation. The combination of machine vision and deep learning has revolutionized the automated detection and assessment of pests and diseases in agriculture. Palm leaf diseases pose a considerable challenge to the health and productivity of trees (Resh and Cardé, 2009; Hessane et al., 2023).

Adult, nymph, and egg are the three phases that make up a dubas bug’s life cycle. With its two sets of wings, this insect is hemimetabolous. It undergoes a mating cycle in the spring (February–May) and another in the fall (August–November) of each year. Date palm females use the third and fifth leaf fronds, respectively, to deposit their eggs in the spring and autumn.

The life cycle begins with oviposition, followed by hatching into nymphs and undergoing five molts until the adult form is attained. Adults have a yellowish-brown to greenish coloration, characterized by two black patches on their heads. They generate honeydew and necrotic regions in plant tissues due to their oviposition behavior. Nonetheless, it remains uncertain whether these necrotic lesions result from fungal infections (Elwan and Al-Tamimi, 1999; Thacker et al., 2003; Ba-Angood et al., 2009; Resh and Cardé, 2009; Payandeh et al., 2010).

Computer-based technologies are rapidly revolutionizing agriculture, reducing human labor, and enabling impartial decision-making. Image processing methods are utilized in various computer vision applications for disease diagnosis, identification, and segmentation tasks. This technological development is revolutionizing agriculture, improving efficiency and efficacy. Using a modified MobileNetV2 neural network, the authors (Kong et al., 2022) enhanced the accuracy of cassava leaf disease classification by employing data augmentation methods on lower-quality test images.

With high-quality photographs, they achieved a 97% recognition accuracy; however, this accuracy was significantly reduced with low-quality images. Using a range of classifiers at the image level, including Fine KNN, Cubic SVM, and tree ensemble, the authors in (Li et al., 2023) classified guava plant illnesses with an overall classification accuracy of 99%. Plant disease classification is achieved by a hybrid wrapper model that combines CNN classifiers with FPA-SVM (Bayomi-Alli et al., 2021). This methodology yielded a classifier with an accuracy of 99% through feature selection using FPA and SVM in a wrapper approach. The authors (Almadhor et al., 2021) proposed a deep learning (DL) model for disease detection on cucumber and potato leaves, utilizing an optimization approach. A 99% success rate was achieved by optimizing the DFs obtained from the global pooling layer using an upgraded Cuckoo search strategy. For disease categorization in plant leaves, Yağ and Altan (2022) presented the EfficientNet DL architecture. They employed a transfer learning approach to train their model and other deep learning models, and both performed well. To classify citrus diseases, Zia Ur Rehman et al. (2022) developed a new DL model. The accuracy rate was 95% because the Whale Optimisation Algorithm (WOA) was used to retrain two pre-trained models, DenseNet 201 and MobileNetv2, so that they could produce feature vectors. A DL model for guava disease identification has achieved a 97% accuracy rate by utilizing enhanced data supplemented with color-histogram equalization and unsharp masking techniques (Atila et al., 2021). The primary contributions of this research are as follows.

Make a substantial contribution to the categorization of palm leaf diseases by using cutting-edge deep learning architecture. This paper presents a method for the automated detection and enumeration of palm leaves, disease identification, and assessment of palm health from high-resolution images using deep learning and Vision Transformer models. The training dataset included more than 1600 images of individual dubas and healthy classifications. We assessed the model via training assessment and by comparing prediction outcomes with visual and ground inspections. The model was also evaluated using images captured at varying elevations. It can achieve elevated accuracy with less labeled data by using pre-trained models. This method enhances classification efficacy while reducing training duration and computational expenses. Additionally, it streamlines the process of disease identification, providing a scalable and rapid option for early detection. This is essential for agricultural disease management.

Several new approaches to identifying and categorizing plant diseases have recently been developed. The methods were tested on a variety of datasets, each with its distinct features. Representing plant photos using practical and discriminative characteristics is essential for creating a system to identify plant diseases. There are two primary schools of thought regarding feature extraction methods: those that rely on manually created features and those that utilize deep learning techniques (Shah et al., 2016; Sethy et al., 2020; Iqbal et al., 2023; Tholkapiyan et al., 2023).

Using CNN with SVM, Kamal et al. (2018). We were able to differentiate between Chimara (the most prevalent date palm spot leaf disease) and Anthracnose (the least frequent), achieving accuracy rates of 97% and 95%, respectively. To attain a classification accuracy of 99.67% with an artificial neural network (ANN) classifier, Hamdani et al. (2021) suggested an approach that makes use of a color histogram feature and a dataset consisting of 300 laboratory photos. With an overall accuracy of above 96%, Liu et al. (2021) used a DL-based Faster RCNN to identify and quantify oil palm plants in UAV pictures. Abu-zanona et al. (2022) achieved the best accuracy for the Kaggle dataset, classifying four types of sick palm trees with 97% accuracy using VGG 16 and MobileNet.

Septiarini et al. (2021) presented a method for detecting diseases in oil palm leaves. Otsu thresholding was employed in the Lab color space to identify Regions of Interest (ROIs), followed by preprocessing and classification using k-nearest neighbors (KNN) (Eunice et al., 2022). The identification of leaf diseases in tomato plants was conducted by Sunil S. Harakannanavar et al (Harakannanavar et al., 2022). This approach combines multiple techniques, such as SVM, KNN, and CNN, for the detection of palm diseases. Recent studies have developed a CNN for disease classification in palm trees (Abu-zanona et al., 2022).

Authors (Nusrat et al., 2020) were able to detect and categorize wheat illness leaves with a 98% success rate using SVM and GoogLeNet, two ML models. A prior study by (Too et al., 2019) compared several DL models for plant disease detection. This study utilized 121-layer customized models, including VGG-16, Inception V4, ResNet, and DenseNets, to classify plant species. This study used the model to detect and classify healthy leaves, as well as four prevalent diseases that may affect palm trees: bacterial leaf blight, brown spots, leaf smut, and white scale. When tested against VGG-16 and MobileNet, two prominent CNN models, the suggested model outperformed them both with an accuracy rate exceeding 99%. In a separate study, Hamdani et al. (2021) examined methods for identifying and categorizing diseases that affect palm trees. The authors of this study achieved a 99% success rate in classifying palm diseases using an artificial neural network (ANN) classifier and principal component analysis (PCA) to extract color features.

Masazhar and Kamal (2017) employed an automated system to detect and categorize disease indicators in palm oil leaves. Two palm oil illnesses were successfully detected using k-means clustering, an SVM classifier, and leaf symptomatology. Thirteen novels were produced from k-means clustering for disease classification. Ashqar and Abu-Naser (2018) utilized a CNN model for identifying tomato leaf diseases, demonstrating improved performance with full-color images compared to grayscale images. Shruthi U et al. highlighted the efficacy of convolutional neural networks in identifying specific agricultural diseases through machine learning methodologies (Shruthi et al., 2019).

The experiments were evaluated using a GPU P100 Kaggle system as the baseline. Operating Windows 11, the machine features 16 GB of RAM and a 9th-generation Core i7 CPU. Aiming to enable deep learning applications with reduced memory usage and improved execution speed, the software implementation included libraries such as Anaconda, Keras, OpenCV, NumPy, and CuDNN. For every experiment carried out, this work assessed the training and testing accuracy. Every model has calculated losses throughout the testing and training periods. Training the models on the Palm tree dataset helped to improve the learning speed of the transformer and transfer learning models. EfficientNetV2B0, DenseNet121, and ViT models were used for this work.

Each of the two dataset classes corresponded to a different disease. Since the Palm dataset’s color images worked well with the DL and ViT models, we used them in our experiments. The images were downscaled to a uniform pixel format since different pre-trained network models require varied input sizes. Input dimensions of 224 × 224 × 3 (height, width, and channel depth) are used by EfficientNet V2B0, DenseNet 121, and ViT. Data augmentation after preprocessing is a regularization strategy that is used to reduce the impact of overfitting. This method of model augmentation makes the model more robust, which in turn enhances its ability to categorize images of real plant diseases while reducing the likelihood of overfitting and model loss.

This study investigated a transfer learning and transformer methodology to provide an edge computing solution for the identification and detection of palm leaf diseases. Python supports three pre-trained models. This study presents a novel method for automatically detecting palm leaf disease against a natural background. This enables the differentiation between groups of infected and healthy leaves. We developed the technique for EfficientNetV2B0, DenseNet12, and the transformer paradigm. The dataset comprises 1600 images of palm leaves, with 800 depicting healthy and 800 depicting Dubas. To reduce computing time, pre-processing was performed, including image resizing and normalization, followed by augmentation. Augmentation was implemented by rotation, flipping, shearing, and zooming methods. The models were used to identify palm leaf disease using the TensorFlow framework with an input dimension of 224 × 224 × 3. The suggested approach achieved superior performance, as indicated by an accuracy value. The experiment demonstrated that the ViT model outperformed the other three models, achieving a validation accuracy of 99.37%, which is comparable to previously published techniques. The developed model successfully sustained elevated recall values, accuracy, and F1 scores. Although several automated detection models for palm leaf disease have been developed, their efficacy has often proven insufficient due to the resemblance of class attributes. This study primarily focused on detecting Dubas insect-related diseases and healthy leaves. This limitation of the study did not include other diseases, such as Brittle Leaves and Brown Leaf Spot.