Minimizing energy consumption and carbon emissions while maintaining system performance is a critical challenge in cloud task scheduling. This paper presents a multi-objective scheduling framework based on a Memetic Algorithm (MA) designed to optimize task-to-VM mapping with respect to energy efficiency, carbon footprint, and throughput. The algorithm employs a weighted fitness function that integrates actual and idle energy usage, simulated time-varying carbon intensity, and task throughput. To enhance solution quality, MA combines global evolutionary operations (selection, crossover, mutation) with local search heuristics that adaptively refine candidate solutions based on workload characteristics and green energy opportunities. The carbon emission model incorporates dynamic emission factors (γ) derived from location- and time-sensitive datasets, reflecting real-world variability in grid carbon intensity. The proposed method is evaluated using the NASA Ames iPSC/860 workload under both low and high resource utilization scenarios. Comparative results demonstrate that the proposed MA approach achieves reduces the carbon emission by 20.2%, minimizes energy consumption by 17.7%, and enhances throughput by 21.2% over conventional techniques such as HDDPGTS and RAPTS, while also ensuring competitive performance in terms of makespan and resource utilization. These improvements underscore the potential of memetic-based hybrid scheduling to support environmentally sustainable and performance-efficient cloud infrastructures. The findings highlight the importance of integrating eco-aware intelligence into task scheduling policies, particularly for mission-critical and energy-intensive cloud applications.
Population growth and pandemics like COVID-19 have led to the depletion of natural resources and an increase in hospital waste generation. This issue is particularly pressing in developing countries, where innovative solutions are needed to address the environmental and health risks associated with improper waste disposal. Traditional waste sorting methods, which rely on human intervention, are time-consuming and pose a significant risk of infection. Moreover, different categories of hospital waste require specific treatment methods. This study proposes an artificial intelligence-based approach for classifying and sorting hospital waste using Convolutional Neural Networks (CNNs). The proposed CNN model effectively identifies and categorizes various types of hospital waste, providing a sustainable solution that enhances regulatory compliance. The model is developed using an approach which leverages K-fold cross-validation, data augmentation, and a publicly available, modest-sized dataset tailored for variability in waste categories. The model achieved a peak classification accuracy of 97.08%, along with strong precision, recall and F1-score, despite class imbalances and the presence of visually similar and hard-to-distinguish waste categories. These results highlight the potential of the model to improve hospital waste management practices, thereby reducing environmental impact and health risks.
This study presents the Multi-Modal Deep Fusion Network to identify cotton leaf diseases. Initially the images are collected from Kaggle cotton disease dataset. The dataset is preprocessed, and data augmentation is applied exclusively to the training set to prevent data leakage. The VGG-16-based Faster Region-based Convolutional Neural Network model is used for lesion detection and region of interest localization by generating bounding boxes around diseased areas. Both the handcrafted features, shape descriptors and color moments and deep learning features are used in feature extraction. The extracted features are optimized using the Snowy Wolf Optimization algorithm which combines Snow Leopard Optimization and Grey Wolf Optimization. The proposed achieved 98.4% accuracy, 98.6% sensitivity, and 98.8% F-score, consistently outperforming existing methods under identical experimental settings. While the proposed framework demonstrated promising performance on the evaluated dataset, further validation on larger and more diverse field datasets is required to comprehensively assess its generalization capability.
Machine learning models, particularly deep learning architectures, achieve high performance in prediction tasks but remain susceptible to adversarial attacks. This study aims to enhance the robustness of Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), and Recurrent Neural Networks (RNNs), thereby improving the security of machine learning systems. A three-step approach is adopted. First, benign sample classification is performed using the MNIST benchmark dataset. Second, adversarial attacks, namely Projected Gradient Descent (PGD), DeepFool (DF), and the Fast Gradient Sign Method (FGSM), are launched on the trained models, resulting in significant performance degradation. Based on the biased outputs induced by adversarial perturbations, an adversarial detection model is subsequently established. Third, to counteract these attacks, various defense strategies, including adversarial training, defensive distillation, autoencoder-based denoising, ensemble methods, and feature squeezing are employed and evaluated using standard performance metrics and graphical analyses. The results indicate that, in the absence of defense mechanisms, PGD attacks lead to accuracy drops of approximately 27% in CNNs, 83% in DNNs, and 90% in RNNs, demonstrating severe model vulnerabilities. However, when defense strategies are applied, all models recover to an accuracy of at least 98.9%, with adversarial training improving performance under attack by up to 90%. Among the evaluated models, CNNs exhibit the highest baseline robustness, whereas DNNs and RNNs rely more heavily on defense mechanisms to maintain performance. These findings provide valuable insights into the development of secure and resilient machine learning systems capable of mitigating adversarial threats.
Social media’s worldwide expansion over the past two decades has significantly altered the dissemination of extremist narratives, creating both challenges and opportunities for counterterrorism efforts. Addressing critical gaps in the detection and classification of extremist content on social media platforms, this research supports earlier-stage analytical assessment for law enforcement and security agencies. Using datasets from the publicly available Global Terrorism Database (GTD, n > 209,000 incidents) and a curated corpus of labeled tweets (n = 17,410), a hybrid framework integrating machine learning and deep learning models through a late-fusion stacking architecture is developed. The proposed ensemble leverages contextual indicators derived from historical terrorism data alongside linguistic and behavioral signals from social media content to distinguish extremist from non-extremist activity. Evaluated under strict temporal validation, the model achieves an accuracy of 98.52%, precision of 97.01%, recall of 99.66%, and an AUC of 0.92 under controlled experimental conditions. To address ethical and transparency considerations, Shapley Additive exPlanations (SHAP) are employed to enhance collectively indicate that integrating interpretability in automated decision-making. While the reported results reflect dataset-specific evaluation, the findings historical terrorism intelligence with temporally ordered social media analysis can support counterterrorism efforts by mitigating digital radicalization pathways and associated downstream physical security risks linked to terrorism and extremism through earlier analytical intervention.
In real-world clinical settings, the growing number of patients and the shortage of experienced ophthalmologists make early and accurate diagnosis of retinal diseases increasingly challenging. Cataracts, diabetic retinopathy, and glaucoma are some of the most common causes of lifelong blindness around the world. This is why there is a need for automated diagnostic systems that can accurately diagnose and interpret clinical data. The major goal of this work is to find out if a deep learning architecture based on EfficientNetB3 and Explainable Artificial Intelligence (XAI) can accurately classify multiple types of retinal diseases while still being clear to doctors. The proposed system categorizes retinal fundus images into four groups: cataract, diabetic retinopathy, glaucoma, and normal. The dataset consisted of a balanced and publicly accessible collection of 4,217 retinal fundus pictures, processed using standard preprocessing techniques to enhance their generalizability. We chose EfficientNetB3 as the main architecture since it is better at extracting features, and we compared it to the standard convolutional neural network baselines to show how useful it is. The suggested model was 97% accurate in classifying better than Residual Network 50 (ResNet50) is 91% and Visual Geometry Group 16 (VGG 16) is 87%. The high precision, recall, and F1-scores (0.94 – 1.00), the Cohen’s kappa of 0.95, and the low logarithmic loss of 0.10 all point to reliable predictions. The receiver operating characteristic analysis yielded an AUC of 1.00 across all illness categories. Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to address the interpretability deficit in deep learning-based medical systems and to pinpoint clinically significant retinal regions that influence model predictions. The results indicate that employing XAI alongside EfficientNetB3 enhances both diagnostic precision and interpretability, hence validating its suitability as a transparent decision-support system for the automated screening of retinal disorders.
Accurate detection of ECG arrhythmias plays a critical role in enabling timely diagnosis and treatment of cardiovascular diseases, which remain the leading cause of mortality worldwide. However, achieving high classification performance remains challenging due to class imbalance, signal variability, and resource constraints in real-time deployments. This study aims to enhance ECG arrhythmia detection accuracy through an optimized ensemble approach combining multiple CNN models with a novel stage-wise sieving strategy. Methodology: Three lightweight CNN models (ShuffleNet, MobileNet-v2, ResNet-18) were integrated into a multi-stage binary classification framework. Each stage systematically eliminated accurately classified arrhythmia classes. The novelty of the proposed approach lies in introducing a stage-wise sieving strategy that incrementally removes well-classified classes, combined with an optimized ensemble fusion of multiple CNN models guided by metaheuristic optimization techniques to boost performance. Optimization techniques, including Particle Swarm Optimization, Whale Optimization Algorithm, Grey Wolf Optimizer, Ant Colony Optimization, and Firefly Algorithm, were applied to improve model fusion. The approach was validated using combined public datasets (PTB-XL, MIT-BIH, and Shaoxing ECG databases). Results: The proposed stage-wise sieving ensemble significantly improved overall classification accuracy by 17.95%, reaching 96.29% accuracy using the Grey Wolf Optimizer. Classes previously misclassified, such as Conduction Disturbance and Hypertrophy, exhibited accuracy improvements of up to 32.44% and 25.19%, respectively. Conclusion: The proposed optimized ensemble approach significantly enhances ECG arrhythmia detection performance and demonstrates feasibility for real-time deployment on resource-constrained platforms such as Raspberry Pi.
It is well known that diagnosing Alzheimer's disease (AD) accurately and early is a major clinical challenge, especially when using brain MRI data to differentiate between subtle stages of cognitive decline. This study investigated the efficacy of two deep learning models for the classification of AD stages: Vision Transformer (ViT), a transformer-based architecture, and EfficientNetB7, a convolutional neural network. To enhance classification performance and address class imbalance, extensive data preprocessing and augmentation techniques were employed on the publicly accessible 'Alzheimer’s Dataset (4 class of Images)' from Kaggle. This dataset comprises 6,400 brain MRI images categorized into four AD stages: Non-Demented, Very Mild Demented, Mild Demented, and Moderate Demented. Techniques applied included cropping, horizontal and vertical flipping, 20-degree rotations, histogram equalization, Gaussian noise addition, Gaussian blurring, and thresholding, aimed at improving the representation of underrepresented classes. Hyperparameter optimization was executed via a two-phase methodology: an initial grid search to determine parameter ranges, succeeded by Bayesian optimization employing an upper confidence bound acquisition function to refine learning rates, batch sizes, momentum, and weight decay values. Experimental results indicated that EfficientNetB7 attained a classification accuracy of 93.5% with F1-scores surpassing 92% for early-stage classes, whereas Vision Transformer (ViT) recorded a lower accuracy of 88.7% and exhibited diminished sensitivity to early-stage instances. The performance disparity is due to ViT's dependence on extensive training datasets, which may restrict its generalization when utilized on comparatively smaller medical imaging datasets. The results indicate that, in dataset-constrained scenarios, CNN-based architectures such as EfficientNetB7 may provide more consistent and effective performance. Using distinct training, validation, and test datasets, the model's generalization, training stability, and computational efficiency were assessed. With an intuitive user interface, the top-performing model, EfficientNetB7, was implemented as a web-based application to facilitate real-time supportive predictions for research demonstration. This comparative analysis demonstrated that the CNN-based EfficientNetB7 exhibited more robustness with constrained medical imaging data and was computationally economical, but the transformer-based ViT displayed increased sensitivity to dataset size and necessitated extended training to attain similar convergence. The development of a validated and deployable AI-based Alzheimer's disease diagnostic solution showed great promise for clinical use.
In this study, a novel decision support model integrating spherical fuzzy sets enhanced with autoencoder-based dimensionality reduction, MEREC weighting, and CODAS ranking methods is proposed for high-dimensional, uncertain multi-criteria decision problems. The spherical fuzzy set structure allows decision makers to express their evaluations using three levels of membership (membership, non-membership, and hesitation). Thus, it produces linguistic evaluations appropriately to the nature of uncertainty. In the numerical analysis, five elective courses, Python Programming, Java Programming, C# Console Programming, Visual Programming with C#, and Web-Based Programming, were evaluated based on 41 selection criteria. The latent structures among these criteria were analyzed using the Autoencoder architecture, yielding 17 latent features with a reconstruction mean squared error of 0.016 as determined by an elbow-based reconstruction loss analysis, indicating negligible information loss beyond this dimension. The weights for these dimensions were objectively calculated using the MEREC method, which is based on the distinctiveness of each dimension in the decision process. The CODAS method was applied to rank the courses and provide decision support using the calculated weights. In the final stage, a comprehensive sensitivity analysis was performed to test the impact of changes in both dimension weights and decision-maker weights on the results Sensitivity analysis further confirmed the robustness of the proposed framework, with the top-ranked alternative preserved under ±10% criteria weight perturbations. The numerical results illustrate the practical applicability of the proposed framework and validate its effectiveness in handling complex evaluation structures. Although the proposed framework is demonstrated through a programming course selection problem, the methodology is generic and can be readily applied to other complex decision-making scenarios involving high-dimensional, uncertain, and interrelated criteria.
This study investigates interaction effects among rule sets, sensor fusion strategies, and membership functions on the navigational performance of a nonholonomic wheeled mobile robot in static, unknown environments using fuzzy logic controller. Employing a 3×3×3 factorial design, factors including rule set size (27, 18, and 14 rules), fusion level (minimal, moderate, and dense), and membership function shape (triangular, trapezoidal, and Gaussian) were varied. Each of the 27 configurations were evaluated in triplicate using a MATLAB/CoppeliaSim co‐simulation, with traversal time as the performance metric. An analysis of variance (ANOVA) revealed that each of the three main factors had a significant impact on traversal time (p < 0.001). Notably, there were also meaningful interactions between rule set size and membership function, as well as between rule set size and sensor fusion (p < 0.01), suggesting that system performance is closely tied to how these parameters are combined. Among the tested configurations, setup with a 14-rule base, Level 2 sensor fusion, and a triangular membership function consistently achieved the fastest average traversal times. These interactions likely arise from computational perceptual trade-offs. Increasing rule set size enhances decision granularity but introduces inference delay, whose effects vary depending on how smoothly membership functions partition the input space and how densely sensor data are fused. In practice, this implies that controller performance depends on achieving a balance between linguistic complexity, sensor integration depth, and fuzzification. The findings therefore emphasize the importance of joint parameter tuning and offer design insight for balancing computational cost against navigational precision in embedded fuzzy logic controllers.
Detecting and classifying brain tumours is essential for early diagnosis and effective treatment planning, significantly enhancing patient outcomes. This research presents a deep learning-based approach that utilizes T1-weighted MRI data to automatically identify and classify brain tumours, distinguishing between normal and abnormal cases. The proposed methodology consists of four key steps: pre-processing, segmentation, feature extraction, and classification. In the pre-processing stage, image quality is enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE) to boost contrast, along with a Gaussian filter to reduce noise. Tumour segmentation is achieved through thresholding, which effectively isolates the tumour regions. For feature extraction, a Convolutional Neural Network (CNN) captures high-dimensional features that are vital for classification. To accurately differentiate between normal and abnormal tumours, an Artificial Neural Network (ANN) is employed for classification. The effectiveness of the proposed technique is evaluated based on performance metrics such as time, accuracy, and peak signal-to-noise ratio (PSNR). The obtained parameters are compared with existing techniques to highlight improvements in detection and classification performance. Among the tested images, the best result achieved a PSNR of 13.015 dB, an accuracy of 99.231%, and a computational time of 1.267 ms, demonstrating the efficiency and reliability of the proposed method for brain tumor detection and classification. Overall, this approach provides an effective and automated method for detecting brain tumours, aiding in clinical decision-making and diagnosis.
The rapid advancements in deep learning and classifier fusion techniques offer promising solutions to enhance the accuracy and robustness of biometric authentication systems in this paper we propose the integration of these methodologies, specifically in multimodal biometric systems that utilize face and fingerprint recognition. The research investigates various deep learning architectures, highlighting their effectiveness in processing diverse biometric datasets. Additionally, it examines classifier fusion techniques, which combine multiple classifiers to improve person identification performance. A significant focus of this research is on spoofing and anti-spoofing measures. Biometric systems, especially those involving facial and fingerprint recognition, are vulnerable to spoofing attacks such as the use of photographs, videos, or artificial fingerprints to impersonate legitimate users. We developed various anti-spoofing strategies that are integrated into the biometric authentication process to mitigate these risks. These include techniques like texture analysis, motion analysis, and liveness detection, which help differentiate between genuine biometric traits and spoofed samples. We benchmarked a comparative analysis of deep learning models and classifier fusion, demonstrated their strengths, weaknesses, and best practices. Additionally, performance evaluations focus on key metrics such as accuracy, computational efficiency, scalability, and the system’s ability to resist spoofing attacks. Ultimately, the paper emphasizes the potential of these advanced techniques to revolutionize biometric systems, with a particular focus on future research directions for optimizing these methodologies, particularly in the context of improving robustness against spoofing and enhancing the overall security of biometric authentication systems. overall system Equal Error Rate (EER), the True Acceptance Rate at a specified False Acceptance Rate (e.g., TAR @ 0.1% FAR), and the accuracy of the anti-spoofing module.
Public transport (PT) users often experience instances of leaving items behind in the public transport system. Finders who come across these items may choose to keep them maliciously or, out of goodwill, decide to return them. This paper aims to utilize six (6) machine learning models, including LR, SVM, DT, RF, NB, and KNN, to predict the ability of finders to return found items. Nine (9) features, comprising four (4) demographic parameters (age, gender, income, and education), were used in the models’ prediction process. The study involved a total of 603 PT users in the Accra cosmopolitan area of Ghana to assess finder’s decision regarding returning found item(s). The classification success rates were obtained as follows: 86.740% (LR), 87.293% (SVM), 82.873% (DT), 85.083% (RF), 85.083% (GNB), and 87.845% (KNN) using Python codes. The RF model also performed well, considering the balance of performance with the desired precision and recall. RF, GNB, and LR achieved the highest AUC values (0.78), demonstrating strong discriminative ability in predicting user honesty.
Electronic devices and internet purchasing are more common today. For online line shopping most of people are using internet banking and credit for doing payment for purchasing. For time saving and various offers on credit card and debit card customer prefer on line shopping like various platform Amazon, Flip cart, big basket etc. For online transaction security is prime concern. There is various type of attack possible during online transaction, stealing of password, fraud transaction, and meet in middle attack etc. During the online transaction stealing confidential information like OTP, transfer money from someone account to another account is a crime. In the digital world fraud during the online transaction day by day increases exponentially. To detect the unauthenticated transaction and fraud during online used various methods. Data is playing very important role during the online fraud. So, knowledge discovery is most frequently used to protect online fraud. In this paper suggested a technique based on knowledge discovery and machine learning methods, we strive to develop the best model possible in this research study to predict transactions involving fraud and transactions involving no fraud. Fraud detection uses a variety of machine learning techniques, including K-Means clustering methods, Support Vector Classifier, Logistic Regression, and Anomaly Detection Algorithm Techniques. After analysis it was found that Anomaly Detection Algorithm Techniques gives best accuracy for fraud detection 99.85%.
This paper proposes a drift-aware off-policy deterministic actor–critic framework for constrained continuous resource allocation in non-stationary environments. Feasible allocations are ensured by a simplex-parameterized policy using softmax normalization with budget scaling, avoiding projection or Lagrangian tuning. The reward integrates Nash social welfare via mean log-utility, efficiency, fairness, and constraint-violation penalties with adaptive weights. To improve sample efficiency, we adopt prioritized experience replay based on TD error and state novelty. Non-stationarity is detected by KL divergence between recent and historical state-visitation distributions; detected drift triggers buffer refresh and incremental fine-tuning, while Elastic Weight Consolidation mitigates catastrophic forgetting. Experiments across six application-motivated domains (food, medical, housing, education services, employment support, and elderly care) demonstrate improved utilization and welfare with reduced inequality and low decision latency compared with optimization, heuristic, and DRL baselines. Results are reported over multiple runs with mean ± standard deviation and corrected significance tests.
Forecasting time series data especially in volatile sectors like financial markets, shows significant challenges due to non-linearity, non-stationarity and noise in the data. Traditional forecasting models most likely fail to generalize effectively across varying tasks without extensive retraining. This study investigates the application of meta learning techniques, particularly First-Order Model-Agnostic Meta-Learning (FOMAML) and Reptile, to make adaptability and generalization better in time series forecasting tasks. An extensive empirical study was done using three neural networks as base models, namely Long Short Term Memory (LSTM), Gated Recurrent Unit (GRU) and Feed Forward Neural Network (FFNN) applied to four real-world stocks: TCS, TATASTEEL, GRASIM and DJIAHD. The models were evaluated under few-shot learning(defined here as 211-shot learning using sliding window samples) conditions with varying iteration counts(outer loops or epochs) and their effectiveness was checked using some common standard metrics like RMSE(Root Mean Squared Error), MAE(Mean Absolute Error) and R²(Coefficient of Determination). Outcomes have shown that meta-learning approach notably performs much better than traditional models with MAML(First Order) in particular showing quicker task adaptation as well as stable convergence behavior, especially when it used with GRU and LSTM as base models, as validated empirically on the GRASIM dataset where the MAML with LSTM configuration attained around 81.9\% reduction in RMSE (dropping the value from 622.94 to 112.60 over the iterations). In all four stocks, reptile shows relatively steady performance. The study validates the potential of meta-learning as a powerful framework for time series forecasting problem in dynamic settings which offers robust algorithmic foundation for numerous future financial modeling applications.
Accurate prediction of cardiovascular disease (CVD) is essential for timely intervention and improved patient outcomes. This paper presents a hybrid model, BPSO-RAF-CNN that integrates Binary Particle Swarm Optimization (BPSO) with a Regularized Accuracy-Based Fitness Function (RAF) and a Convolutional Neural Network (CNN) to improve prediction performance through optimized feature selection. The approach begins with feature engineering on cardiovascular data, followed by BPSO-RAF to identify the most important, predictively salient and compact feature subset, lowering dimensionality and improving generalization. These selected features are then fed into a CNN for final classification. Extensive experiments demonstrate that BPSO-RAF-CNN outperforms traditional classifiers (Logistic Regression, SVM, Naive Bayes, Decision Tree, Random Forest) achieving an accuracy of 87.05%, Precision 89.71%, Recall 83.77%, F1-score of 86.05%. And Specificity 90.22%, all with a standard deviation 0.5%. The model also shows good performance across 10-fold cross-validation, indicating strong generalization.
Accurate and immediate incident identification is essential in the cybersecurity area, as it allows the timely detection of threats, along with countermeasures and mitigation, ensuring security for organizations and individuals. This reduces false positives and enables efforts to be concentrated on real risks. This paper presents a framework that integrates ontologies and Large Language Models (LLMs) to identify incidents from events within the context of security threats. Ontology rules are employed to infer probable incidents, resulting in an initial set of incidents for analysis. Furthermore, ontologies provide contextual information, which is combined with event data to formulate queries for LLMs. These interactions with LLMs produce a second set of probable incidents. The outputs from ontology-based inferences and LLM-driven responses are then compared, and the discrepancies are leveraged to refine ontology rules and adjust LLM responses. Experimental results, focusing on context generation and incident detection, demonstrate that the integration of ontologies and LLMs significantly enhances the accuracy of incident identification when compared to using only LLMs.
Community engagement is essential to social service delivery, yet traditional community needs assessment remains time-consuming and poorly suited for timely monitoring. This study proposes a semi-supervised learning framework to identify emerging community needs and service gaps from massive, mostly unlabeled, unstructured text. We construct an explicit heterogeneous text graph where each record is a document node linked to keyword and need-category nodes; document–document edges are built using a weighted combination of semantic similarity (BERT cosine), lexical overlap (keyword Jaccard), and temporal proximity. A graph neural network with iterative self-training leverages 3% expert-labeled seed data and the remaining unlabeled corpus to classify records into a 10-category need taxonomy. On 176,602 records, the proposed model achieves F1 = 0.895 and Recall = 0.899, outperforming supervised baselines trained on the same labeled ratio by 23.8% (macro-F1). Post-hoc quarterly aggregation of predictions enables trend monitoring and prioritization of service-gap severity for decision support.
Plant diseases adversely affect the quantity and quality of food production, contributing to food insecurity. Prompt identification, diagnosis, and intervention can significantly minimize economic and ecological losses. By reducing the use of agrochemicals through timely disease detection, the environmental impact can be mitigated. Traditional manual methods for recognizing plant diseases are prevalent but are often limited, time-consuming, costly, and ineffective. Convolutional Neural Network (CNN) architectures have demonstrated excellent capabilities in detecting plant diseases and other complex images, but they lack spatial or rotational invariance and require extensive data in various forms to be effective. This is typically achieved by applying data augmentation, as the datasets in the field of agriculture are often limited. Capsule Networks address CNN's limitations, but their encoder network is inefficient at feature extraction, hence does not perform well on complex images. This study seeks to modify and improve CapsNet by combining a Color Difference Histogram (CDH) with a Capsule Network that includes extra two convolutional, three max pooling layers, three batch normalization layers, and reduced the primary capsule channels in the original CapsNet to 16 from 32 for efficient plant disease detection in apples, bananas, grapes, corn, mangoes, pepper, potatoes, rice, tomato, and on the CIFAR-10 dataset. This approach improved the original CapsNet in terms of validation accuracies by 5.83%, 14.82%, 5.9%, 4.42%, 20.87%, 40.12%, 4.41%, 0.76%, 9.49%, and 13.97% on apple, banana, grape, corn, mango, pepper, potato, rice, tomato, and CIFAR-10 datasets respectively. The CDH-CapsNet achieved better results in terms of accuracy, sensitivity, F1-Score, precision, specificity, Receiver Operating Characteristic (ROC), Precision-Recall (PR) values, parameter count, and disk size, surpassing the original CapsNet and CapsNet models presented in available research. The original CapsNet and CDH-CapsNet exhibited strong performance on datasets such as the Rice dataset, possibly because of high-quality images and low intra-class variance. The findings suggest that this efficient and computationally less demanding supportive tool can significantly enhance plant disease classification by offering a lightweight, scalable solution that can be adapted for field use in resource-constrained settings, contributing to efforts aligned with the SDG 2 goal. However, environmental factors such as inconsistent lighting and complex backgrounds encountered in practical scenarios may affect the model's effectiveness. Subsequent studies will aim to overcome these issues and broaden the model's applicability.
