Tanzania’s economy is predominantly centered around agriculture, and the country gains from a wide range of agricultural activities, such as livestock, essential food crops, and many cash crops (Oxfordbusinessgroup, 2018). In Tanzania, agricultural output accounts for about 29.1% of the country’s Gross Domestic Product (GDP). It also employs 67% of the labor force, a paramount supplier of food, raw materials for industry, and foreign exchange (International Trade Administration, 2021). Moreover, as agronomy production is far too low, food demand is increasing dramatically (Dewbre et al., 2014). Farmers, scientists, researchers, analysts, specialists, and the government are working hard to enhance agricultural production to meet growing needs (Panigrahi et al., 2020). However, crop diseases continue to be a challenge affecting major food security crops like maize (Savary and Willocquet, 2020). Maize is a very crucial and important crop in Tanzania, contributing significantly to the country’s agricultural sector (Maiga, 2024). However, maize leaf diseases such as Maize Streak Virus and Maize Lethal Necrosis, pose a severe threat to maize production with the potential to reduce yield (Shepherd et al., 2010; Mahuku et al., 2015; Kiruwa et al., 2020). Early detection of these diseases is crucial for implementing timely preventive measures and mitigating yield losses (Boddupalli et al., 2020; Haque et al., 2022). Traditional visual analysis methods for disease detection in crops are prone to errors, labor-intensive, and time-consuming. Moreover, these methods have been observed to identify diseases at a later stage, potentially leading to more harm to the crops (Toseef and Khan, 2018; Gong and Zhang, 2023). These traditional methods rely heavily on the expertise of farmers, plant pathologists, and agriculture experts. Additionally, the subjective nature of these methods can lead to inconsistent diagnoses among different experts.

Recently, technology has been used to improve yields in agriculture, whereby researchers have devised several solutions, including image processing and object detection using deep learning models (Panigrahi et al., 2020). Deep learning (DL) is a branch of machine learning that involves training artificial neural networks to learn from large volumes of data and make predictions. Moreover, it is known for its ability to use many processing layers to discover patterns and structures in large datasets (Rusk, 2015). It moreover automatically extracts features from the data, making them suitable for various applications, such as image recognition, natural language processing, speech recognition, and autonomous systems (Ho, 2016). It has become widely known for its potential and advanced ability to efficiently process large numbers of images, yielding reliable outcomes. It is doing very well in many fields, including agriculture (Kamilaris and Prenafeta-Boldú, 2018). During the last few years, many crops have become accustomed to detecting, classifying, and assessing a broad spectrum of diseases, pests, and stresses (Singh et al., 2016; Panigrahi et al., 2020; Haque et al., 2022). For the past several years, deep learning achievements in computer vision tasks have strongly depended on Convolutional Neural Networks (CNNs) (Raghu et al., 2021). CNNs prevail in the domain of computer vision as a foundation for various applications, such as image classification (Sibiya and Sumbwanyambe, 2019; Darwish et al., 2020; Syarief and Setiawan, 2020; Atila et al., 2021; Chen et al., 2021; Liu and Wang, 2021; Haque et al., 2022), object detection (Zhang et al., 2020; Liu and Wang, 2021; Maxwell et al., 2021; Roy et al., 2022) and image segmentation (Gayatri et al., 2021; Liu and Wang, 2021; Loyani and Machuve, 2021; Maxwell et al., 2021; Sibiya and Sumbwanyambe, 2021). The CNN architecture consists of components such as a convolutional layer, a pooling layer, a fully connected layer, and activation functions (Bharali et al., 2019; Francis and Deisy, 2019; Jasim and Al-Tuwaijari, 2020), as shown in Figure 1.

Natural language processing has been performed using transformer architecture, and vision transformers have produced outstanding outcomes compared to CNNs (Vaswani et al., 2017; Qi et al., 2022). Researchers have recently adapted transformers to computer vision applications, inspired by the significant success of transformer architectures in the field of NLP. The Vision Transformer (ViT) has achieved cutting-edge performance on various image recognition benchmarks. In addition to image classification, transformers have been used to solve a variety of computer vision problems, including object identification, semantic segmentation, image processing, and video interpretation. Because of their superior performance, an increasing number of academics are proposing transformer-based models for improving a wide range of visual tasks (Han et al., 2023). ViT works by implementing a transformer-like architecture over image patches. Images are divided into fixed-size patches, which are then linearly embedded. Position embeddings are then added, then the resulting vector sequence is fed into a standard transformer encoder. The standard approach of adding an extra learnable classification token to the sequence is used to perform classification (Vaswani et al., 2017; Dosovitskiy et al., 2020). The sequence of the 1D array is passed to the transformer structure. To process 2D image patches, the 2D patches are extracted from the first, and then they are reshaped to create 1D arrays that are suitable for the ViT structure. They are added to the positional encoder to finish preparing the patch embedding for the next layer. The positional encoder aids the network in remembering the relative position of the patches with one another. Inputs are then normalized with the normalization layer before entering the transformer block. The multi-head attention layer is the most important aspect of this block. The multi-head attention layer calculates weights to assign higher values to the more important areas. In other words, network attention is focused on the most important parts of the network. The output of the multi-head attention layer is a linear combination of each head (Borhani et al., 2022). Figure 2 shows the ViT architecture inspired by Vaswani et al. (2017).

Both the ViT and CNN models have achieved state-of-the-art results in various computer vision tasks, including plant disease detection. However, the relative performance of the model would depend on the specific dataset, model architecture, and training hyperparameters used in a certain study. A lot of various techniques have been developed and proposed for the detection of diseases in general. The most adopted techniques CNN and ViT have shown great performance when used separately Therefore, this study aimed to develop combined deep-learning models for the early detection of Maize Streak Virus (MSV) and Maize Lethal Necrosis (MLN) diseases in maize plants based on images obtained and collected directly from the field, allowing the model to be trained with real data. The grand purpose is to utilize the maize imagery datasets collected from farms and made available in open source to the research community for future studies on MLN and MSV infections, by introducing, an approach that enhances the effectiveness and efficiency of these diseases in maize. Hence this paper fills a gap existing in a debate between the most quality and reliable model for detection of maize diseases.

The diagnosis of a wide variety of plant diseases and pests has shown encouraging and remarkable results when employing deep learning techniques in computer vision, such as CNNs. A convolutional neural network deep learning model was developed to analyze images of healthy and unhealthy plant leaves. A total of 87,848 images in an open database with 25 distinct plants in 58 distinct categories of healthy and unhealthy images were trained using five model architectures, AlexNet, AlexNetOWTBn, GoogLeNet, Overfeat, and VGG. VGG was the most common architecture for detecting plant diseases, with a higher success rate. Implementation was performed using the Torch71 machine learning computational framework, which uses the LuaJIT programming language. The model’s exceptionally excellent performance makes it suitable as a vital early warning or advising tool (Ferentinos, 2018). This study was conducted in Athens, Greece, to detect many plant diseases and not specifically for the detection of maize streak virus and lethal maize necrosis.

Another deep-learning model was developed to detect maize diseases in Indonesia. The study used a classification approach to detect 3 diseases, Cercospora, northern leaf blight, and common rust. A support vector machine, k-nearest neighbor, and decision tree were used to classify the maize leaf images, and seven other CNN architectures were used to analyze the maize leaf images. The architectures used included ResNet50, GoogleNet, VGG19, AlexNet, Inception-V3, VGG16, ResNet110 and VGG19. The data consisted of 200 images that were divided into 4 classes, 50 images per class with a size of 256×256 pixels. However, AlexNet and SVM were the best methods for feature extraction and image classification of maize leaf diseases. This study used fewer samples (200 images), which were collected in Asia (Syarief and Setiawan, 2020).

Additionally, a Mobile-DANet model was developed to identify 8 maize crop diseases, gibberella ear rot, maize eyespot, crazy top, gray leaf spot, Goss’s bacterial wilt, common smut, phaeosphaeria spot, and southern rust. Except for some samples, the results of the Mobile-DANet model demonstrated that the majority of the images and maize diseases were correctly identified. Mobile-DANet correctly detected samples with phaeosphaeria spots with a probability of 0.71. Similarly, the model accurately detected gibberella ear rot and southern rust disease, with probabilities of 0.83 and 0.93, respectively. China served as the study location, and this study focused on maize images other than MSV and MLN images. The model employed in the study is Mobile-DANet (Chen et al., 2021).

Furthermore, another study from India proposed a deep convolutional neural network to detect healthy and diseased images of maize leaves. The dataset contained 5,939 images of maize leaves. The dataset consisted of images of three diseases, Maydis leaf blight (MLB), Sheath blight (BLSB), Turcicum leaf blight (TLB), and banded leaf, as well as healthy maize leaves. The study employed the Inception-v3 network structure, as well as three more different models were developed using the normal training procedure (Haque et al., 2022).

In Cairo, Egypt, a classification model for the identification of common rust, northern leaf blight, healthy maize leaves, and gray leaf spots was developed. To identify plant diseases, an ensemble model composed of two pre-trained convolutional neural networks, VGG19 and VGG16, was used to distinguish between the leaves in healthy photos and the leaves in unhealthy photos. The outcomes show how well the suggested strategy works, outperforming alternative methods for VGG19. Even though the created model performed well, this study struggled with the categorization of unbalanced data, and the dataset employed lacked sufficient images to properly train CNNs that were created from scratch (Darwish et al., 2020).

A model for the recognition of common rust (Puccinia sorghi), gray leaf spot (Cercospora), and northern corn leaf blight (Exserohilum) from healthy leaves was developed due to the impacts of these diseases on the majority of the maize plantations in South Africa. Neuroph was used for training the convolution neural network to recognize and classify images of maize. CNN was quite correct in identifying these diseases. This research was restricted to the neuroph framework of the Java neural network, which is an integrated environment for developing and deploying neural networks to Java programs, despite the model’s strong performance (Sibiya and Sumbwanyambe, 2019).

A similar study was conducted by Sibiya and Sumbwanyambe (2021) to develop a CNN deep learning model. The diseased leaf area was calculated using segmentation by the threshold on diseased images of leaves of maize impacted by common rust disease. This information was used to create ambiguous decision guidelines in assigning common rust images to severity groups with images created using this proposed approach. The VGG-16 network, trained with images generated using this suggested method, achieved a higher testing and validation accuracy when tested on photos of common rust illness in 4 stages of severity (early stage, middle stage, late stage, and healthy stage). Despite the good performance of the developed model, this study was limited to only the image segmentation approach, which tends to partition a digital image into multiple segments. Furthermore, the study used a CNN architecture, which lacked a detailed description.

Arnaud et al. (2022) from Kenya developed a deep learning model to examine, in contrast, 6 convolutional neural network architectures. Transfer learning was employed for model training, and the architectures used included EfficientNet b7, VGG19, SqueezeNet, GoogleNet, AlexNet, and DenseNet. The study analyzed four hyperparameters: the batch size, learning rate, number of epochs, and number of optimizers. An open-source dataset with 4,082 photos was used. DenseNet121 outperformed other models by achieving a higher accuracy and F1 score. DenseNet121 was trained with batch 32, a learning rate of 0.01, and stochastic gradient descent (SGD) as the optimizer. In general, various techniques for detecting plant diseases have been proposed. These techniques have shown good performance; however, no studies have focused on building a combined deep-learning model for the detection of MSV and MLN together, and there is no publicly available dataset containing images of maize leaves infected by MSV and MLN. Moreover, several studies have used a large number of images from online sources, which might not accurately represent field scenarios. As a result, this study aimed to develop a combined deep learning model for MSV and MLN detection based on images collected directly from the field, allowing the model to be trained with real data. The dataset will be made available in open source to the research community for future studies on MLN and MSV infections. Furthermore, the majority of the studies employed transfer learning methods, and the scope of their studies was not in Tanzania.

This study developed two deep learning models, CNN and ViT. Both models performed well in detecting MSV and MLN diseases in maize plants. The ViT model achieved a validation accuracy of 93.1%, whereas the CNN model achieved an overall average validation accuracy of 90.97%. These results suggest that both models are capable of detecting the presence of diseases in maize plants. Furthermore, these results are considered to be among the best examples of a good model, as a good model is expected to have an accuracy greater than 70% (Maxwell et al., 2021). However, deep learning models also perform very well when trained with larger datasets. The CNN model for this study was trained with 27,588 data samples compared to Syarief and Setiawan (2020) who used a few data samples (200) for model training in the detection of maize diseases. The majority of the studies have employed transfer learning to train deep learning models for maize diseases detection and their scope is not focused on Tanzania (Darwish et al., 2020; Syarief and Setiawan, 2020; Chen et al., 2021; Arnaud et al., 2022; Haque et al., 2022), unlike the study where both CNN and ViT deep learning models were developed from scratch and the study area is Tanzania. Another study by Sibiya and Sumbwanyambe (2021) developed a deep learning model for early detection of maize disease using a segmentation approach, while the approach of the study for our case was classification. Furthermore, none of the studies has come up with a combined deep-learning model for the early detection of MSV and MLN diseases in maize. Additionally, when the developed deep learning models were compared, the ViT model had somewhat greater accuracy than the CNN model. According to Dosovitskiy et al. (2020), the ViT model’s key design which includes the ability to capture global dependencies through self-attention mechanisms gives it an advantage in detecting and classifying various plant diseases with higher accuracy than the CNN model. Furthermore, ViT divides the input image into patches and processes these patches as sequences, enabling the model to learn a high-resolution and systematic representation of the image data. However, when the prediction speed for both models per image is compared. CNN is 10 milliseconds faster than ViT which is 20 milliseconds per image.

This study has shown that early maize disease detection is possible in Tanzania, with a specific focus on the Maize Streak Virus (MSV) and Maize Lethal Necrosis (MLN). The study collected a substantial dataset comprising 26,913 field-acquired images and 675 wrong images acquired from open-access databases. The dataset’s availability as an open-source resource will facilitate further research on MSV and MLN infections. Deep learning models, namely, convolutional neural networks (CNNs) and vision transformers (ViTs), were developed to address the challenge of early disease detection. Both models were developed from scratch, with CNN demonstrating its ability to extract local image features, while ViT demonstrated proficiency in understanding the global image context. ViT achieved a validation accuracy of 93.10%, while CNN achieved a validation accuracy of 90.96%. This highlights the value of deep learning models in the early diagnosis of plant diseases in maize.