Attention enhanced hybrid deep learning architecture with PCA-based feature fusion for banana leaf disease detection.

Attention enhanced hybrid deep learning architecture with PCA-based feature fusion for banana leaf disease detection - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Sci Rep . 2026 Apr 14;16:17291. doi: 10.1038/s41598-026-47092-7 Attention enhanced hybrid deep learning architecture with PCA-based feature fusion for banana leaf disease detection Hongyi Du 1, ✉ , Hua Yang 2 , Zhidong Zhang 1 , Jinjing Wang 3 Author information Article notes Copyright and License information 1 Department of Big Data and Intelligent Engineering, Shanxi Institute of Technology, Yangquan, 045000 Shanxi China 2 School of Information Science and Engineering, Shanxi Agricultural University, Taigu, 030801 Shanxi China 3 Shanxi Wanhong Technology Co., Ltd., Taiyuan, 030000 Shanxi China ✉ Corresponding author. Received 2025 Sep 30; Accepted 2026 Mar 30; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13234160  PMID: 41981085 Abstract Banana is a key staple crop in the world, but its yield is very desperately affected by the leaf diseases like Sigatoka, Fusarium Wilt, and Cordana, which bring serious losses of yield and economy. The manual disease detection methods deployed in the past are time and labor-intensive and cannot be effective in the field, so powerful automated solutions are sought. In this paper, we have presented a hybrid deep learning architecture which combines MobileNetV2 and ResNet101 with attention, dilated convolutions, multi-scale feature pooling and PCA-based feature fusion to classify banana leaf disease accurately and efficiently. The Banana and Banana Leaf, Banana Disease Recognition, and Banana LSD three benchmark datasets have been trained and evaluated following an innovative preprocessing pipeline that consists of illumination correction, background suppression, denoising, and hi-tech data augmentation. The experimental findings have shown that the presented hybrid model is always better than ten state-of-the-art deep learning models, such as VGG16, ResNet50, DenseNet121, EfficientNet-B0, and Vision Transformer (ViT). It has reached an optimal accuracy of 98.28, precision of 98.18, recall of 98.77 and F1-score of 98.43, with only 12.7 M trainable parameters and convergence rate of only 15 epochs, which makes the model both high-accuracy and computationally efficient. Keywords: Banana leaf diseases, Deep learning, Convolutional neural networks, Transfer learning, MobileNetV2, ResNet101 Subject terms: Computational biology and bioinformatics, Engineering, Mathematics and computing, Plant sciences Introduction Bananas, a staple food and a major source of income to millions of farmers in various parts of the world, is a heavily cultivated crop, especially in tropical and subtropical areas. Bananas, which are among the top five agricultural products, contribute significantly to the agricultural economy as well as the global food security, according to the Food and Agriculture Organization (FAO) 1 . Nevertheless, several foliar diseases, which affect yield and post-harvest value, are a major threat to banana crop productivity and quality. Some of the most devastating banana leaf diseases include the black Sigatoka, Panama disease, Fusarium wilt, and Banana streak virus, which destroys the photosynthetic ability of the leaf, causing stunted growth, ripeness, and loss of large amounts of money 2 . Effective disease management and timely intervention coupled with sustainable cultivation of these diseases is therefore heavily dependent on early detection and proper classification of such diseases. In conventional systems of disease diagnosis, manual examination by agriculturalists can be subjective, time-consuming and not applicable to large plantations 3 . In banana plantations, as well as in other places across the world, serious economic and yield losses are caused by banana leaf diseases, especially Black Sigatoka (Mycosphaerella fijiensis) 4 . The leaves of infected plants lose their photosynthetic ability causing abnormal ripening of fruits, low bunch size and low quality of fruit. Research has documented that Black Sigatoka by itself can lead to yield losses of 35–50% or in severe instances, yield losses of 100% in the event of uncontrolled cases. These diseases make production more expensive, as they highly depend on fungicides, which tend to be costly and polluting. Due to the predominance of banana production among smallholder farmers in some areas, such losses directly affect household income, food security, and export potential, and disease control is a crucial issue in maintaining banana production 5 . Recent improvements in artificial intelligence and image processing have resulted in the creation of powerful banana leaf disease classification methods, and there is a heavy focus on deep learning models and transfer learning models. Convolutional Neural Networks (CNNs) 6 are the most commonly used because they can be used to extract features of leaf images automatically without feature engineering. Transfer learning has been applied to state-of-the-art architectures like ResNet 7 , DenseNet 8 , VGGNet 9 , and Inception to reach high accuracy in classifying various common banana diseases, including Black Sigatoka, Fusarium wilt, and Banana streak virus. Moreover, it has been noted that hybrid methods using CNNs with standard machine learning classifiers like Support Vector Machines (SVM) and Random Forests have been investigated to enhance the resistance to classifications. In addition to CNNs, recent research has examined the use of Vision Transformers (ViTs) and attention mechanisms 10 , which improve the capabilities of the model in targeting disease-related regions of the leaf. Improving the dataset imbalance as well as enhancing generalization has been achieved by data augmentation methods (rotation, flipping, synthetic image generation with Generative Adversarial Networks (GANs)) that augment the dataset with additional images 11 . In addition, lighter deep learning algorithms and mobile disease recognition systems are under development to facilitate real-time disease detection under field conditions making these methods more affordable to smallholder farmers. Taken together, these strategies are a step toward scalable, precise, and farmer-friendly solutions to early banana disease detection 8 . Although there are various automatic techniques to classify banana leaf disease, several issues prevent the implementation of these techniques in the real agricultural field conditions. Available datasets usually are of limited size and are skewed in representation among disease groups and this limits the generalization capabilities of trained models. Further, the environment (including variations in lighting, background noise, the orientation of leaves, and occlusion) can greatly influence classification accuracy when models are evaluated under field conditions. Even though deep learning algorithms such as CNNs and Vision Transformers have shown to be very high-performing in controlled settings, their complexity prevents their implementation on resource-intensive systems such as smartphones, which are vital resource-heavy devices needed by smallholder farmers. Moreover, most of the existing techniques have difficulty in detecting the disease at its early stages when timely treatment is essential and do not usually work in cases of multiple diseases on the same leaf. Lastly, there is a gap between development of technology and its actual application because many of the available solutions are limited to scholarly research and are not converted into usable and affordable and easy-to-use tools. To solve these problems, the present study will suggest a powerful and scalable banana leaf disease classification framework based on advanced deep learning methods and with a practical applicability in real-world agricultural settings. The methodology makes use of a highly edited and augmented dataset to address the drawbacks of data imbalance and optimize model generalization in different field settings. It uses a lightweight but precise deep learning model, optimized to run on mobile and edge devices, to deliver real-time disease detection to smallholder farmers. The structure also focuses on pre-empting the disease at an early stage through a combination of fine-grained feature extraction and attention, which secures the timely intervention of the disease before it can cause extensive crop damage. What is more, the model is created to deal with multi-disease classification, allowing the detection of multiple infections on the same leaf at the same time, which is closer to the real-life situation. The main contributions of this work include: development of an optimized deep learning model for accurate and efficient classification of banana leaf diseases. incorporation of advanced augmentation and attention-based strategies to improve robustness in diverse field environments. introduction of a lightweight design for real-time, mobile-based deployment. demonstration of a practical, farmer-friendly solution that bridges the gap between academic research and agricultural practice. The rest of this paper is divided in the following way. Section “ Literature review ” provides a review of related banana leaf disease detection works, including the strengths and weaknesses of the available solutions. Section “ Material and methods ” includes a detailed description of the proposed methodology, such as preparing the dataset, augmentation strategies, and the design of a deep learning framework. Section “ Results and discussion ” presents the experimental design and measure of evaluation, reporting the findings and discussing the results in detail, comparing with baseline methods. Lastly, Sect. “ Conclusion and future work ” summarizes and wraps up the study by outlining the key contributions and specifying possible future research directions. Literature review Various studies have also investigated various image processing and machine learning approaches in the realization of detecting banana leaf diseases. One was a method that would combine color segmentation, the Otsu method, with histogram thresholding to deal with images that had highly complex natural backgrounds. This procedure performed well in isolating healthy green leaf areas and disease spots, weeds, soil, and other noise with a segmentation accuracy of approximately 97% and an error of 2.3%, showing its consistency in future segmentation and recognition studies. Other research combined CNN-based features extraction with the Support Vector Machines (SVM) to classify. This started with image preprocessing (denoising, resizing, and image quality) methods. The convolutional and pooling layers were used to extract features, and SVM was used to make the classification. This CNN-SVM hybrid method achieved approximately 94% in the detection of all common diseases such as Black Sigatoka, Cordana leaf spot and banana leaf speckle 7 . Another study project created a pipeline with image stitching, contrast equalization, segmentation and feature extraction. Experiments with the Scale-Invariant Feature Transform (SIFT) and BruteForce (BF) matching identified the disease regions with the Otsu thresholding method. The HSV space color features were then used to detect brown spots, which is often a feature of Sigatoka disease 12 . The system used in conjunction with SVM was able to detect Sigatoka spots with approximately 90% accuracy. In other studies, K-means clustering was used to do segmentation, extract features using color, shape, and texture, and ultimately classify using SVM. This method proved efficient in identifying several diseases such as Panama disease, Bacterial Wilt, Cucumber Mosaic Virus (CMV) and Sigatoka with approximately 85% precision. More recently, techniques that use a combination of histogram equalization with Fuzzy C-means clustering have been used to diseases like Black Sigatoka, Anthracnose and Freckle disease. Image quality was boosted by preprocessing including resizing, filter and elimination of Gaussian noise techniques and then a segmentation was performed using clustering followed by pattern recognition to classify the image. This technique, combined with Artificial Neural Networks (ANN), showed good results in identification of various diseases of banana leaves 13 . As deep learning advances rapidly, Convolutional Neural Networks (CNNs) have become popular in measuring the symptoms of plant diseases, which are much more accurate and more efficient than alternative methods. As an example, Arun et al. 14 trained CNNs on MalayaKew, the dataset of 44 plant categories, and showed that CNN-based models significantly outperformed conventional hand-crafted feature methods. In the same light, Yan et al. conducted a comparative study between manually engineered features and deep CNNs, which again confirmed the superiority of the latter 15 . Expanding on this, Nagachandrika et al. 16 suggested a multi-scale fusion CNN architecture, which utilizes the resizing of images to various resolutions and then pass through independent CNNs; the learned features were then fused, and the network was trained to capture more visual information. In addition to single-network approaches, ensemble-based approaches have also been proposed to enhance the robustness of classification. Pham et al. 17 tested different CNN backbones on coffee leaf classification implementing early and late fusion strategies and noting that results were affected by the fusion scheme. Sarawathi et al. 18 and Nagu et al. 19 also investigated the concept of ensemble learning, with the latter used four CNN models in training, out of which two best were selected to create a weighted ensemble, providing more credible predictions. Because high-quality labeled data acquisition is still one of the most significant challenges, scientists also explored algorithms that can work with a few annotations. Few-shot learning (FSL) has been suggested as an effective solution because it can train models to make predictions based on a limited number of samples that are labeled 20 . Thirumeninathanet al. showed this by testing a CNN with an SVM classifier on a source data set and freezing the CNN with fine-tuning of the SVM on the target data set only. They found that their ablation experiments were more accurate given more fine-tuning samples, with an accuracy of close to 90% when 80 samples were provided 21 . By applying a meta-learning method that employed a feature-level attention mechanism to a Vision Transformer (ViT), Saini 22 , further cut the number of samples per class to just five and therefore increased model robustness in the face of severe data scarcity. The recent developments in deep learning in terms of detecting plant and leaf diseases have paid more emphasis on enhancing the classification accuracy, robustness, and interpretability of the models across a wide range of crop species. Some papers have investigated optimization and hybrid approaches to improve deep learning, e.g. wild horse and hybrid crossover oppositional firefly algorithms to optimize CNN parameters and promote feature extraction in plant disease classification, which has been shown to be more robust than standard deep learning-based networks 23 . In another study, it was suggested to increase classification accuracy and sustain the agricultural practice by developing gradient-weighted DenseNet models that can be easier to interpret and achieve higher performance in relation to their computational costs 24 . Attempts to design light and efficient models are also becoming more common; recent research in grapevine leaf disease recognition showed new methods to downsample features and to do hyperparameter optimization to obtain high detection but at a lower computational cost, suggesting the prospect of real-world and edge deployment 25 . Besides this, studies that use explainable AI methods like SHAP on deep neural networks have furthered the explanation of the plant disease diagnostics, showing significant visualization effects that govern the choice made by the model when faced with complicated field conditions 26 . Moreover, the use of hybrid machine-deep learning reviews highlights the combination of conventional ML and DL schemes to exploit global and local characteristics, which is a tendency in more comprehensive, data-adaptive models in precision agriculture 27 . The detection of banana leaf disease has been investigated using a wide variety of image processing and machine learning methods, including color segmentation by Otsu thresholding and K-means clustering, as well as hybrid pipelines such as SIFT and HSV features and SVM classifiers with accuracies of between 85% and 97%. More complex methods used CNN feature extraction with SVM or CNN variants including dual-branch networks to learn shape and texture jointly, multi-scale fusion CNN to learn features at multiple scales, and an ensemble model to make stronger predictions. More recent studies have also examined approaches to few-shot learning and meta-learning to address the scarcity of labelled agricultural data, with encouraging results with CNN-SVM fusions and attention-enhanced Vision Transformers. Even with these developments, issues such as reduced performance when working in a complex field, reliance on large and heavily labeled datasets, lack of simultaneous performance across multiple diseases, and high computational complexity remain common, as do computational requirements that make the methods inaccessible to resource-limited farming systems. Materials and methods In this section, the datasets, preprocessing, and methodology framework used to classify banana leaf disease are described. The experiment used image data of healthy and infected banana leaves covering typical infections, including Black Sigatoka and Fusarium Wilt, and other foliar diseases. To train the model, several preprocessing procedures, such as resizing, noise reduction, and normalization, were utilized to standardize the input images and enhance feature extraction. Moreover, to augment the data used in improving the accuracy of the trained model, rotation as well as flipping and scaling were used to ensure the datasets were diversified and to help in reducing overfitting. The classification system was trained in deep learning architectures and mostly lightweight but accurate models that can be used in real-time detection. The subsections below describe the dataset, data preprocessing pipeline, model architecture, training configuration and evaluation measures in this study. Figure 1 shows the overall framework for proposed hybrid deep method for banana leaf disease classification. The architectural improvements utilized in this paper such as SE attention, dilated convolutions, GeM pooling, and PCA-based feature fusion were chosen with cautiousness to use their strengths. Channel-wise calibration of features is increased by SE attention, which increases the model capability to concentrate on informative features. The dilated convolutions increase the receptive field, which captures multi-scale contextual images without augmenting the computation cost. The feature maps are of variable size, which is well aggregated by GeM pooling, and multi-modes or high-dimensional features are well integrated by PCA-based fusion to eliminate redundancy. These elements were chosen over other methods that utilized early experiments which showed better performance in accuracy and computational efficiency hence the extension in the proposed architecture. The proposed model was made with lightweight and real-time constraints to facilitate the application. They have the benefit of a smaller number of parameters and the ability to extract features quickly and efficiently, which allows them to infer rapidly on resource-constrained edge and mobile devices without the need to have high-end GPUs. The simplified architecture and streamlined operations are likely to ensure the model has low inference latency and moderate power consumption, therefore, it can be used in continuous or near-real-time monitoring. These features indicate the possibility of the implementation of the offered method in practice where computation power, response speed, and energy consumption are very important factors. Fig. 1. Open in a new tab Proposed hybrid framework for banana leaf disease classification. Image preprocessing A multi-stage preprocessing pipeline was designed to implement a strong and correct classification of banana leaf diseases beyond the traditional resizing and normalization methods. The pipeline under development starts with lighting correction based on adaptive histogram equalization to reduce the impact of light variation in field-acquired images. Background suppression follows, using a series of procedures that include conversion to color space (RGB to HSV) followed by selective masking to isolate the leaf area in an image with respect to soil, weeds, or other extraneous objects often found in the natural scene. A contrast-limited adaptive histogram equalization (CLAHE) filter 28 is also used to maximize disease conspicuity (i.e., increase the strength of lesion patterns) without over-saturating the healthy areas. Subsequently, a wavelet-based multi-scale denoising methodology is used to remove high-frequency noise but retain texture features essential to detecting disease. Morphological operations are then applied to the segmented leaf regions to eliminate small artifacts and emphasize the boundaries of lesions. Advanced data augmentation techniques (geometric transformations (rotation, flipping, scaling), photometric variations (brightness, contrast, hue changes), lesion-specific (region-masking) augmentation are used to balance the datasets and enhance generalization of the model, as well as, more importantly, learn more discriminative features 29 . Lastly, to find consistency in training, validation, and testing data, all images are standardized to a common resolution and scaled to a standard pixel intensity range. This new preprocessing pipeline does not only clarify the features of diseases but also boosts the strength of the classification structure during various real-life situations, including variations in light conditions, background noise and partial hiding. The suggested system allows real-time monitoring of the banana leaf disease with the help of a mobile device. The smartphone camera captures images, and they pass through the basic preprocessing stage and are sent to the lightweight attention-enhanced hybrid model (Fig. 2 ). The model conducts inference at the device to determine the condition of the disease, and the inferred outcome is shown immediately to the user. This test process illustrates the effectiveness and feasibility of the suggested solution to the real-world use, mobile-based agricultural applications. Fig. 2. Open in a new tab Real-time mobile deployment workflow of the proposed banana leaf disease detection system. Datasets The datasets used in this research are summarized in Table 1 , and, together, they cover a variety of banana leaf conditions and types of diseases. The Banana and Banana Leaf Dataset 30 is well distributed with 1057 samples in five classes (Chewing Insect, Healthy, Sigatoka, Skipper, and Wilt). This balance is very appropriate in training deep learning models and reducing problems with class bias. The Banana Disease Recognition Dataset 31 , however, also has a certain level of class imbalance as Wilt is only represented by 287 samples as opposed to over 600 of the rest of the categories. This imbalance makes it difficult to train models reliably and requires the application of data augmentation, oversampling, or adjustments in class weight to promote fair learning across classes. Banana LSD (Leaf Spot Diseases) 32 data has 400 samples under each class of Cordana, Pestalotiopsis, Sigatoka and Healthy leaves. It appears to be relatively small, but clear structure based on leaf spot diseases makes it useful in measuring disease-specific classification performance. In general, a combination of these datasets allows extensive experimentation, which means that the given framework could be tested in both balanced and imbalanced conditions and, therefore, become more robust and applicable in practice. Table 1. Dataset details used for experiments. Sr. Dataset Classes Number of samples 1 Banana and banana leaf dataset Chewing Insect 1057 Healthy 1057 Sigatoka 1057 Skipper 1057 Wilt 1057 2 Banana disease recognition dataset Healthy 602 Insect 602 Sigatoka 630 Wilt 287 3 Banana LSD (leaf spot diseases) Cordana 400 Healthy 400 Pestalotiopsis 400 Sigatoka 400 Open in a new tab Hybrid deep feature extraction This paper utilized deep feature extraction techniques with two currently state-of-the-art convolutional neural network (CNN) models, MobileNetV2 33 and ResNet101 34 , to obtain discriminative features of banana leaf diseases. The two models were pre-trained on ImageNet and then fine-tuned on the banana leaf datasets to fit the target domain. MobileNetV2 based features The efficiency and accuracy of MobileNetV2 make it one of the backbones of feature extraction because they are especially well-suited to agricultural settings, where time and resource constraints can be a limitation. In contrast to conventional CNNs, MobileNetV2 uses depth wise separable convolutions and inverted residual blocks with linear bottlenecks that significantly reduce the quantity of operations and parameters but have no effect on discriminative power. Figure 3 shows modified MobileNetV2 architecture for deep feature extraction from banana leaf images. Fig. 3. Open in a new tab Modified MobileNetV2 architecture for feature extraction. A typical convolution involves computation costs proportional to , where is the kernel size, M is input channels and N the output channels. By factorizing this operation into depth wise and pointwise convolutions, MobileNetV2 reduces the cost to: 1 To improve the sensitivity of MobileNetV2 to recognize the subtle disease symptoms on banana leaves, the model was altered in three important aspects: Each inverted residual computing block was enhanced with Squeeze-and-Excitation (SE) attention mechanism, which enabled the network to re-tune channel-wise feature responses. The recalibration of a given feature map f is given by: 2 where GAP denotes global average pooling, and are weight matrices of the fully connected layers, δ is the ReLU activation, and σ is the sigmoid function. This ensures that more emphasis is placed on disease-relevant channels, while redundant activations are suppressed. To capture larger contextual details as extended areas of lesions, dilated depthwise convolutions were added, selected bottleneck layers. Dilated convolution of an input x and a kernel w can be defined as: 3 where r is the dilation rate. This modification enables the network to increase its receptive field without greatly increasing the number of parameters. Rather than collecting features solely on the last layer, early and mid-level bottleneck intermediate outputs were pooled by multi-scale global average and concatenated. The representation of fused features is: 4 implicitly, ∥ implicitly means concatenation. This guarantees that low level texture detail (e.g., spot shapes and edges) and high-level semantic detail (patterns of diseases in general) all play a role in the final feature vector. After applying the above modifications, the extracted feature vectors from MobileNetV2 were L2-normalized to maintain consistency: 5 Resnet101 based features ResNet101 was used as a second backbone feature extractor to supplement MobileNetV2 with more interesting and deeper representations. ResNet101 has 101 layers in the form of residual blocks such that each block has an identity shortcut connection to allow the gradient flow as well as avoiding vanishing gradients in extremely deep networks. The fundamental building block of a residual block is: 6 where x is the block input, is the learned residual mapping that is expressed as convolutional layers stacked on top of each other with weighted parameters. and the addition is the shortcut connection. This architecture allows the network to acquire low and high-level information with competence and can thus be used more appropriately in learning the trendy patterns of diseases among banana leaves. Figure 4 shows modified ResNet101 architecture for deep feature extraction from banana leaf images. Fig. 4. Open in a new tab Modified ResNet101 architecture for feature extraction. In order to further improve the baseline ResNet101 to the task of classification of banana leaf disease, three significant changes were implemented: Every residual block was also adjusted to incorporate an attention gate that selectively focuses on disease-relevant spatial areas. Attention-modulated residual block may be written as: 7 where α ∈ [0,1] is a spatial attention coefficient learned dynamically for each feature map. This ensures that lesions and infected regions contribute more strongly to the learned features. The features were not only extracted at the last layer but also at the intermediate layers (conv3_x, conv4_x, conv5_x) and included multi-scale details of banana leaf diseases including small lesions, elongated streaks, and general leaf degradation. The model of aggregation is developed as: 8 and GAP is global average pooling and ∥ is concatenation. Such a multi-level fusion enables both fine-grained and semantic information to be maintained in the final representation. In order to enhance the discriminative ability, the standard global average pooling was substituted with the Generalized Mean (GeM) pooling, which is: 9 that N is the number of spatial locations, are responses to features, and p is a trainable parameter. GeM pooling, in contrast to average pooling, can dynamically adjust its aggregation, paying more attention to the high-activation areas, e.g., disease spots. The extracted multi-scale GeM-pooled feature vectors from ResNet101 were normalized using L2-normalization to ensure stability across samples: 10 These enriched features capture a comprehensive set of disease characteristics, from local lesion details to overall infection patterns, making ResNet101 a powerful complement to MobileNetV2 in the proposed feature extraction framework. Attention weighted feature fusion MobileNetV2 and ResNet101 learn to extract different but complementary features of the images of banana leaves. MobileNetV2 is small and capable of capturing finer lesion features with computational efficiency whereas ResNet101 offers more contextual information using its residual learning framework. In order to leverage both models, a feature fusion plan was developed such that the resulting representation is sensitive to local detail as well as at the semantic level. 11 During the first stage, feature vectors of both networks were normalized and added together to create a joint representation. In particular, the MobileNetV2 1280-dimensional vector and the ResNet101 2048-dimensional vectors were fused to form a 3328-dimensional fused one. Such a concatenation is guaranteed to retain the richness of both models, but it also opens the risk of overlap in the feature space which can adversely affect the efficiency of the classifier. To overcome this constraint, a modification was proposed which is the use of an attention-weighted fusion mechanism. Rather than averaging the two sets of features, a learnable coefficient was used to dynamically trade-off the contribution of MobileNetV2 and ResNet101 in training. By doing this, the mechanism facilitates the fusion process to focus on most applicable features on a specific disease category. As one example, in instances where finer lesion textures are more discriminatory, MobileNetV2 features are prioritized and those diseases are characterized by a larger change in leaf pattern, the ResNet101 feature is more heavily weighted. 12 where β ∈ [0,1] is a trainable coefficient that dynamically balances contributions from MobileNetV2 and ResNet101 during training. This ensures that the final representation emphasizes the most informative features depending on the dataset and disease type. Lastly, dimensionality reduction was also done to the fused vector using Principal Component Analysis (PCA) to further enhance efficiency. PCA, having retained 95% of the variance, has not only reduced the computational complexity, it also assists in removing the redundant information, thus, the resultant feature space is more compact and discriminative. The result is a strong representation with the complementary strengths of the two architectures to achieve better classification performance in a wide range of conditions. In order to guarantee the reproducibility of the experiment, a constant random seed of 42 was used in all experiments. Principal Component Analysis (PCA) was used and 50 components were used which captured 95% of the total variance in the data. These measures bring about transparency and enable others to reproduce the experiment by other researchers. Proposed deep neural network In the present study pre-trained models were modified by freezing and altering the final classification layers. The completely connected layers of the base networks were substituted with redesigned ones depending on the set of data. Data preprocessing and augmentation methods were used to enhance robustness and generalization prior to the images (224 × 224 × 3) being fed to the pre-trained networks to extract features automatically. The tailored classification head, as shown in Fig. 1 , is made up of a series of layers: the first layer is a batch normalization layer to normalize the input distribution, then there is a flattening layer to transform the feature maps into one-dimensional vectors. To increase flexibility of the model, two layers of 512 units with ReLU activation functions were added, and another layer of batch normalization was added to stabilize and speed up training. A dropout rate of 0.2 was added to minimize the overfitting, randomly deactivating neurons throughout the training. Lastly, the output layer consisted of two neurons whose activation function was a Softmax, which related to the categorization between banana leaf diseases. Experimental setup All the experiments were performed using a workstation with the NVIDIA GPU (e.g., RTX 3080 with 10 GB VRAM), Intel i7 processor, and 32 GB RAM with the Linux-based operating system. It was implemented in Python with TensorFlow and PyTorch frameworks alongside the supporting libraries of scikit-learn and OpenCV to preprocess and evaluation. The input images were scaled to 224 × 224 × 3 and normalized and then were fed into the models. To reduce the issue of class imbalance and improve generalization, data augmentation techniques of rotation, horizontal and vertical flipping, scaling and brightness adjustments were used. The training was done with the Adam optimizer, an initial learning rate of 0.0001, a 32-sized batch, and terminated by early stopping to avoid overfitting with a maximum number of 100 epochs. The proposed models, i.e., modified MobileNetV2, ResNet101 and the fusion framework were trained and tested in the same conditions to allow fair comparison. Stratified 80:20 train-test splits were used to evaluate model performance, and 5-fold cross-validation was performed to guarantee robustness and reproducibility. Measures of evaluation accuracy, precision, recall, F1-score, and confusion matrices were used to give a detailed evaluation of the performance in classification. A deployment workflow figure has been included and accompanied by a deployment feasibility table. The hybrid model has 12.7 M parameters, which is much less than the traditional architectures, and hence, it is applicable in real-time mobile inference. This system pipeline illustrates image capture, preprocessing, inference within the device, and disease prediction (see Table 2 ). Table 2. Model suitability for mobile deployment. Model Parameters (M) Inference suitability Deployment target VGG16 138.4 Not suitable Server ResNet101 44.5 Limited Edge GPU ViT 86.5 Not suitable Server Proposed hybrid 12.7 Highly suitable Mobile / Edge Open in a new tab Evaluation policy The developed models have been tested within a standardized experimental set up to achieve reliability and equity. The data sets are classified into stratified training, validation and testing sets with the same class balance. Besides this, a 5-fold cross-validation approach has been used to prevent overfitting and ensure strength in splits 35 , 36 . Standard classification metrics, such as accuracy, precision, recall and F1-score have been used to quantify performance and are defined as: 13 14 15 16 where TP, FP, TN and FN are the true positives, true negatives, false positives and false negatives, respectively. Confusion matrices, to display a detailed account of misclassifications and per-class predictions, have been created. In addition to this, the Receiver Operating Characteristic (ROC) curve and Area Under the Curve (AUC) have been applied to determine the discriminative capability of the models in various threshold settings. Results and discussion This section has reported the experimental findings obtained through the proposed framework and it has also discussed the implications in detail. The modified MobileNetV2, ResNet101, and the fusion-based model performance have been compared with each other based on several evaluation measures such as accuracy, precision, recall, F1-score and AUC. Analysis of the findings has been done quantitatively, using metric comparisons and statistical validation, and qualitatively by way of confusion matrices and visual inspection of correctly and incorrectly classified samples. Moreover, the effects that occur when preprocessing, model modifications, and feature fusion are used on classification performance have been mentioned. The findings have also been presented in relation to current state of the art methods with emphasis especially on the strength of the proposed strategy in cases of class imbalance and complicated field conditions. Results sections are described in the following sections: Performance evaluation of proposed hybrid model. Comparison with different recent models. Ablation experiments. Comparison with state-of-the-art methods. Performance analysis of proposed hybrid model In Fig. 5 , the training and validation performance of the proposed hybrid model on each of the three single banana disease datasets and the combined dataset is shown. In the case of Banana and Banana Leaf data (top-left), the model has converged fast with training and validation accuracy approaching close to 100% after the initial 10 epochs and the losses reduced sharply to close to zero indicating successful information learning and generalization. Equally, on the Banana Disease Recognition dataset (top-right), the model has had the strongest performance with training and validation curves overlapping nearly at 100%, which implies that there is little overfitting and that the optimization is stable. The same trend has been identified with the Banana LSD dataset (bottom-left) where the accuracy curves peaked at saturation at 98% with a validation performance following closely behind the training accuracy. Excellent convergence has also been exhibited in the combined dataset (bottom-right), albeit with slightly increased initial variance because of the integration of the various samples; nevertheless, the model has settled very fast, with almost ideal accuracy and the minimal loss observed in all folds. All these findings corroborate the fact that the hybrid model not only has generalized well to personal datasets but has also shown high robustness to generalized aggregated data and this highlights its versatility and applicability to real-world banana disease detection problems. Fig. 5. Open in a new tab Training and validation performance of proposed hybrid model on different datasets. The testing confusion matrices of the proposed hybrid model are shown in Fig. 6 on both individual and combined banana leaf disease datasets. In the case of Banana and Banana Leaf (top-left) dataset, the model has made almost perfect classification with some few cases of misclassification evident between Sigatoka and Wilt classes. Most of the classes such as Chewing Insect, Healthy, and Skipper have been determined precisely, and this proves the strength of the model on this dataset. Similar high results have been achieved in the Banana Disease Recognition dataset (top-right), with some confusion between Sigatoka and Wilt, which can be explained by the visual similarity of the disease symptoms in these classes. Fig. 6. Open in a new tab Testing confusion matrices of proposed hybrid model on different datasets. The Banana LSD data (bottom-right) has similarly shown a high classification accuracy in all four disease categories but some overlaps between Cordana and Sigatoka and Pestalotiopsis and Sigatoka have been noted in a few cases. Such misclassifications are not surprising because these leaf spot diseases have similar images, and it can be difficult even to a human professional. The overall dataset (bottom-left) has confirmed the generalization capacity of the model on a variety of diseases. Although there has been unwavering success on the recognition of Healthy, Insect, and Wilt using the model, there is a slight confusion when distinguishing between Sigatoka and Pestalotiopsis and Cordana. Comprehensively, the confusion matrices have indicated that the proposed hybrid model has obtained reliable classification in all the datasets with few confusions between the disease classes which are inherently highly similar visually. These findings support the high generalization abilities of the model to a wide range of datasets and give good per-class results. Figure 7 presents ROC curves of the proposed hybrid model on four banana-related datasets, which demonstrate that it is effective with the classification of various disease types. With the Banana and Banana Leaf Dataset (top-left), the model has near-perfect performance, with AUC values of 1.000 on the majority of classes, and very high values (0.9988–9997) on the rest, which means that the model is confident in the detection of chewing insect damage, sigatoka, wilt, and skipper. In the case of the Banana Disease Recognition Dataset (top-right), the ROC curves are nearly on the top-left side, with all the classes having in the Banana Disease Recognition Dataset (top-right) showing a perfect AUC (1.000), indicating an optimal discrimination of the healthy and diseased leaves. As per the data that is more diverse in the Combined Dataset (bottom-left), the model has maintained a high performance with AUC values approaching the value of 1.000 at six classes, which is indicative of a robust model even in mixed conditions. Lastly, the Banana LSD (Leaf Spot Diseases) Dataset (bottom-right) has a little bit lower but still very high AUC values (0.997–0.998), which implies a little variation in classification yet the overall excellent generalization to disease-specific cases. Generally, the figure illustrates that the hybrid model is nearly perfect in its classification activities in various datasets, which proves its efficacy and versatility in banana disease recognition activities. Fig. 7. Open in a new tab Testing ROC Curves of proposed hybrid model on different datasets. In order to counter possible overfitting issues related to almost perfect accuracy and the values of the areas of the receiver using the cost-effective models, the generalization capability of the proposed model was thoroughly analyzed using testing on various independent and non-homogeneous datasets. The model was always stable in its performance across these unseen datasets, which means that it did not overfit in the characteristics of the dataset but acquired robust and transferable feature representations. Also, conventional regularization techniques and validation measures were used in training to minimize risks of overfitting. These findings can be taken to indicate that the suggested framework is valid and can be successfully applied in practice, though further efforts will be devoted to the validation of the framework by other large-scale real-field datasets. Comparison with other deep learning models A comparative study of the suggested hybrid model with ten popular deep learning architectures to classify banana disease is provided in Table 3 . Conventional designs like the VGG16 and VGG19 demonstrated a lower level of accuracy (94.8 and 95.2) and a great quantity of free parameters (138 M and higher) to train, resulting in longer convergence periods. Advanced models, such as ResNet50, ResNet101, DenseNet121, and InceptionV3, have provided moderate gains on accuracy (96 97) with a smaller computational cost. MobileNetV2 and EfficientNet-B0 are two of the lightweight architectures that have been shown to achieve performance competition by reaching above 97% accuracy with less parameters (3.4 M and 5.3 M, respectively), and less time (1820 epochs) to converge. ViT has also shown good performance (97.8% accuracy), however it has many parameters (86.5 M) and it requires a long time to converge thus not efficient to be used practically. Table 3. Comparative analysis of proposed method with other deep learning models. Model Accuracy (%) Precision (%) Recall (%) F1-score (%) Trainable parameters (M) Convergence time (Epochs) VGG16 94.8 94.5 94.2 94.3 138.4 35 VGG19 95.2 95.0 94.7 94.8 143.7 37 ResNet50 96.1 95.9 95.7 95.8 25.6 28 ResNet101 96.8 96.6 96.3 96.4 44.5 32 DenseNet121 97.0 96.8 96.6 96.7 8.0 26 InceptionV3 96.5 96.3 96.0 96.1 23.9 30 Xception 96.9 96.7 96.5 96.6 22.9 29 MobileNetV2 97.2 97.0 96.8 96.9 3.4 18 EfficientNet-B0 97.6 97.5 97.2 97.3 5.3 20 Vision transformer (ViT) 97.8 97.6 97.4 97.5 86.5 34 Proposed Hybrid (MobileNetV2 + ResNet101, Attention + PCA Fusion) 98.28 98.18 98.77 98.43 12.7 15 Open in a new tab Conversely, the proposed hybrid model that incorporates MobileNetV2 and ResNet101, along with attention mechanisms and feature fusion via PCA, has provided the best results of 98.28% accuracy, 98.18% precision, 98.77% recall, and 98.43% F1-score. Notably, this high precision has been achieved using just 12.7 M parameters and the shortest convergence time (15 epochs) reflects a good performance/computational efficiency trade off. These findings underscore the usefulness of the suggested structure in capitalizing lightweight as well as deep feature extraction coupled with counteracting the problem of overfitting and convergence in complicated banana disease classification problems. Ablation experiments The findings of a study on ablation that investigated the effect of fine-tuning depth on the model performance in terms of accuracy and F1-score are presented in Fig. 8 . With a frozen backbone, the model is associated with 94.5% accuracy and F1-score of 0.9201, which means a lack of flexibility. Semi-fine-tuning increases the accuracy to 95.23% and F1-score to 0.9432 which is a better feature refinement. Interestingly, full fine-tuning has a slightly higher accuracy, 96.01% but has a lower F1-score of 0.9152, which implies that it may be overfitting or imbalance handling. Hybrid fine-tuning to selectively fine-tune certain layers with frozen layers gives the highest accuracy of 98.28% and the maximum F1-score of 0.9843. These results demonstrate that hybrid fine-tuning is the best tradeoff approach in terms of generalization and specialization, as it works better than partial or full fine-tuning in the sense of utilizing pre-trained knowledge and adapting to task-related characteristics. Fig. 8. Open in a new tab Effect of fine tuning on model performance. Hybrid fine-tuning is an extension of the freezing technique where the low-level layers of the network (representing general features such as edges, shapes, and textures) are frozen, and the deeper task-specific layers are learned with discriminative high-level features. The method maintains the strength of pre-trained representations and customizes the model to novel domain-specific features without overfitting, which may happen during full fine-tuning. This selective adjustment enables this model to optimize the trade-off resulting in a higher level of accuracy and F1-score as shown in Fig. 5 than other fine-tuning approaches. Figure 9 illustrates that input image resolution has an impact on model performance in accuracy and F1-score. However, the results indicate a clear rising pattern: performance is the weakest in 128 × 128 with an accuracy of 91.2. and F1-score of 89.0, probably because there is no detail in low-resolution pictures. The accuracy also increases to 93.5% and 95.1 as it reaches the resolution of 160/160, 192/192 with the respective F1-score of 90.5 and 94.3, suggesting that the finer disease patterns are recognized better. This is highest at 224 × 224 resolution where the accuracy and F1-score are 98.2 and 98.4%, respectively, which indicates that increased resolution is more informative to yield more dependable classification. These results point to the relevance of input resolution in improving the performance of the model which means that the bigger the image size, the more discriminative features are resolved by the network, and ultimately, the model will have a higher accuracy to detect the disease. Fig. 9. Open in a new tab Effect of input image resolution on model performance. Ablation study was used to determine the effect of different data augmentation methods on the performance of the model in terms of accuracy and F1 score. As Fig. 10 shows, data augmentation is an important factor in enhancing the generalization of models. The model was the least accurate (92.00%) and F1 score (89.00) without any augmentation. Geometric transformations and added noise of color were incrementally improved to a maximum of 99.0% accuracy and 98.8% F1 score. The overall performance of MixUp was the highest among all the methods considered, with 99.2% accuracy and 99.0% F1 score, which is why it can be said to be effective at improving the precision and recall. More complicated augmentation techniques like AutoAugment and Full Settings did not achieve a high performance, but its performance was not lower than that of MixUp, which is why simpler ones can be even more effective in the right settings. The results highlight the need to pick more complex and effective augmentation strategies to achieve maximum model robustness. Fig. 10. Open in a new tab Effect of data augmentation methods on model performance. The results obtained in the ablation studies indicate that the attention mechanism and feature fusion using PCA are independent of each other and that they perform better than the baseline model as shown in Table 4 . The attention module is more beneficial with more gain of the module by increasing the learning of discriminative features with adaptive channel weights whereas PCA fusion is complementary in that it reduces redundancy of features and gives better generalization. It is worth noting that the combination of attention and PCA 1 has the best performance in all the evaluation metrics, which is the effect of synergy. Through this analysis, the need for each component is confirmed and the reason why they must be incorporated in the proposed hybrid framework justified. Table 4. Ablation study evaluating the individual and combined effects of the attention mechanism and PCA-based feature fusion on model performance across standard classification metrics. Configuration Attention module PCA fusion Accuracy (%) Precision (%) Recall (%) F1-score (%) Baseline model ✗ ✗ 91.8 91.2 91.5 91.3 Baseline + Attention ✓ ✗ 93.1 92.8 92.9 92.8 Baseline + PCA fusion ✗ ✓ 92.6 92.1 92.3 92.2 Proposed hybrid model ✓ ✓ 98.28 98.43 98.2 97.1 Open in a new tab Comparison with state-of-the-art methods Table 5 offers a comparative study of the proposed hybrid framework and some of the state-of-the-art (SOTA) approaches that have been reported over the last few years. Conventional deep CNN models like those used in 35 performed better with 96.22, 87.50, respectively but poorer generalization and poorer robustness performance as measured by their F1-scores and their AUROC values. Banana-specific models, like BananaLeaf-Net 34 have shown improvement with accuracy of 97.00 and F1-score of 96.00, but its performance is still worse than that of the AUROC (0.9400). Transformer based models, such as the ViT model 22 , also report a competitive accuracy (96.88) and F1-score (96.21) but their values of AUROC are lower than the proposed model, showing difficulties to make sure and hard decisions between challenging classes. More sophisticated algorithms, such as KVA and GR-ARNET 32 and ensemble deep learning 24 demonstrated good performance with the accuracy being over 96.9% and with an AUROC of 0.9836 respectively. Nonetheless, these methods are quite efficient, but very costly in calculation and longer to converge. Likewise, ResNet50 36 and VGG19 37 offer decent performance but their outcomes indicate that single-network systems tend to fail when trying to find the intricate changes in the diseases of banana leaves. The reason behind the implementation of an attention-enhanced hybrid network was to attain high accuracy in a robust framework in challenging field environments as opposed to using handcrafted feature engineering that can be dataset specific and can be vulnerable to noise. Table 4 demonstrates that the accuracy of a typical CNN using hand-optimized features was 94.1%, and the accuracy of the proposed attention-based hybrid model was 98.28% with computational efficiency (12.7 M parameters). The attention mechanisms enabled the model to pay attention to the lesion-specific regions, which is hard to realize consistently with manual feature engineering. Table 5. Comparative analysis of proposed method with SOTA methods. Ref, Year Method Accuracy F1-score AUROC Parameters (millions) 37 , 2024 KVA and GR-ARNET 96.98 94.25 0.9668 ~ 18.2 38 , 2024 CNN framework 96.22 95.36 0.9465 ~ 9.5 39 , 2025 BananaLeaf-Net 97.00 96.00 0.9400 ~ 11.3 22 , 2024 Vision transformer (ViT) model 96.88 96.21 0.9653 ~ 86.5 40 , 2025 CNN framework 87.50 91.52 0.9068 ~ 7.8 29 , 2025 Ensemble deep learning 98.02 95.64 0.9836 ~ 45.0 41 , 2025 ResNet50 96.25 97.85 0.9768 25.6 42 , 2025 VGG19 88.90 88.58 0.9048 143.7 Proposed Hybrid (MobileNetV2 + ResNet101, attention + PCA fusion) 98.28 98.43 0.9997 12.7 Open in a new tab Contrastingly, the hybrid model proposed has the greatest overall performance, accuracy of 98.28, F1-score of 98.43, and the AUC of 0.9997. This almost ideal AUROC demonstrates the capability of the model to differentiate diseased and healthy samples with great confidence. It is important to note that the proposed framework not only outperforms ensemble models but also has fewer trainable parameters and converts more quickly, hence could be considered accurate and computationally efficient. These findings affirm the strength and excellence of the hybrid MobileNetV2- ResNet101 architecture with attention and PCA fusion when dealing with the problem of banana disease detection. The research has certain limitations associated with the usage of curated sets, RGB-only images, and the absence of large on-device benchmarking. Such restrictions are a good incentive in the future to work with real-field data, multimodal sensing, and optimization of mobile hardware. In order to prove the effectiveness of the proposed model in more detail, we also tested the performance of its inference with desktop and mobile hardware. The model can run at an average inference speed of 75 FPS on a standard NVIDIA GTX 1080 GPU and used 220 MB of memory and reached 18 FPS with 150 MB used on a mid-range mobile device (Snapdragon 8 series). These findings substantiate the fact that the model can be used in real-time applications and deployed in edge devices in addition to being computationally efficient with a low parameter count (12.7 M) and with rapid convergence (15 epochs) in addition to being fast. Conclusion and future work This paper suggests a hybrid deep learning model that comprises MobileNetV2 and ResNet101 with attention-enhanced bottlenecks, dilated convolutions, multi-scale pooling, and feature fusion using PCAs to effectively and efficiently classification of banana leaf disease. The framework has been strictly tested on three benchmark datasets, such as Banana and Banana Leaf, Banana Disease Recognition and Banana LSD with a well-crafted preprocessing pipeline. The experimental outcomes have shown that the given method works much better than ten popular deep learning structures and some state-of-the-art (SOTA) methods. The hybrid model, with accuracy of 98.28, F1-score of 98.43 and AUROC of 0.9997 is proven to be highly accurate and computationally efficient as it only needs 12.7 M parameters, converging with only 15 epochs, thus making it appropriate in the real-world application in the precision agriculture system. Although the suggested framework has been delivered to promise, it still has some weaknesses that leave open for further studies. To begin with, the experiments that have been conducted to date have been based on curated datasets; to achieve a wider applicability, it will be required to apply the method to large-scale field images captured in diverse lighting, occlusion and noise levels. Second, generative models, e.g., GANs or diffusion models, may be used to tackle the issues of data scarcity and class imbalance, which would also improve robustness. Third, the framework can be further simplified to a lightweight and edge-optimized implementation to be used in real-time on a mobile system, and IoT-based monitoring systems by farmers. In addition, it might be possible to involve multimodal sources of data, e.g., hyperspectral imaging, soil health indicators, and climatic conditions, to make disease prediction more robust than using leaf symptoms. Lastly, it can also be investigated in the future that self-supervised learning and domain adaptation techniques can be used to improve transferability between various varieties of crops and geographical locations. Author contributions Hongyi Du conceived the study, designed the methodology, supervised the research process, and prepared the initial draft of the manuscript. Hua Yang contributed to data curation, software implementation, formal analysis, and participated in reviewing and editing the manuscript. Zhidong Zhang was responsible for validation, visualization, experimental investigations, and analysis of the results. Jinjing Wang provided resources, assisted in dataset preparation, contributed to funding acquisition, and offered technical support throughout the study. Funding This research is not funded by research groups or funding agencies. Data availability The datasets used in this study are publicly available through Mendeley Data. The sources are: Banana and Banana Leaf Dataset for Classification and Disease Detection (DOI: 10.17632/5nfjzntwd8.2, https:/data.mendeley.com/datasets/5nfjzntwd8/2), Banana Disease Recognition Dataset (DOI: 10.17632/79w2n6b4kf.1, https:/data.mendeley.com/datasets/79w2n6b4kf/1), and Banana Leaf Spot Diseases (BananaLSD) Dataset (DOI: 10.17632/9tb7k297ff.1, https:/data.mendeley.com/datasets/9tb7k297ff/1). Declarations Competing interests The authors declare that they have no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Awuchi, C. G. HACCP, quality, and food safety management in food and agricultural systems. Cogent Food Agric. 9 (1), 2176280. 10.1080/23311932.2023.2176280 (2023). [ Google Scholar ] 2. Rashid, J., Khan, I., Ali, G., Rehman, S. U. & Alturise, F. Real-time multiple guava leaf disease detection from a single leaf using hybrid deep learning technique. Comput. Mater. Contin . 74 (1), 1235–1257 (2023). [ Google Scholar ] 3. Raja, N. B. & Selvi Rajendran, P. An efficient banana plant leaf disease classification using optimal ensemble deep transfer network. J. Exp. Theor. Artif. Intell. 37 (4), 585–608. 10.1080/0952813X.2023.2241867 (2025). [ Google Scholar ] 4. Yumang, A. N., Baguisi, J. M., Buenaventura, B. R. S. & Paglinawan, C. C. Detection of Black Sigatoka Disease on Banana Leaves Using ShuffleNet V2 CNN Architecture in Comparison to SVM and KNN Techniques, in 15th International Conference on Computer and Automation Engineering (ICCAE) , 281–286. 10.1109/ICCAE56788.2023.10111367 (2023). 5. Madhusudhan, C. K., Mahendra, K., Raghavendra, N., Revanasiddappa, M. & Faisal, M. Corrosion-resistant polypyrrole-banana carbon (PPy-BC) nanocomposites for protection against electromagnetic interference: a green approach. J. Mater. Sci. Mater. Electron. 33 (3), 1366–1382. 10.1007/s10854-021-07466-1 (2022) [ Google Scholar ] 6. Arslan, M. et al. A deep features based approach using modified ResNet50 and gradient boosting for visual sentiments classification, in IEEE 7th International Conference on Multimedia Information Processing and Retrieval (MIPR) , 239–242. 10.1109/MIPR62202.2024.00045 (2024). 7. Ahmed, M. & Ahmed, A. Palm tree disease detection and classification using residual network and transfer learning of inception ResNet. PLOS ONE . 18 (3), e0282250. 10.1371/journal.pone.0282250 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Akram, A., Tariq, A., Ali, M. S., Tariq, M. U. & Altaf, A. Recognizing facial expressions across cultures using gradient features. J. Innov. Comput. Emerg. Technol. 3 , 1 (2023). [ Google Scholar ] 9. Huang, Y. et al. VGG16-AttnNet: A banana leaf lesion recognition model combining VGG16 and self-attention mechanism, in 5th International Conference on Big Data & Artificial Intelligence & Software Engineering (ICBASE) , S745–748. 10.1109/ICBASE63199.2024.10762417 (2024). 10. Saini, A. Efficient banana leaf disease classification using Vision Transformer (ViT) model, in 4th International Conference on Technological Advancements in Computational Sciences (ICTACS) , 176–180. 10.1109/ICTACS62700.2024.10841002 (2024). 11. Sahu, P., Singh, A. P., Chug, A. & Singh, D. A Systematic literature review of machine learning techniques deployed in agriculture: A case study of banana crop. IEEE Access 10 , 87333–87360. 10.1109/ACCESS.2022.3199926 (2022). [ Google Scholar ] 12. Chong, M. Z., Taujuddin, N. S. A. M., Sari, S., Tukiran, Z. & Ghani, A. R. A. Banana leaf disease classification using SqueezeNet, AlexNet and MobileNet. J. Electron. Volt Appl. 6 (1), 48–58 (2025). [ Google Scholar ] 13. Rajalakshmi, N. R. et al. Early detection of banana leaf disease using novel deep convolutional neural network. J. Data Sci. Intell. Syst. 3 (3), 192–199. 10.47852/bonviewJDSIS42021530 (2025). [ Google Scholar ] 14. Arun, A., Sharma, S., Singh, B. & Hazra, T. Identification of plant species using convolutional neural network with transfer learning. J. Phytopathol. 173 (1), e70032. 10.1111/jph.70032 (2025). [ Google Scholar ] 15. Yan, K., Shisher, M. K. C. & Sun, Y. A. Transfer learning-based deep convolutional neural network for detection of fusarium wilt in banana crops. AgriEngineering 5 (4), 2381–2394 10.3390/agriengineering5040146 (2023). [ Google Scholar ] 16. Nagachandrika, B., Prasath, R. & Praveen Joe, I. R. An automatic classification framework for identifying type of plant leaf diseases using multi-scale feature fusion-based adaptive deep network. Biomed. Signal. Process. Control . 95 , 106316. 10.1016/j.bspc.2024.106316 (2024). 17. Pham, T. C., Nguyen, T. N. & Nguyen, V. D. Ambiguity-aware semi-supervised learning for leaf disease classification. Sci. Rep. 15 (1), 14070. 10.1038/s41598-025-95849-3 (2025). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Saraswathi, M. S. et al. Production of quality planting material in commercial banana cvs. Rasthali (Silk, AAB) and Neypoovan (Neypoovan, AB) through farmer-friendly macropropagation technique and their field evaluation. South. Afr. J. Bot. 167 , 410–418. 10.1016/j.sajb.2024.02.047 (2024). [ Google Scholar ] 19. Nagu, B., Kaur, G. & Shanmugavelu, M. & MRM, V. Prediction of banana leaf disease by intelligent algorithm through digital image using neural network technique. Multimed. Tools Appl. 84 (28), 34249–34265. 10.1007/s11042-024-20529-9 (2025). [ Google Scholar ] 20. Yan, W., Feng, Q., Yang, S., Zhang, J. & Yang, W. Heterogeneous metric fusion network-based few-shot learning for crop disease recognition. Agronomy 13 (12), 2876. 10.3390/agronomy13122876 (2023). [ Google Scholar ] 21. Thirumeninathan, V., Vijayalakshmi, S. & Palathara, S. T. Enhancing banana cultivation: Disease identification through CNN and SVM analysis for optimal plant health, in International Conference on Trends in Quantum Computing and Emerging Business Technologies , 1–6. 10.1109/TQCEBT59414.2024.10545264 (2024). 22. Saini, A. Efficient banana leaf disease classification using vision transformer (ViT) model, in 4th International Conference on Technological Advancements in Computational Sciences (ICTACS) , 176–180. 10.1109/ICTACS62700.2024.10841002 (2024). 23. Rameshkumar, R., Arunkumar, G., Devi, B. A. & Pandi, S. S. Enhanced plant disease classification via wild horse optimizer and convolutional attention-bidirectional long short term memory. Signal Image Video Process. 19 (12), 1010. 10.1007/s11760-025-04556-z (2025). [ Google Scholar ] 24. Senthil Pandi, S., Reshmy, A. K., Muruganandam, S. & Manju, I. Hybrid Crossover oppositional firefly optimization for enhanced deep transfer learning in plant leaf disease classification. J. Crop Health . 77 (5), 146. 10.1007/s10343-025-01205-w (2025). [ Google Scholar ] 25. Sellam, V., Kannan, N., Pandi, S., Manju, I. & S. & Enhancing sustainable agriculture using attention convolutional bidirectional Gated recurrent based modified leaf in wind algorithm: Integrating AI and IoT for efficient farming. Sustain. Comput. Inf. Syst. 47 , 101160. 10.1016/j.suscom.2025.101160 (2025). [ Google Scholar ] 26. S, S. P., Pounambal, M., Vellingiri, J. G. J. & K, A. Advancing agricultural sustainability with gradient weighted Densenet-201 model for accurate detection of plant leaf diseases. Environ. Dev. Sustain. 10.1007/s10668-025-06314-0 (2025). [ Google Scholar ] 27. Senthil Pandi, S., Reshmy, A. K., Elangovan, D. & Vellingiri, J. AI-driven environmental monitoring for hydroponic agriculture: ExCNN-LFCP approach. Earth Sci. Inf. 18 (1), 73. 10.1007/s12145-024-01516-y (2024). [ Google Scholar ] 28. Mathew, D., Kumar, C. S. & Cherian, K. A. Integration of nondecimated quaternion wavelet transform and neighborhood texture patterns for disease classification in banana (Musa spp.) foliage. Multimed. Tools Appl. 82 (24), 37327–37349 10.1007/s11042-023-14869-1 (2023). [ Google Scholar ] 29. Rehman, A. et al. An intelligent deep augmented model for detection of banana leaves diseases. Microsc Res. Tech. 88 (1), 53–64. 10.1002/jemt.24681 (2025). [ DOI ] [ PubMed ] [ Google Scholar ] 30. Das, U., Azam, S. & Al Kafi, M. A. Banana and banana leaf dataset for classification and disease detection, 2 10.17632/5nfjzntwd8.2 (2025). 31. Mafi, M. M. H. M., Sifat, R. M., Moazzam, M. G. M. & Uddin, M. S. Banana disease recognition dataset. Mendeley Data 10.17632/79w2n6b4kf.1 (2023). 32. Arman, S. E., Baki Bhuiyan, M. A., Abdullah, H. M., Islam, S., Chowdhury, T. T. & Hossain, M. A. Banana leaf spot diseases (BananaLSD) dataset for classification of banana leaf diseases using machine learning. Data Brief 50 109608. 10.17632/9tb7k297ff.1 (2023). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 33. Haq, M. K. et al. Enhancing banana leaf spot disease classification using dense Mobilenet V2, in International Conference on Advances in Computing, Communication and Applied Informatics (ACCAI) , 1–6. 10.1109/ACCAI61061.2024.10602228 (2024). 34. Shu, Y., Zhang, J., Wang, Y. & Wei, Y. Fruit freshness classification and detection based on the ResNet-101 network and non-local attention mechanism. Foods 14 (11), 1987 10.3390/foods14111987 (2025). [ DOI ] [ PMC free article ] [ PubMed ] 35. Akram, A. et al. Recognizing breast cancer using edge-weighted texture features of histopathology images. Comput. Mater. Contin . 77 (1), 1081–1101 10.32604/cmc.2023.041558 (2023). [ Google Scholar ] 36. Akram, A., Rashid, J., Jaffar, A., Hajjej, F., Iqbal, W. & Sarwar, N. Weber law based approach for multi-class image forgery detection. Comput. Mater. Contin . 78 (1), 145–166. 10.32604/cmc.2023.041074 (2024). [ Google Scholar ] 37. Deng, J., Huang, W., Zhou, G., Hu, Y., Li, L. & Wang, Y. Identification of banana leaf disease based on KVA and GR-ARNet. J. Integr. Agric. 23 (10), 3554–3575 10.1016/j.jia.2023.11.037 (2024). [ Google Scholar ] 38. Shukla, A. N., Joshi, K., Singh Yadav, A. P., Kukreja, V. & Mehta, S. Precision phytopathology in agriculture: A federated learning CNN framework for banana leaf disease classification, in 11th International Conference on Reliability, Infocom Technologies and Optimization (Trends and Future Directions) (ICRITO) , 1–7 10.1109/ICRITO61523.2024.10522433 (2024). 39. Sum, A. S. I., Saha, A. K., Nur, M. K. & Hasan, K. T. BananaLeaf-Net: A proposed deep learning model for accurate classification of banana leaf diseases, in Machine vision in plant leaf disease detection for sustainable agriculture (eds Mridha, M. F. & Dey, N.) 41–56 (Springer Nature, 2025). [ Google Scholar ] 40. Ancheta, J. B., Santos, L. Q., De Jesus, L. C. M., David, J. V., Avelino, A. M. & Grande, M. E. A. Development of a classification model for banana leaf disease using google teachable machine, in International Conference on Artificial Intelligence in Information and Communication (ICAIIC) , 0258–0263. 10.1109/ICAIIC64266.2025.10920850 (2025). 41. Singla, S., Jagadish, S. S., Saxena, M., Mahajan, S., Singh, A. R. & Gupta, R. Automated detection of banana leaf diseases using ResNet50, in 6th International Conference for Emerging Technology (INCET) 1–5 10.1109/INCET64471.2025.11140293 (2025). 42. Jiménez, N., Orellana, S., Mazon-Olivo, B., Rivas-Asanza, W. & Ramírez-Morales, I. Detection of leaf diseases in banana crops using deep learning techniques. AI 6 (3), 61. 10.3390/ai6030061 (2025). Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The datasets used in this study are publicly available through Mendeley Data. The sources are: Banana and Banana Leaf Dataset for Classification and Disease Detection (DOI: 10.17632/5nfjzntwd8.2, https:/data.mendeley.com/datasets/5nfjzntwd8/2), Banana Disease Recognition Dataset (DOI: 10.17632/79w2n6b4kf.1, https:/data.mendeley.com/datasets/79w2n6b4kf/1), and Banana Leaf Spot Diseases (BananaLSD) Dataset (DOI: 10.17632/9tb7k297ff.1, https:/data.mendeley.com/datasets/9tb7k297ff/1). Articles from Scientific Reports are provided here courtesy of Nature Publishing Group ACTIONS View on publisher site PDF (3.3 MB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Deep learning enables edge deployment for citrus leaf disease recognition in natural orchard scenes.#465298
• Multisource neural network feature map fusion: An efficient strategy to detect plant diseases#367427
• A self-attention-based deep learning model for identifying key genes in insect pupal metamorphosis.#1134805
• Beyond the Visual Spectrum: From RGB-Based Learning to Hyperspectral Intelligence for Plant Disease Detection-Challenges and Opportunities.#318005
• AgriPestDatabase-v1.0: A Structured Insect Dataset for Training Agricultural Large Language Model#183488