The mango is of great agronomic and economic importance, particularly in tropical and subtropical regions, where it is one of the most widely cultivated fruit crops (Zhang et al., 2025). However, mango plants’ productivity and health are persistently threatened by foliar diseases, which impair photosynthetic efficiency and lead to significant reductions in yield and fruit quality. Although traditional approaches to disease diagnosis are still widely used, they often involve subjective assessments, delayed response times and a heavy dependence on expert knowledge. These limitations highlight the urgent need for automated, accurate and scalable diagnostic tools to support the timely and objective management of mango diseases.

In recent years, deep learning (DL) techniques have transformed plant disease detection by enabling complex, hierarchical patterns to be extracted directly from image data. Unlike conventional handcrafted approaches, DL models have demonstrated superior performance in various agricultural vision tasks. However, classifying mango leaf diseases is challenging due to their intricate visual symptoms, similarities between classes and variations within classes that frequently occur across disease types (Shehu et al., 2025; Wei et al., 2025). These complexities necessitate the development of more advanced architectures that can effectively learn both fine-grained local textures and high-level semantic features.

In response to these challenges, we propose MangoLeafCMDF-FAMNet (cross-modal dynamic fusion (CMDF) with feature attention module (FAM) network), a novel hybrid DL framework specifically designed for multi-class mango leaf disease classification. This architecture integrates ConvNeXt and Vision Transformer (ViT) as dual feature extractors, combining the strengths of convolutional inductive biases and transformer-based global attention mechanisms. The proposed model uses a CMDF strategy to combine texture- and semantic-level information into a coherent, enriched feature representation space. To further enhance feature expressiveness, a FAM, inspired by squeeze-and-excitation (SE) networks, is incorporated at each stage. This module adaptively recalibrates channel-wise feature responses to prioritize disease-relevant patterns while suppressing irrelevant background noise.

To comprehensively evaluate the performance of MangoLeafCMDF-FAMNet, we conduct extensive experiments on three publicly available mango leaf disease datasets—MangoLeafDataset1 (8 classes), MangoLeafDataset2 (5 classes), and MangoLeafDataset3 (8 classes). The model’s effectiveness is quantified using multiple evaluation metrics, including classification accuracy (CA), recall (RCL), precision (PRC), Matthews correlation coefficient (MCC), and Cohen’s kappa score. The experimental results consistently demonstrate that the proposed method significantly outperforms conventional baseline models across all datasets, achieving high CA and strong correlation measures.

The main contributions of this work are as follows:

The remainder of this paper is structured as follows: Section 2 provides a thorough review of existing research for plant disease detection. Section 3 outlines the materials and methods employed in this study, providing detailed descriptions of the dataset and the proposed hybrid deep learning and feature selection framework. Section 4 reports the experimental results, alongside performance evaluation metrics and comparative analyses. Section 5 concludes the paper by summarizing the main findings and highlighting potential future research directions. Finally, Section 6 critically discusses the study’s limitations and underlying assumptions, as well as its practical implications for real-world agricultural applications.

Recent advances in computer vision and DL have significantly accelerated the development of automated tools for plant disease diagnosis. convolutional neural networks (CNNs) and transformer-based models, especially ViTs, have emerged as dominant paradigms in plant pathology research due to their ability to extract discriminative spatial and semantic patterns from complex visual data (Chen et al., 2024). However, existing studies on mango leaf disease classification have faced multiple challenges that limit their practical utility and generalizability. Among these studies, Rao et al. (2021) initiated one such effort by leveraging the PlantVillage dataset, applying the AlexNet architecture, and achieving classification accuracies of 0.9900 for grape leaves and 0.8900 for mango leaves. Building on this, Arivazhagan and Ligi (2018) utilized a CNN-based approach on a six-class mango dataset and attained a commendable accuracy of 0.9667. In contrast, Mia et al. (2020) combined artificial neural networks and support vector machines, achieving 0.8000 accuracy in detecting four disease classes and healthy leaves. A more advanced ensemble strategy was introduced by Gautam et al. (2024), who developed a Stacked Ensemble Deep Neural Network that integrated multiple DNNs with classical ML classifiers, yielding a high accuracy of 0.9857 across eight disease classes. Similarly, Saleem et al. (2021a) investigated disease detection using canonical correlation analysis (CCA)-based feature fusion and found cubic SVM to deliver the highest performance. In a follow-up study, Saleem et al. (2021b) further proposed the FrCNet model for lesion segmentation and, after combining it with CCA feature fusion and classification via quadratic and cubic SVMs, achieved 0.9890 accuracy for binary disease-versus-healthy discrimination. Continuing the exploration of CNN variants, Varma et al. (2025) benchmarked several pretrained architectures, with InceptionV3 achieving the highest accuracy at 0.9987. Meanwhile, Patel et al. (2024) introduced a hybrid framework combining Total Variation Filter-based variational mode decomposition with DenseNet121 and VGG-19, achieving 0.9885 CA. This fusion approach notably improved feature interpretability and robustness against noise. Hossain et al. (2024) evaluated ViTs against well-established CNNs and proposed an optimized DeiT-based model, which outperformed all compared methods with a CA of 0.9975. Similarly, Mahmud et al. (2024) proposed DenseNet78, a lightweight variant of DenseNet tailored for mango leaf disease classification, reporting accuracies of 0.9947 for healthy and 0.9944 for diseased leaves. In practical implementations, Puranik et al. (2024) retrained MobileNetV3 on the MangoLeafBD dataset and embedded it within a mobile application, reaching 0.9800 accuracy and enabling real-time field diagnosis. Singh et al. (2024) adopted a transfer learning approach and proposed the DTLD model, which demonstrated strong multi-class classification performance with a peak accuracy of 0.9976 on a 4000-image dataset. Expanding on comparative model analysis, Bairwa et al. (2024) assessed multiple deep networks, finding ResNet50 to deliver the highest accuracy at 0.9912. Complementarily, Pratap and Kumar (2024) designed a CNN-based system incorporating transfer learning from VGG-16, GoogLeNet, MobileNet, YOLOv8, and EfficientNet, enabling effective classification of several mango diseases including Anthracnose, Gall Midge, and Powdery Mildew. Finally, Pahati et al. (2025) trained a Google Teachable Machine model on 4000 annotated images, obtaining an accuracy of 0.9960 and demonstrating high potential for democratized, user-friendly disease recognition platforms.

Most conventional CNN-based models, although capable of capturing local textures, fall short in modeling long-range dependencies—a critical requirement for accurately distinguishing visually similar diseases with subtle morphological variations. Transformer-based methods, while excellent at global context modeling, often lack the inductive biases necessary for fine-grained feature localization. As such, stand-alone CNN or ViT models struggle to deliver optimal performance across varying environmental conditions and disease stages observed in agricultural settings. For example, the study by Alamri et al. (2025) proposed a dual-branch architecture combining ConvNeXt and ViT to detect mango leaf and fruit diseases separately using the MangoLeafBD and SenMangoFruitDDS datasets. Their model achieved promising accuracy levels of 99.87% and 98.40% respectively, demonstrating the value of hybrid architectures in plant disease classification. However, their method did not incorporate any explicit attention mechanism to recalibrate the feature importance across network layers. Furthermore, their architecture processed the outputs of ConvNeXt and ViT using a static fusion approach, which may limit the adaptability of feature interactions during training.

By contrast, our proposed MangoLeafCMDF-FAMNet framework introduces several significant improvements to the original design. Firstly, inspired by SE networks, we incorporated a FAM at each stage to dynamically recalibrate channel-wise features. This enables the model to selectively emphasize disease-relevant information and suppress background noise. Secondly, instead of using a static feature aggregation strategy, our model uses a CMDF mechanism to adaptively combine spatial and semantic cues extracted from ConvNeXt and ViT backbones. This significantly improves the model’s representational richness and robustness.

Moreover, MangoLeafCMDF-FAMNet was rigorously evaluated on three distinct datasets encompassing both 5-class and 8-class classification tasks. Experimental results demonstrated that our model consistently outperforms traditional CNNs, ViTs, and hybrid baselines—including the model by Alamri et al. (2025)—not only in terms of CA but also across comprehensive evaluation metrics such as MCC and kappa. The superior performance of our model, particularly under multi-class, real-world conditions, underscores its potential as a scalable and generalizable solution for precision agriculture.

Importantly, foliar disease diagnosis remains a crucial but underexplored area in the literature, especially concerning tropical crops such as mango. Leaf diseases are often early indicators of plant stress and can significantly affect fruit development and overall yield. Therefore, developing robust, accurate, and field-deployable diagnostic systems for leaf disease identification is critical for achieving sustainable agricultural outcomes. Our contribution lies not only in achieving state-of-the-art performance but also in offering a practical architecture that balances accuracy, computational efficiency, and adaptability, setting a new benchmark in artificial intelligence (AI)-assisted mango disease diagnosis.

In this study, the proposed MangoLeafCMDF-FAMNet architecture was rigorously evaluated using a 5-FCVP across three publicly available datasets: MLD1, MLD2, and MLD3. The training phase employed the AdamW optimizer with an initial learning rate set to 0.00005 and a weight decay of 0.0001. Cross-entropy loss was used as the objective function to guide the optimization process. Model performance was assessed comprehensively using multiple evaluation metrics, namely CA, RCL, PRC, MCC, and kappa. In addition, confusion matrices and high-dimensional feature distributions visualized through t-SNE were generated to provide further insights into the discriminative capability of the model. Table 4 summarizes the performance metrics obtained for each fold across all three datasets. For MLD1, the model achieved remarkably high performance, consistently exceeding 99.00% CA across all folds. Specifically, Fold 3 yielded a perfect CA, RCL, and PRC of 1.0000, with corresponding MCC and kappa of 1.0000, indicating flawless classification without any mispredictions. Even in the comparatively lower-performing Fold 4, MangoLeafCMDF-FAMNet still maintained an outstanding CA of 0.9938, demonstrating its robustness against potential variability in the data splits. Similarly, for MLD2, the model maintained exceptional performance. Perfect scores were achieved in Folds 3 and 5, mirroring the trends observed in MLD1. Notably, the lowest CA across all folds was 0.9970, which still reflects a near-perfect classification capability. The consistently high MCC and kappa across folds further underline the model’s strong agreement between the predicted and true class labels, confirming its reliability. On MLD3, which is inherently more challenging due to greater symptom variability and inter-class similarity, MangoLeafCMDF-FAMNet continued to demonstrate excellent performance. The CA values across the five folds ranged from 0.9918 to 0.9965, with the highest score achieved in Fold 5. RCL and PRC closely mirrored the trends of CA, and the high MCC and kappa reaffirmed the model’s ability to generalize well even under more complex conditions.

Following, a detailed class-wise performance analysis was conducted to further assess the robustness and generalization ability of MangoLeafCMDF-FAMNet. Specifically, RCL and PRC were calculated for each class across all folds on the MLD1, MLD2, and MLD3 datasets. For the MLD1 dataset, the model demonstrated outstanding classification capabilities. As shown in Figure 7 , the average RCL values were 0.9988 for Anthracnose, 0.9988 for Bacterial Canker, 1.0000 for Cutting Weevil, 0.9987 for Die Back, 0.9951 for Gall Midge, 0.9987 for Healthy leaves, 0.9961 for Powdery Mildew, and 0.9962 for Sooty Mould. In terms of PRC, the averages were equally high, reaching 1.0000 for several classes, with minor reductions to 0.9962 and 0.9935 for Powdery Mildew and Sooty Mould, respectively. These results confirmed that the model could accurately distinguish subtle disease symptoms even under slight class imbalance or symptom similarity.

Similarly, in the MLD2 dataset, the model achieved nearly perfect classification. The average RCL scores were recorded at 0.9991 for Anthracnose, 1.0000 for Die Back, 0.9990 for Gall Midge, 1.0000 for Healthy leaves, and 0.9958 for Powdery Mildew. Correspondingly, the PRC values were consistently excellent, with an average exceeding 0.9990 for all categories. Notably, the model maintained strong performance even in folds where small fluctuations in Powdery Mildew recognition were observed, demonstrating resilience against minor dataset variations.

The analysis of the MLD3 dataset, which is inherently more challenging due to the larger number of disease classes, also revealed robust model performance. The average RCL values across folds were 0.9869 for Anthracnose, 0.9949 for Bacterial Canker, 0.9910 for Cutting Weevil, 0.9992 for Die Back, 0.9909 for Gall Midge, and 1.0000 for Healthy leaves, Powdery Mildew, and Sooty Mould. PRC values aligned closely, maintaining averages above 0.9850 for all categories. While slight drops were noted in classes such as Anthracnose and Sooty Mould, the model overall preserved an exceptional balance between sensitivity and specificity. To visually capture these findings, RCL and PRC radar plots were generated across all folds, as depicted in Figure 7 . In these visualizations, Fold 1 is represented in dark blue, Fold 2 in orange, Fold 3 in gray, Fold 4 in yellow, Fold 5 in dark navy, and the average of all folds is plotted in green.

Furthermore, to comprehensively evaluate the class-level robustness of MangoLeafCMDF-FAMNet, class-wise MCC and Kappa were calculated and illustrated in Figure 8 for the MLD1 dataset across the 5-FCVP. Specifically, MCC scores remained extremely high for all disease categories, often reaching perfect agreement across most folds. Minor variations were observed only in the Gall Midge and Sooty Mould classes, where the MCC values slightly dropped but still remained above 0.98, demonstrating the strong generalization capacity of the model without signs of overfitting. Similarly, the Kappa mirrored the MCC trends, with near-perfect agreement across all folds and classes. On average, both MCC and Kappa exceeded 0.99 for nearly every class, highlighting the model’s consistent ability to correctly classify diverse disease symptoms under varying validation conditions.

In addition, we further computed the MCC and Kappa for the MLD2, as visualized in Figure 9 . As illustrated, the MCC scores for all classes consistently achieved near-perfect values, with almost every fold yielding scores of 1.0000 across the Anthracnose, Die Back, Healthy, and Powdery Mildew categories. A slight deviation was observed in the Gall Midge class during Fold 1 and Fold 2, where the MCC values dropped marginally but still remained exceedingly high, thereby underscoring the model’s remarkable stability even in the presence of subtle intra-class variations. Similarly, the Kappa mirrored these trends, maintaining values close to 1.0000 across all classes and folds, reaffirming the excellent agreement between predicted and true labels. The minimal variability observed in the Gall Midge class reflects realistic complexities inherent in agricultural imaging datasets, yet the exceedingly high average scores across all classes strongly indicate that MangoLeafCMDF-FAMNet successfully mitigates overfitting and maintains robust generalization.

Furthermore, to ensure a comprehensive performance analysis, we computed the class-wise MCC and Kappa of the MangoLeafCMDF-FAMNet on the MLD3, as presented in Figure 10 . The results demonstrated that the proposed model consistently achieved exceptionally high MCC values across all classes and folds. Specifically, MCC scores for classes such as Die Back, Healthy, Powdery Mildew, and Sooty Mould remained at or extremely close to 1.0000 across all folds. Although slight variations were noted in classes like Anthracnose and Bacterial Canker, the MCC scores still hovered around 0.98–0.99, reflecting highly reliable performance even in more challenging classes. Similarly, the Kappa closely followed the MCC trends, indicating outstanding agreement between predicted and ground truth labels across all folds and classes.

Furthermore, the feature distributions fused by MangoLeafCMDF-FAMNet were qualitatively analyzed using t-SNE, as illustrated in Figure 11 for each dataset individually. t-SNE serves as a powerful non-linear dimensionality reduction technique that projects high-dimensional feature representations into a two-dimensional space, facilitating the visualization of complex feature relationships. The t-SNE plots reveal that features extracted by MangoLeafCMDF-FAMNet exhibit clear, compact, and well-separated clusters for different disease classes across all three datasets. This visual evidence supports the numerical findings, indicating that the model effectively learns discriminative and disease-specific representations without significant overlap among categories. Moreover, the distinct cluster formations further demonstrate the absence of overfitting, affirming the model’s strong generalization capability to unseen samples.

The average confusion matrices for the MLD1, MLD2, and MLD3 datasets, computed across the 5-FCVP and illustrated in Figure 12 , provide critical insights into the class-specific discriminative capabilities of MangoLeafCMDF-FAMNet. These matrices (vertical axis: true labels; horizontal axis: predicted labels) reveal near-perfect diagonal dominance, underscoring the model’s ability to minimize misclassifications while maintaining high intra-class consistency. For MLD1, the matrix demonstrated exceptional precision, with Anthracnose and Bacterial Canker—classes often confused due to overlapping lesion patterns—achieving 159.80 correct predictions, respectively. Only minor off-diagonal errors were observed: 0.20 of Anthracnose samples were misclassified as Sooty Mould, while 0.20 of Bacterial Canker cases were incorrectly assigned to Healthy. The Cutting Weevil class exhibited flawless performance, with all 160 samples correctly identified. Similarly, Die Back and Gall Midge achieved near-perfect classification, with diagonal values of 159.80 and 159.40, respectively.

In MLD2, the matrix highlighted robust performance under increased symptom variability. Gall Midge, a class with subtle morphological features, achieved 199.80 correct predictions, with only 0.20 confusion with Healthy leaves. Die Back and Healthy were resolved with high precision, as evidenced by diagonal entries of 200.00. The most notable misclassification occurred in the Powdery Mildew class, where 0.60 samples were erroneously predicted as Gall Midge, likely due to shared textural patterns in late-stage infections.

For MLD3, the matrix further validated the model’s generalization capability. The Die Back class, often prone to misclassification due to its overlapping early-stage symptoms with Gall Midge, achieved a strong diagonal value of 255.80 correct predictions, reflecting the model’s ability to discern subtle differences in lesion distribution. Powdery Mildew, a class with visually ambiguous fungal patterns, was correctly classified in 154.80 instances, with only 0.20 samples misassigned to Die Back—a negligible error likely attributable to shared textural features in advanced infection stages. Gall Midge, despite its complex morphological variations across growth cycles, demonstrated exceptional performance with 442.60 accurate predictions. A minimal leakage of 3.80 samples to Sooty Mould was observed, potentially stemming from similarities in necrotic patterning under low-light imaging conditions. The Healthy class once again exhibited flawless discriminative capability, with all 250.00 samples correctly identified, underscoring the model’s precision in isolating disease-specific features from healthy tissue. Sooty Mould, a class frequently confused with Powdery Mildew in conventional methods, achieved a near-perfect diagonal score of 264.80, further highlighting the architecture’s proficiency in resolving spectral ambiguities. These results collectively affirm that MangoLeafCMDF-FAMNet generalizes robustly across diverse data distributions, making it particularly suitable for real-world agricultural applications where symptom variability and class overlap are prevalent.

In order to robustly validate the effectiveness of the proposed MangoLeafCMDF-FAMNet, an extensive comparative analysis was conducted against prominent baseline models, including MangoLeafCMDF-Net, ViT, and ConvNeXt, across the MLD1, MLD2, and MLD3 datasets. The results, summarized in Table 5 , clearly demonstrate the superiority of MangoLeafCMDF-FAMNet across all evaluation metrics. Specifically, on MLD1, MangoLeafCMDF-FAMNet achieved a CA of 0.9978, outperforming MangoLeafCMDF-Net with 0.9961, ConvNeXt with 0.9939, and ViT with a considerably lower score of 0.9256. In terms of RCL and PRC, MangoLeafCMDF-FAMNet consistently achieved 0.9978 for both, substantially higher than the competing models. On MLD2, the proposed model maintained its leading performance, reaching a CA of 0.9988, while MangoLeafCMDF-Net, ConvNeXt, and ViT achieved 0.9960, 0.9814, and 0.9176, respectively. Even on the more challenging MLD3 dataset, MangoLeafCMDF-FAMNet maintained its superiority, recording a CA of 0.9943, whereas ConvNeXt achieved 0.9864 and ViT remained at 0.9111. When considering robustness metrics, MangoLeafCMDF-FAMNet attained an MCC of 0.9975 and a Kappa of 0.9975 on MLD1, again consistently exceeding those of the other methods. This outstanding performance was mirrored across MLD2 and MLD3, highlighting not only accuracy but also high reliability and class-wise agreement.

To rigorously evaluate whether the performance differences between the proposed MangoLeafCMDF-FAMNet and baseline architectures were statistically significant, we conducted a Tukey Honest Significant Difference (HSD) post-hoc test based on the CA values obtained across five folds for each dataset. The results revealed that MangoLeafCMDF-FAMNet significantly outperformed the ViT model across all three datasets (MLD1, MLD2, and MLD3), with adjusted p-values consistently below 0.01. Likewise, when compared to the ConvNeXt backbone, the proposed method demonstrated statistically significant improvements in MLD1 (p = 0.024) and MLD3 (p = 0.038), while achieving a strong yet marginally nonsignificant advantage in MLD2 (p = 0.067). These findings underscore the consistent superiority of MangoLeafCMDF-FAMNet in terms of CA, particularly highlighting the added value of its attention-enhanced, cross-modal fusion strategy. Furthermore, the absence of overlapping confidence intervals supports the robustness of the proposed model’s improvements and suggests that the observed performance gains are unlikely due to random variation or overfitting.

This study introduced MangoLeafCMDF-FAMNet, a novel attention-augmented hybrid DL architecture tailored for robust multi-class mango leaf disease classification. By synergistically combining ConvNeXt and ViT backbones through a CMDF strategy, and further enhancing their output with FAMs, the proposed method effectively captured both fine-grained local patterns and global semantic context. Extensive experiments on three publicly available datasets (MLD1, MLD2, and MLD3) demonstrated the model’s superior classification performance across various evaluation metrics, including CA, RCL, PRC, MCC, and Kappa.

The proposed architecture consistently achieved exceptionally high accuracies across all datasets, reaching 0.9978 on MLD1, 0.9988 on MLD2, and 0.9943 on MLD3. These results notably outperformed competing baselines such as ViT, ConvNeXt, and MangoLeafCMDF-Net. Class-wise evaluations further confirmed the model’s capacity to distinguish complex and visually similar disease symptoms with high RCL and PRC. Visualizations via t-SNE and confusion matrices affirmed the learned feature separability and robustness against inter-class confusion, while the statistical analyses via Tukey HSD post-hoc testing verified that the observed improvements were statistically significant (p < 0.05) in most comparisons, particularly over the ViT and ConvNeXt baselines.

Furthermore, the architecture demonstrated a remarkable ability to generalize without overfitting, even on MLD3—a dataset characterized by greater class imbalance and symptom variability. The inclusion of FAMs was instrumental in adaptively amplifying disease-relevant features while suppressing irrelevant or redundant information, thereby enhancing class separability across diverse visual domains.

Despite these promising outcomes, the study is not without limitations. First, although the model was tested across three comprehensive datasets, all samples were derived from controlled imaging conditions. Future work should explore the model’s applicability to in-field images collected under varying lighting, occlusion, and background clutter. Second, while the proposed model achieved excellent results in disease identification, it currently does not support disease severity estimation, which is crucial for more nuanced decision-making in real-world scenarios.

As future research directions, we aim to extend the MangoLeafCMDF-FAMNet architecture for real-time mobile deployment in smart agriculture systems, incorporate multimodal inputs such as hyperspectral or thermal imagery to improve resilience under environmental variations, and explore the integration of explainability modules to foster model transparency for end-users such as farmers and agronomists. In conclusion, MangoLeafCMDF-FAMNet represents a scientifically grounded, practically scalable, and statistically validated advancement in automated plant disease recognition.

The experimental findings obtained in this study demonstrate that MangoLeafCMDF-FAMNet offers a highly effective solution for the complex task of multi-class mango leaf disease classification. By integrating ConvNeXt and ViT within a unified hybrid architecture and augmenting their capabilities with FAM, the proposed model achieves a refined balance between local feature extraction and global semantic understanding. This synergy enables precise discrimination of disease types, particularly in cases where subtle morphological differences challenge conventional classifiers.

The model’s high performance across three publicly available mango leaf datasets confirms its robustness and generalizability under controlled conditions. CA approaching 0.999, alongside strong MCC and kappa, indicate the architecture’s capacity to produce stable, reliable, and interpretable predictions. Furthermore, the CMDF mechanism contributes significantly to the enrichment of feature representations, enabling more resilient learning from heterogeneous visual patterns.

Despite these strengths, it is essential to contextualize the results within the scope of the datasets utilized. All datasets in this study were collected under relatively uniform environmental settings, characterized by consistent lighting, minimal occlusions, and simplified backgrounds. While such conditions are favorable for model training and benchmarking, they may not fully reflect the variability encountered in operational agricultural environments. In practice, field images often contain challenges such as partial leaf visibility, shadowing, cluttered scenes, and inconsistent illumination, which may affect the model’s generalization performance.

This observation highlights the importance of future work focused on validating the proposed framework using in-field image datasets collected in diverse and uncontrolled environments. Incorporating real-world variability into the training and evaluation pipeline will facilitate the development of more adaptive and field-deployable models. In addition, real-time image acquisition technologies, such as mobile devices and drone platforms, present promising avenues for extending the system toward scalable agricultural decision support.

Another important consideration relates to the clinical utility of disease severity estimation. While the present model effectively identifies the disease type, it does not explicitly address the severity or progression stage of the infection. In real-world agricultural applications, the intensity of disease symptoms is a critical factor influencing treatment strategies and resource allocation. Thus, expanding the model to support ordinal or regression-based predictions for disease severity would significantly enhance its applicability. Although the datasets used in this study did not provide severity annotations, future efforts will focus on curating such datasets and developing models capable of jointly performing disease identification and severity grading.

Moreover, it is worth considering that the absence of diverse environmental conditions and severity-level annotations in the training data may limit the interpretability and practical utility of the current system. Addressing these challenges through targeted dataset development, attention to domain adaptation, and auxiliary prediction tasks will be essential for realizing the full potential of DL-based disease diagnosis in agricultural practice.

In conclusion, MangoLeafCMDF-FAMNet offers a robust and scalable architecture for automated mango leaf disease classification. By leveraging multi-level attention mechanisms and cross-modal fusion, the model provides a strong foundation for high-accuracy plant disease recognition. Future directions should emphasize improving real-world generalization and enhancing the interpretability of the system through disease severity assessment, ultimately supporting the broader goals of precision agriculture and sustainable crop management.

To evaluate the practical applicability of the proposed MangoLeafCMDF-FAMNet in real-world settings, we report key computational characteristics, including model complexity and inference efficiency. The model contains approximately 46.9 million trainable parameters, representing a balanced architectural design that ensures high discriminative power while maintaining computational feasibility. All training and evaluation experiments were performed using the PyTorch DL framework on a standard workstation equipped with an Intel(R) Core(TM) i7–9700 CPU and 8 GB of RAM, without access to GPU acceleration. Under this configuration, the average inference time for a single 224×224-pixel image was observed to be approximately 180–200 milliseconds, depending on system load and batch scheduling. These results suggest that MangoLeafCMDF-FAMNet remains computationally viable even in resource-limited environments, which is particularly beneficial for field-deployable plant disease diagnosis systems where high-end hardware may not be available.