Accurate and early prediction of Cardiovascular Disease (CVD) plays an important role in improving patient survival and enabling timely preventive or therapeutic interventions. In this study, we propose an enhanced Crow Search Algorithm (eCSA) that integrates feature selection and hyperparameter optimization into a unified framework to improve the performance of a Bidirectional Long Short-Term Memory (Bi-LSTM) based heart disease prediction network. By employing a hybrid binary-continuous search strategy, the proposed eCSA jointly identifies discriminative features and optimizes Bi-LSTM hyperparameters, resulting in improved predictive performance and enhanced feature-level interpretability. Experimental evaluations were conducted on two benchmark datasets, namely the Cleveland Heart Disease dataset and the Framingham Heart Study dataset, under multiple train-test split conditions (80:20 and 70:30). Comparative analysis against state-of-the-art approaches, including Stochastic Configuration Network–Deep Bidirectional Long Short-Term Memory (SCN-Deep Bi-LSTM), Fitness-based Horse Optimization–Bidirectional Long Short-Term Memory (FHO-Bi-LSTM), Quantum Hybrid Deep Network (QHDN), and Convolutional Transformer Network (CTN-Trans), demonstrated that the proposed eCSA-BiLSTM model consistently outperformed competing methods across multiple evaluation metrics. Performance was assessed using accuracy, macro-averaged F1-Score (Macro-F1), precision, recall, Matthews Correlation Coefficient (MCC), and Area Under the Receiver Operating Characteristic Curve (AUC-ROC). The proposed model achieved an accuracy of up to 0.95 on the Cleveland dataset and 0.91 on the Framingham dataset, while maintaining stable and consistent performance across stratified 5-fold cross-validation and repeated train–test split experiments.
In this paper, we present a comprehensive approach for depth map completion using a variable pipeline model. The pipeline is designed to be flexible, and its structure is dynamically determined during the training phase by minimizing a fitness function. To achieve this objective, we used the Particle Swarm Optimization (PSO) technique, which efficiently ex-plores the solution space to optimize key parameters. The pipeline itself consists of three main stages: a convolutional stage (Lab Convolution), an Interpolation Model (IM), and a post-Convolutional Stage (SC2). During training, our model automatically adjusts several aspects, including the selection of filter parameters, interpolator settings, and the number of filters used at each stage. Moreover, the pipeline is capable of interchanging the execution order between the stages, either processing in the sequence of Lab Convolution-IM-SC2 or SC2-IM-Lab Convolution. This flexible structure allows for the discovery of better configurations that significantly im-prove the performance of depth map completion tasks. Our approach demonstrates superior results by adapting to the specific characteristics of the input data.
This paper explores and evaluates advanced signal processing techniques for blind channel identification and equalization in modern 5G networks. The focus is on traditional statistical methods and emerging artificial intelligence based on the autoencoder method. The study concentrates on two principal frameworks: high-order cumulants, specifically the fourth-order and sixth-order cumulants, which are capable of exploiting non-Gaussian and higher-order statistical dependencies in digitally modulated signals, and Artificial Intelligence (AI)-based models, particularly autoencoders, which use deep learning architectures to reconstruct or denoise received signals without prior knowledge of the transmission channel. To systematically assess the capabilities of each method, three distinct algorithms were developed and implemented: the fourth-order cumulant, the sixth-order cumulant, and the autoencoder algorithm. These algorithms were applied to 5G signals transmitted over Rayleigh multipath fading channels, modeled using standard propagation profiles such as BRAN-A, BRAN-B, and Proakis B. Simulations were conducted over a wide range of Signal-to-Noise Ratios (SNRs) and varying levels of noise and channel dynamics. Key findings from the simulation results are as follows: The sixth-order cumulant demonstrates superior noise resilience compared to the fourth-order cumulant, making it more suitable for environments with stronger interference or non-Gaussian noise. The autoencoder algorithm achieves the highest estimation and identification accuracy after sufficient training, due to its ability to model complex, nonlinear transformations. However, this comes with increased data requirements and computational cost. The fourth-order cumulant algorithm is computationally efficient, theoretically tractable, and performs well under stationary and low-noise conditions. However, it exhibits a notable performance decline in time-varying or high-noise scenarios. In contrast, the autoencoder algorithm shows enhanced robustness and adaptability to nonlinear, time-varying, and imperfectly modeled environments. Its strength lies in learning the underlying channel effects implicitly from raw data, without requiring a structured analytical model. Nevertheless, this benefit is offset by higher computational complexity and a lack of interpretability compared to the cumulant-based methods. In conclusion, while cumulant-based techniques remain valuable in structured or lightly impaired channels, AI based on autoencoder models, especially autoencoders, offers a more powerful and flexible solution for blind equalization under realistic and challenging 5G conditions.
Grading tobacco leaves is crucial for ensuring fair pricing and quality control, however, the process is still largely carried out manually, resulting in a slow, subjective, and often inconsistent outcome. In this work, we present a multi-task deep learning approach designed to automate the grading of air-cured Burley tobacco leaves in controlled settings. The model is constructed with shared convolutional layers and separate task-specific branches, allowing it to predict stalk group, quality, and color at the same time, in line with the hierarchical grading system. To improve consistency, images were preprocessed using coin-based size normalization, rotation alignment, and segmentation. In our experiments, the multi-task model with EfficientNetB0 achieved an accuracy of 94.82% and significantly outperformed the multi-class and single-task baselines, while reducing both training time and inference delay. These findings suggest that multi-task learning can be a valuable and robust method for automated tobacco grading, showing gains in accuracy, speed, and scalability compared to other algorithms.
Accurate ambient temperature prediction is essential for climate monitoring, urban planning, and environmental management, particularly in regions experiencing rapid climatic variability such as Sri Lanka. This study investigates the application of explainable machine learning models for short-term ambient temperature prediction in Battaramulla, Sri Lanka. Five regression algorithms-K-Nearest Neighbors (KNN), Decision Tree (DT), Random Forest (RF), Support Vector Regression (SVR), and Histogram-based Gradient Boosting Regressor (HGBR) were evaluated using 14 meteorological and environmental predictors, including temporal variables, relative humidity, solar radiation, rainfall, wind speed, and air pollutant concentrations (CO2, NOₓ, CH4, O3, CO, PM2.5, and PM10). Among the models tested, HGBR demonstrated superior predictive performance, achieving R² values of 1.00 (training) and 0.96 (testing), with corresponding mean squared error values of 0.01 and 0.11. Model interpretability was examined using SHapley Additive exPlanations (SHAP), Partial Dependence Plots (PDP), Individual Conditional Expectation (ICE) analyses, and Accumulated Local Effects (ALE), which identified several features as the most influential predictors. Model validation using 192 real-time observations showed close agreement between predicted and measured temperatures, although the evaluation was limited to a single location and time period. A web-based application, ‘Therma’, was developed to facilitate practical deployment of the model for localized temperature estimation. Overall, this study demonstrates the utility of explainable machine learning for localized climate prediction while highlighting the need for broader spatiotemporal validation in future work.
Privacy issues and communication overhead are bottlenecks to the machine learning implementation in IoT sensor network. In this paper, we introduce a lightweight Edge-enabled Federated Learning (EFL) framework utilizing Personalized Federated Learning (FedPer) over edge computing for privacy-preserving collaborative training. Our solution mitigates the challenges of non-independent and identically distributed (non-IID) data, using client-specific personalization and secure aggregation without raw data exchange. Extensive experiments over five real-world IoT datasets (UCI HAR, Ambient, DOO-RE, SHL and WISDM) show that FedPer can achieve up to 96% accuracy—surpassing baseline methods (FedAvg, FedProx) by 2–12% within non-IID scenarios—meanwhile decreasing communication overhead by up to 20%. The hardware evaluation on Raspberry Pi 4 and Jetson Nano validates the realization of real-time inference (<30 ms/sample) for Decision Tree (DT) with compact model sizes (<500 KB). Our system offers a large-scale, privacy-assured way of intelligent sensing in smart home, wearables and industrial IoT.
Deep Learning (DL)-based systems, employing advanced growths, pave the way for bioinspired methods in almost all domains of life. Healthcare organizations can employ DL approaches due to their precision in recognizing and identifying distinct diseases. The coronavirus disease (COVID-19) epidemic has emerged as the most dangerous disease in recent times, posing a significant burden on health organizations worldwide. Medical imaging and PCR testing have the potential to analyze COVID-19. Given the high spreadability of COVID-19, Chest X-Ray (CXR) analysis is considered safe under various conditions. DL systems are capable of enhancing medical imaging tools and supporting radiologists in making medical decisions for the analysis, diagnosis, and monitoring of distinct diseases. With this motivation, this study presents a novel bioinspired red kite optimizer with a DL-based COVID-19 classification (BRKODL-COVIDC) method on CXR images. The BRKODL-COVIDC method intends to recognize and categorize the presence of COVID-19 using DL models. In the presented BRKODL-COVIDC method, the Bilateral Filtering (BF) model can be used for image pre-processing tasks. In addition, the complex patterns and features in the images can be derived from the DenseNet121 model. For optimal hyperparameter selection of the DL techniques, the Random Key Optimizer (RKO) model can be employed in this research. Last, the BRKODL-COVIDC technique makes use of the Dilated Convolutional Auto-Encoder (DCAE) model and is used for identification purposes. Thus, with this research study, we are contributing to the National Priorities for Research, Development, and Innovation (RDI) in health and wellness to provide and maintain a sustainable environment in the health sector. The simulation outputs of the BRKODL-COVIDC technique can be examined on a standard dataset. The simulated outputs demonstrated the enhanced performance of the BRKODL-COVIDC technique in the COVID-19 detection procedure.
The article presents an algorithmic model for decision-making on selecting the optimal strategy for modernizing a seagoing vessel via nonlinear multi-parameter optimization. A set of interrelated technical, economic, and operational factors that influence the feasibility of modernization measures within the vessel’s life cycle is considered. It is shown that traditional approaches to maintenance and periodic assessment of a vessel’s condition are insufficient for strategic planning, as they do not account for the long-term economic consequences, the level of risk, and the investment attractiveness of modernization. A model is proposed that integrates investment cost indicators, modernization parameters, characteristics of the modernization contractor, the projected service life of the vessel after modernization, expected annual profit, risk level, and life cycle indicators (Life Cycle Cost (LCC) and Life Cycle Assessment (LCA)). Net Present Value (NPV) is used as the main optimization criterion, and a system of constraints is formed that takes into account financial, time, and safety factors. The proposed algorithmic system allows for a reasonable choice of modernization type, parameter configuration, and contractor, ensuring maximum economic efficiency under existing conditions. The results indicate the promise of applying optimization algorithms in marine engineering to improve the reliability, profitability, and long-term sustainability of ship operations.
Packing presents a major challenge in cybersecurity, as it complicates malware analysis and extends the operational lifespan of malicious software. This study addresses the issue by developing a robust framework designed to detect packed executable files and identify the specific packers used. The proposed framework leverages 20 optimally selected features extracted from Portable Executable (PE) files to detect packing and recognize packer signatures. A series of extensive experiments was conducted to determine the most effective combination of classification model and feature set. The extreme gradient boost algorithm was selected based on its superior performance. The proposed model achieved a high detection accuracy of 99.27% and an F1-score of 98.84%, outperforming recent methods in the field. In addition, the study introduces a publicly accessible dataset containing 213,784 PE samples and 125 features to facilitate future research. The framework provides a practical tool for security analysts, improving their ability to identify and respond to PE file-based malware in real-world environments. This study focuses exclusively on a static analysis pipeline; no dynamic execution is performed. We also describe how the framework could interface with sandbox-derived dynamic behavioral signals in future work without extending the current study’s scope. Overall, this research contributes a static feature-based approach for packer detection and signature identification, together with a large-scale open dataset that supports ongoing advances in malware classification and analysis.
In intralogistics, accurate 3D object detection is essential for enhancing robotic perception and enabling autonomous navigation. This paper presents a novel hybrid method that combines 2D object detection with depth-based point cloud segmentation for efficient and real-time 3D object pose estimation. The key contribution of this paper is the Cumulative Histogram Depth Filter (HDF), a lightweight algorithm that segments dominant depth regions corresponding to detected objects. This approach uniquely enables the reuse of existing 2D-labeled datasets, eliminating the need for extensive 3D annotations. The proposed method was evaluated using both simulated and real-world data, obtaining high detection accuracy. In the simulated environment, it achieved a mean distance error of 0.14 m, an Intersection over Union (IoU) of 0.90, and a mean Average Precision at 50 (mAP@50) of 0.95. Real-world experiments using Azure Kinect and ZED2 cameras yielded an average distance error of 0.13 m, IoU of 0.78, and mAP@50 of 0.65. Additionally, the system runs at 105 FPS, significantly outperforming more complex hybrid architectures in terms of computational efficiency, making it particularly suitable for real-time robotic applications.
This paper presents Zero-Knowledge Federated Learning Guard (ZK-FLGuard), a privacy-preserving and verifiable federated learning framework for real-time anomaly detection in Fifth-Generation Mobile Network (5G)-enabled Internet of Things (IoT) environments. Building on the integration of zero-knowledge proofs (zk-SNARK—Zero-Knowledge Succinct Non-interactive Argument of Knowledge) and blockchain-based access control, ZK-FLGuard ensures the integrity of model updates without exposing private data. Using real-world intrusion detection datasets (CICIDS2017—Canadian Institute for Cybersecurity Intrusion Detection System 2017, TON_IoT—Telecommunications Organisation of the National Security—IoT) and a synthetic adversarial dataset, our evaluation shows that ZK-FLGuard achieves up to 0.96 F1-score (harmonic mean of precision and recall), improves recall in low-frequency attack detection, and introduces less than 10% additional latency overhead compared to standard Federated Learning (FL). Compared with centralized Long Short-Term Memory (LSTM) and FL without Zero-Knowledge Proof (ZKP), ZK-FLGuard provides competitive accuracy while ensuring verifiable computation and strong privacy guarantees. We address the critical challenge of securing federated anomaly detection in 5G-enabled IoT systems against data leakage, model poisoning, and unauthorized access. While FL preserves privacy by keeping raw data local, it remains vulnerable to gradient leakage and adversarial manipulation. Our hypothesis is that combining zero-knowledge proofs and blockchain with FL can deliver a scalable, tamper-resistant, and privacy-preserving detection pipeline suitable for resource-constrained edge environments.
The design of optimal Convolutional Neural Network (CNN) architectures has become increasingly complex as networks have grown deeper and more diverse. Neural Architecture Search (NAS) has emerged as a powerful technique to automate the discovery of high-performance architectures, thereby reducing reliance on manual expertise. In this work, we propose a novel approach that integrates NAS with a Multi-Layered Population Structure (MLPS), an effective evolutionary scheme originally developed for Genetic Programming (GP). We term this method Multi-Layered Population Structure Neural Architecture Search (MLPS-NAS). Our approach leverages the hierarchical, pyramid-like population management of MLPS to maintain a diverse set of partial solutions, while systematically exploring the vast search space of CNN architectures. A key feature of MLPS-NAS is its ability to construct complex architectures by repeatedly incorporating effective building blocks. To significantly accelerate the search process, we introduce a weight inheritance mechanism between generations, which drastically reduces the computational cost of training new candidate architectures. Experimental results on a real-world image classification task demonstrate that MLPS-NAS achieves accuracy comparable to an established method, CNN Architecture Design Using Cartesian Genetic Programming (CGP-CNN), while substantially reducing the required search time. This synergy between NAS and MLPS offers a promising direction for the efficient, automated design of high-performance CNNs.
Mitral Valve (MV) pathologies such as Mitral Valve Prolapses (MVP), Mitral Stenosis (MS), and type III regurgitation should be diagnosed as early as possible for the better management of the patients. This research work presents a method for the classification of MV diseases using echocardiographic images and texture analysis of the images using Gray Level Co-occurrence Matrix (GLCM) features in conjunction with Machine Learning (ML) classifiers. Initially, a Convolutional Neural Network (CNN) was employed to categorize echocardiographic images into two standard views: Apical Four-Chamber (A4C) and Parasternal Long-Axis (PLA). Next, the energy, contrast, correlation, and the entropy of GLCM-based texture features were obtained. The features were then fed into ML classifiers such as Random Forest (RF), Neural Networks (NN), Ensemble models to classify MV conditions. In the A4C view, the Neural Network Classifier (NNC) obtained an accuracy of 85% while in the Parasternal Long Axis (PLA) view, the accuracy was 84%. Some of the features of GLCM that were deemed important in the performance of the model were revealed. The results show that combining GLCM texture analysis with ML provides a potential way of improving the precision and reliability of MV disease diagnosis. The findings of this study contribute to advancements in cardiovascular disease detection by integrating machine learning techniques with echocardiographic analysis, ultimately supporting efforts to enhance public health and early disease diagnosis, in alignment with global healthcare initiatives.
The effectiveness of system analysis and design methodologies plays a pivotal role in the success of software development projects. This research conducts a comparative analysis of various methodologies, including the Waterfall Model, Spiral Model, Prototyping Model, Iterative Model, Unified Process, Object-Oriented Analysis and Design, Joint Application Design, Computer Aided Software Engineering Tools, Rapid Application Development / Rapid Systems Development, Feature- Driven Development, Extreme Programming, Agile Method, and DevOps, to explore their respective strengths and considerations. The study examines key aspects such as flexibility, documentation, risk management, complexity, and suitability across these methodologies. Findings reveal that each methodology offers unique advantages and challenges, influencing their applicability in different project contexts. The research underscores the importance of aligning methodology choices with project-specific requirements, organizational culture, and environmental factors to optimize software development outcomes. Moreover, it suggests that hybrid approaches, combining elements from multiple methodologies, may offer a balanced approach to address diverse project needs and maximize effectiveness. By providing insights into the comparative effectiveness of system analysis and design methodologies, this research contributes to informed decision- making and improved project success in software development endeavors.
This study presents an innovative programming assistance tool designed to address language barriers faced by Sinhala-speaking novice Java programmers. The tool provides real-time Java code generation and diagram creation based on Sinhala programming queries, enhancing conceptual understanding. Developed using a Design-Based Research methodology, the tool underwent iterative testing with 122 Sinhala-speaking learners, incorporating user feedback to refine usability and performance. Central to the system is Generative Pre-trained Transformer, version 3.5 Turbo, ensuring accurate translations and programming assistance, alongside a transformer-based model that translates Sinhala queries into English for processing. The translation model achieved 91.37% accuracy, with strong Bilingual Evaluation Understudy scores validating its contextual relevance. The tool’s practical applications extend beyond academia, supporting educational institutions, self-learners, and industry professionals in learning and skill development. Statistical evaluation of user performance demonstrated significant improvements in programming comprehension, reinforcing its effectiveness. By promoting inclusivity and expanding access to programming knowledge, this research contributes to the advancement of Sri Lanka’s technology sector and establishes a scalable framework for broader implementation in multilingual programming education.
Machine learning in financial Time Series Forecasting (TSF) has a clear application in investment, where predicting stock price movements can inform investment strategies. The Transformer model has emerged as a powerful tool for this purpose, yet significant research gaps remain. Existing studies often focus on a set of stocks without modeling stock behavior within specific sectors. Notably, there is a lack of research on the Consumer Cyclicals (CC) industry, which includes sectors such as automotive, housing, entertainment, and retail. These industries are highly sensitive to economic conditions, making them crucial for understanding broader economic impacts on stock behavior. Furthermore, we observe that many existing works neglect broader economic contexts, which is particularly important for CC analysis due to its sensitivity to economic trends. Additionally, previous studies on financial Transformer models typically use model tokens as feature vectors of multiple variates at a single timestep. This approach may not adequately capture the important relationships between dataset variates for long-term economic trends. To this end, we present the first study on financial TSF for the CC sector along with economic data. To support the research, we propose the first public benchmark dataset for the CC sector, consisting of traditional stock price time series, technical indicators, and temporal data, enriched with economic indicators. Next, we introduce an alternative tokenization approach to enhance the model’s ability to capture long-term trends by preserving information about nonlinear dependencies between dataset variates. We hypothesize that this approach helps capture long-term signals more effectively. Through a comprehensive data ablation study and benchmark testing, we demonstrate that incorporating economic indicators improves the accuracy of longer-term predictions for the CC sector, and the new tokenization method enhances the performance of Transformer models. The dataset and code are made publicly available at: https://github.com/KasperKrawczyk/econtrans_dataset
Ophthalmologists commonly use retinal fundus images for diagnosis. Recently, automation of this process using deep learning has gained significant attention. Multiclass classification, which distinguishes among multiple eye diseases, is more representative of actual clinical settings, however, it presents challenges such as limited availability of annotated datasets, class imbalance, overlapping clinical features across various eye diseases and disease heterogeneity. This study develops deep learning models for multiclass classification of three major eye diseases—cataracts, diabetic retinopathy, and glaucoma—alongside normal cases. A larger and more diverse dataset was obtained by combining multiple publicly available, well-annotated datasets. Four deep learning models: VGG16, Inception-v3, ResNet50 and EfficientNet-B0, were deployed using a transfer learning approach. These models achieved test accuracies ranging from 74.29% to 78.79%, with ResNet50 performing the best, achieving an accuracy of 78.79%, precision of 80.04%, recall of 78.79%, and an F1-score of 78.76%. The results demonstrate the effectiveness of transfer learning for multiclass classification of eye diseases. Notably, the models were trained and evaluated on a heterogeneous dataset that simulates real-world variability in image acquisition, highlighting their generalization capabilities and robustness to inconsistency. The study provides valuable insights about the performance of pre-trained deep learning models under realistic conditions, supporting their potential as assistive diagnostic tools in actual clinical scenarios.
Aggregation operators are essential in multi-attribute decision-making, particularly for managing uncertainty and risk. Traditional methods, such as the Ordered Weighted Averaging (OWA) operator, typically address either probability-based weighting or uncertainty-based reordering, but rarely combine both within a unified framework. This paper proposes the Ordered Weighted Logarithmic Averaging Weighted Average (OWLAWA) operator, a novel approach that merges the benefits of weighted averaging and ordered reordering with a logarithmic transformation to better reflect decision-maker preferences under uncertainty. The theoretical properties of this operator including monotonicity, boundedness, and commutativity are formally established. A multi-attribute decision-making framework is then presented, integrating recognized expert weighting methods, including an entropy-based approach, to enhance decision robustness. Through comparative analysis and a sustainability-focused case study involving 20 companies, results demonstrate that the proposed approach yields a controlled sub valuation effect, particularly beneficial in risk-sensitive or compliance-driven environments. These findings indicate a more adaptive and structured decision-making process relative to conventional operators, accommodating both structured probabilities and uncertain preferences. By unifying risk-based and uncertainty-based weighting within a logarithmic formulation, this operator offers a versatile and structured tool for applications in financial risk management, policy evaluation, and supply chain optimization. Future research may explore its integration with fuzzy systems and machine learning methods, further expanding its adaptability in complex decision scenarios.
The recent advancements in deep learning have opened avenues for substantial improvements in text extracting accuracy. These technological innovations have significantly enhanced the capabilities of existing tools, particularly in the realm of Optical Character Recognition (OCR). The OCR plays a pivotal role in digitizing image-based texts. To enhance its efficacy, and advanced preprocessing model aimed at improving text clarity and readability should be implemented prior to the text extraction process. Our works explore the application of Generative Adversarial Networks (GANs) for super resolution tasks, aiming to enhance image quality by increasing resolution. Various GAN architectures and training techniques are being experimented including the usage of state-of-the-art model, Super Resolution Generative Adversarial Network (SRGAN). While the results showed some improvements quantitatively, it also highlighted areas for further optimizations. The findings suggest that GAN-based approach holds promise for super resolution tasks in enhancing text extraction result for document image. Significantly, the experiment showed that OCR successfully improved even when dealing with photos that had experienced 75% damage, which in this stage, an image had experienced substantial information loss. Future work will focus on addressing the identified challenges and enhancing model performance.
This article presents a model improvement to increase the efficiency of grape leaf disease classification using the VGG12 (Visual Geometry Group with 12 Layers) with Wide Layer model. It is a new model based on the concept of VGG16 (Visual Geometry Group with 16 Layers) and InceptionV3 Block. It aims to present a small convolutional neural network model. It reduces the number of parameters and computational costs but increases the efficiency of grape leaf disease classification. By reducing the number of layers of the VGG16 model from 16 to 12, horizontal feature maps are forwarded instead of hierarchical feature maps between the second to eighth layers. Feature maps are combined to forward the feature maps hierarchically to the next layer. In addition, the filter size was changed from 3×3 to 1×3 and 3×1 to reduce the number of parameters and help the model process faster. Multiple Dilated Convolution was used to obtain more feature maps and did not increase the parameters. The model’s results were evaluated by experimenting with a four-class grape leaf dataset from the PlantVillage dataset, consisting of three classes of diseased grape leaves and one class of healthy grape leaves. The results showed that the proposed model gave an accuracy of 99.95% in classifying grape leaf disease and comparing the proposed model with the models ResNet50, VGG16, InceptionV3, DenseNet121, and MobileNetV2, whose classification accuracies were 64.14%, 91.16%, 97.92%, 97.71%, and 99.69%, respectively. It was found that the model proposed has the highest accuracy, and it uses 1,571,839 parameters.
