The inability of the general population to understand sign language creates serious communication barriers for hearing-impaired individuals in critical situations such as healthcare emergencies, educational settings, and workplace environments. Addressing this gap requires effective and accurate automated recognition systems. This proposed research presents a real-time American Sign Language (ASL) recognition system for 24 dynamic signs that integrates the MediaPipe frame work with Long Short-Term Memory (LSTM) network. To enhance performance while reducing computational complexity, only the most relevant features are extracted from self-recorded dynamic sign videos: coordinates of 65 key hand and body landmarks, complemented by 41 engineered angle features between joint connections and 26 engineered distance features between specific landmark pairs, yielding a compact 285-feature representation per frame. As a result, LSTM network handles spatiotemporal sequence modeling across 25-frame sequences, effectively capturing the dynamic nature of sign language gestures. The proposed system achieves 98% test accuracy, with precision, recall, and F1-score of 98% across 720 test samples. It also successfully interprets all 24 dynamic signs in real-time testing scenarios using a standard webcam, including visually similar sign pairs such as “Good”/“Bad” and “Mother”/“Not”, demonstrating its practical applicability for assistive communication technologies. The research contribution lies in the systematic integration of MediaPipe Holistic with strategic feature engineering (combining landmark coordinates with computed angles and distances) and LSTM modeling to achieve efficient real-time dynamic sign recognition with reduced computational and data requirements.
Accurate ranking remains challenging in session-based recommender systems, particularly in dynamic domains such as e-commerce and digital news. Neural models can achieve high accuracy but require substantial computational resources, while Association Rule (AR)-based methods are computationally efficient yet may rank relevant items poorly when candidate weights are similar. The research aims to improve the ranking quality of AR-based session recommendations while retaining computational efficiency through Association Rule with Top-k Quartile Filtering (ART-Q). The proposed framework constructs an association-rule dictionary, partitions each consequent set into four quartiles according to rule weights, and independently selects top-k candidates from each quartile. ART-Q is evaluated using two real-world datasets: YooChoose (YOO) and Malang Posco Media (MPM). Performance is evaluated using HR@20, MRR@20, and NDCG@20 against AR, k-Nearest Neighbor (KNN), and neural baselines. On YOO, ART-Q-V6 achieves HR@20 of 0.6899, MRR@20 of 0.7500, and NDCG@20 of 0.5436, representing gains of 11.9% in HR and 31.6% in Normalized Discounted Cumulative Gain (NDCG) over traditional AR, with more than twofold improvement in Mean Reciprocal Rank (MRR). On MPM, ART-Q-V1 achieves HR@20 of 0.7498, MRR@20 of 0.2717, and NDCG@20 of 0.3260, outperforming all evaluated baselines. ART-Q also completes training and inference in less than two minutes on a laptop CPU, demonstrating its computational efficiency. These results indicate that quartile-based filtering effectively reduces ranking ambiguity while maintaining a lightweight recommendation framework.
Convolutional Neural Networks (CNNs) continue to face several challenges in cataract classification using image data, particularly due to limitations in dataset size and variability in color, shape, and position. These constraints arise because CNNs primarily process local features within each layer. The research uses Vision Transformer (ViT) to capture global spatial relationships in front-eye images. ViT can detect cataracts more accurately than CNN-based owing to its capability of modeling long-range dependencies. The research aims to evaluate ViT’s performance on four publicly available cataract datasets using training time, accuracy, precision, recall, and F1-score. To assess the model’s reliability when working with a small number of training datasets, ViT’s performance is also be compared with ResNet-50 and EfficientNet-B7. The results indicate that ViT outperforms ResNet-50 in accuracy by 10%–31% and exceeds EfficientNet-B7 by 30%–41%. However, ViT requires approximately 16 seconds longer to train than ResNet-50 and EfficientNet-B7 due to its deeper architecture. Although ViT’s accuracy is 2.47% lower than previous studies using hybrid deep learning approaches, its parameter structure is simpler. Overall, the findings indicate that, compared with CNN-based models, ViT performs more effectively when trained on smaller datasets. Future research may incorporate segmentation techniques to further validate cataract detection in anterior eye images.
Cardiovascular disease (CVD) is a leading cause of global mortality, demanding accurate and interpretable risk stratification models for early intervention. However, the tradeoff between model performance and clinical interpretability remains a gap. Therefore, the research evaluates lightweight parametric ensemble models against complex black-box alternatives. A comparative analysis of Logistic Regression (LR) and Gaussian Naive Bayes (GNB) is conducted using a publicly available dataset of 1,000 patient records, enhanced with Bagging and AdaBoost ensemble methods. The data undergo standardized preprocessing to mitigate bias, and a fivefold stratified cross-validation protocol ensures model generalizability. The Bagging LR ensemble achieves the best predictive performance, with a mean accuracy of 0.966 (95% Confidence Interval (CI): [0.958, 0.974]) and a Receiver Operating Characteristic-Area Under the Curve (ROC-AUC) of 0.993 (95% CI: [0.990, 0.996]), highly competitive with state-of-the-art baselines including LightGBM and CatBoost. These results show that computationally efficient and interpretable ensembles provide a viable alternative to more complex models, particularly in data-constrained clinical settings. A comprehensive Explainable Artificial Intelligence (XAI) analysis using SHapley Additive exPlanations (SHAP) identifies clinically congruent drivers of prediction, the slope of the peak exercise ST segment, ST depression, and chest pain type. The findings highlight the potential of interpretable ensemble learning to develop dependable, deployable clinical decision support tools for integration into hospital or e-health systems for early CVD risk stratification. Further validation on larger, more diverse datasets is recommended to confirm broader clinical applicability.
Deoxyribonucleic Acid (DNA)-based data storage as a revolutionary approach to long-term, highdensity information storage faces the challenge of high costs and biological constraints in DNA synthesis. Unlike conventional binary encoding, DNA storage requires efficient data compression techniques to minimize the length of the resulting DNA sequence. The research explores the use of a Variational Autoencoder (VAE) as a deep learning-based compression method to reduce image complexity. VAE enables significant data reduction while retaining essential structural features. The dataset used is from the Modified National Institute of Standards and Technology (MNIST) database. The proposed method involves encoding images into latent variables, followed by binarization and translation into DNA sequences while applying biological constraints such as maintaining GC balance and avoiding homopolymers, to ensure stability and sequencing accuracy. Reconstruction quality is assessed using the Structural Similarity Index Measure (SSIM) and Peak Signal-to-Noise Ratio (PSNR), yielding an average SSIM of 0.8017 and an average PSNR of 18.91 dB. Analysis of variance shows that digits with greater variability suffer more noticeable degradation, whereas digits with lower variability preserve their structural integrity. These results suggest that integrating VAE into DNA data storage can reduce sequence complexity while still preserving recognizable image structures. The research contributes to the advancement of DNA-based data storage by presenting optimized coding techniques.
Revision Presently, Internet usage in society has grown considerably. It presents significant opportunities for companies to develop business related to internet service providers (ISPs). In Indonesia, both domestic and foreign ISPs are competing. To survive and remain popular with Indonesian customers, ISPs need to improve their products and services quickly. Hence, companies need to collect user feedback through customer loyalty applications and reviews to improve their service or technology. However, manually auditing thousands of textual reviews is labor-intensive and subject to human error, making it highly inefficient for agile, real-time corporate decision-making. The research aims to address this operational bottleneck by improving the automation process and providing decision-making information to ISP management. More specifically, the research contributes to implementing and comparing two transfer learning models: Bidirectional Encoder Representations from Transformer (BERT) and IndoBERT for tokenization and embedding. The embedding results are then fed into a random forest for classification. The dataset is collected via a scraping process from the Play Store application, focusing on user feedback and ratings. A total of 1,192 records of user feedback reviews from ISP loyalty apps on the Play Store in March 2024 were gathered. These are manually labeled by a general manager from the company into binary and multi-class categories. Splitting the dataset into 70% for training and 30% for testing achieves a classification accuracy of 70% for the problem domain with IndoBERT + Random Forest. However, multi-class classification for categorizing user feedback into a priority score model achieves an accuracy of 58%.
The integration of Generative Artificial Intelligence (Generative AI) into educational websites has expanded digital service capabilities through chatbots and dynamic content systems. However, it has simultaneously introduced novel security vulnerabilities, such as prompt injection, data leakage, and insecure output handling, that remain insufficiently examined, particularly in the Indonesian context where rapid AI adoption outpaces security readiness. The research aims to develop and validate a hybrid security testing framework capable of detecting and mitigating vulnerabilities specific to Generative AI-based websites. The research adopts a Design Science Research (DSR) approach with an experimental mixed-methods strategy, combining quantitative security testing through penetration testing and adversarial red teaming with qualitative analysis based on threat modeling. The evaluation is conducted on Generative AI-enabled educational websites and simulated prototypes using Open Web Application Security Project (OWASP) Top 10 for Large Language Models (LLM) risk mapping, the Spoofing, Tampering, Repudiation, Information Disclosure, Denial of Service, Elevation of Privilege (STRIDE) framework, and adversarial benchmarks. The results indicate that dominant vulnerabilities are associated with prompt injection, information leakage, and insecure model output handling. The proposed framework significantly improves detection effectiveness with an accuracy rate of 92.6% while reducing attack success rates following layered remediation. These findings confirm that conventional web security mechanisms are insufficient for LLM-based systems and require adaptive, AI-specific testing approaches. The research contributes by strengthening the integration of AI security frameworks in the form of a new OWASP-STRIDE hybrid model and practically by providing actionable security testing guidance for developers and educational institutions seeking to deploy Generative AI more securely.
The research aims to develop a hybrid Ant Colony Optimization (ACO)-Deep Neural Network (DNN) for efficient resource allocation and the minimization of path loss in an Internet of Things (IoT)-driven 5G network. IoT-driven data are collected from the Zenodo repository, containing 8,484 datasets with noisy information. The data are preprocessed using Exploratory Data Analysis (EDA) to remove outliers and missing values. The processed data are split into an 80-20% set. The training set (80%) is utilized for the ACO and DNN models, while the testing set (20%) evaluates the system’s performance. ACO is used for feature selection and identification of suitable features for learning and prediction, which feed into the DNN model for learning and prediction. The model is then evaluated with Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and root square (R2) value to determine the optimal path for resource allocation and reducing signal loss in 5G networks. In the results, the ACODNN-based model outperforms the others with the lowest RMSE (0.42) and MAE (0.35). It shows high accuracy and minimal error in optimizing path loss. The ACO-DNNbased model also achieves effective resource allocation with an R2 of 0.92 and low error rates. The results underscore the efficiency of the hybridized ACO-DNNbased model for resource allocation while minimizing the path loss in the transmission link.
Retinal blood vessel segmentation is crucial for detecting and monitoring retinal disorders such as diabetic retinopathy, age-related macular degeneration and glucoma. Automating the segmentation of blood vessels leads to a reduction in the time and cost of manual segmentation, enables large-scale clinical studies, improves accuracy, ensures consistency, allows for realtime analysis, and facilitates early disease detection. The research examines the performance of 7 semantic segmentation architectures, each combined with 10 pretrained backbones, on 5 publicly available fundus image datasets. Models are trained on a NVIDIA GeForce GTX 1080 Graphics Processing Unit (GPU), and key hyperparameters, such as batch size, optimizers, and learning rate schedulers, are systematically optimized. Intersection over Union (IoU), accuracy, sensitivity, and computational time are used as key performance indicators. Approximately 97 experiments are conducted to achieve state-of-the-art accuracies of 97.72%, 98.23%, 97.62%, 97.83%, and 98.42%, along with IoU scores of 67.82%, 66.29%, 63.89%, 71.34%, and 78.45% on the DRIVE, STARE, HRF, HEI-MED-1, and HEI-MED-2 datasets, respectively. The best performance is achieved using the U-Net++ architecture with ResNeSt backbone, RAdam optimizer, and Cosine Annealing scheduler. This combination leverages deep supervision, attention mechanisms, and bottleneck architectures to enhance multiscale feature learning, localization, robustness to image variability, and model generalization. Although the models demonstrate strong performance, challenges remain in addressing dataset imbalance and ensuring generalization to unseen patient populations.
Competency level prediction plays a crucial role in competency-based human resource management such as talent management. Talent management is achieved by identifying individuals’ knowledge, skills, and attitudes through psychological assessment. Recognizing employees as a strategic asset by accurately predicting competencies supports targeted development, boosting individual and organizational performance. Current practices related to competency assessment require expert judgment from psychologists or assessors, which can be time-consuming. The research proposes a machine learning–based approach to predict competency levels using psychological assessment scores as input, designed to operate within digital, network-enabled interview platforms. Several machine learning methods, including Random Forests, k-Nearest Neighbors (KNN), and Support Vector Machines (SVMs), are applied to historical assessment datasets to identify patterns and relationships between psychological assessment scores and competency levels.The dataset comprises 1,220 records from a psychological assessment. The experimental results indicate that the Random Forest model achieves the highest accuracy of 81%, outperforming other models in competency level prediction. The key novelty lies in its data-driven methodology, which enhances the objectivity and efficiency of competency evaluation while reducing reliance on expert interpretation. By enabling automated competency prediction in network-enabled interview environments, the proposed approach supports more efficient talent decision-making, workforce development, and recruitment processes. The findings demonstrate that machine learning can accurately predict competency levels from a clean dataset of psychological assessment scores, achieving accuracy above 80%. Future research may enhance model robustness by incorporating additional assessment center criteria and real-world performance metrics.
In clinical practice, precise and high-quality brain Magnetic Resonance Imaging (MRI) is pivotal for diagnosing and formulating effective treatment strategies. The research objective is to assess the viability of employing a Convolutional Autoencoders (CAE) for the mitigating noise in brain MRI images. The focus is brain MRI images and the various types of noise (Salt and Pepper, Speckle, and Gaussian noise) that typically corrupt images and may lead to inaccuracies in diagnosis. The research also applies methods to artificially generate these noise types to represent real-world scenarios. Specifically, the dataset of brain MRI images is collected, pre-processed, and artificially exposed to various noise types to simulate the real-world conditions after the CAE model is used to reconstruct the corrupted images. The CAE is assessed for its high efficiency and efficacy using Mean Squared Error (MSE) and Peak Signal-to-Noise Ratio (PSNR). The results indicate that the CAE is very effective in removing noise, particularly Salt and Pepper noise. The model achieves a PSNR of 27.0687 dB and an MSE of 0.00216246 at the lowest noise level. The model also demonstrates stability under varying levels of Speckle noise. Although performance degrades as noise increases, the model continues to demonstrate potential for further refinement. The research furthers the CAE’s analytical potential by assessing its denoising capabilities across various noise types and levels. The research adds value by outlining recommendations to the medical imaging community while identifying the need for future research on different classifications of noise and advanced regularization methods.
The research presents a stacking-based hybrid intrusion detection framework for web application attacks, addressing the persistent limitation that minority classes, including Brute Force, Cross-Site Scripting (XSS), and Structured Query Language (SQL) Injection, are frequently underdetected in conventional Intrusion Detection Systems (IDS) due to severe class imbalance. The proposed architecture combines LightGBM and Random Forest as base learners, while a Multi-Layer Perceptron (MLP) functions as the meta-learner. The framework is supported by rigorous preprocessing, ANOVA F-testbased feature selection, and domain-informed augmentation of critical traffic features, such as Flow Inter-Arrival Time (IAT) Min, Init Win bytes forward, and Backward (Bwd) Packets/s, through optimized weighting strategies. Evaluation on the CICIDS-2017 web attack subset using 10-fold stratified cross-validation shows that the proposed model improves the macro F1-Score from 0.62 ± 0.004 to 0.76 ± 0.003 and achieves a binary accuracy of 99.67% with a macro F1 of 0.94. The observed performance gains are statistically significant (p < 0.001), confirming the robustness of the framework. These findings indicate that targeted feature engineering and heterogeneous stacking substantially improve minority-attack detection while preserving majority-class performance. In addition, the framework demonstrates sub-millisecond inference time, highlighting its practical suitability for real-time IDS deployment in resource-constrained and high-throughput operational cybersecurity environments. The proposed design also offers methodological generalizability for broader anomaly detection tasks in dynamic network environments, where reliable recognition of low-frequency but high-impact attack patterns remains increasingly critically important.
The research aims to build a hybrid deep learning model for sentiment analysis of Indonesian ecommerce product reviews, which represent the expressed opinions of customers. A major challenge in the domain is the presence of non-standard language and highly imbalanced sentiment classes, which hinder accurate classification. Most existing Indonesian sentiment analysis studies rely on relatively small and balanced datasets and primarily use attention mechanisms, an ensemble model, as well as a sequential fusion method. In the research, a large-scale dataset of Indonesian product reviews is collected from the largest e-commerce site in the country. The dataset consists of review text and corresponding product ratings. After preprocessing, semantic features are extracted using a pre-trained Indonesia Bidirectional Encoder Representations from Transformers (IndoBERT) model. The features are then fed into a hybrid model combining Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) layers through parallel feature-level fusion. Model hyperparameters are optimized using the Tree-Structured Parzen Estimator (TPE), while data imbalance is addressed through resampling methods. Regularization strategies are also applied to mitigate overfitting, and the model is evaluated using stratified k-fold cross-validation. The model hyperparameters are validated using a learning curve, showing a stable and consistent curve following the trend. The results show that the hybrid CNN-LSTM model, combined with Support Vector Machine Synthetic Minority Oversampling Technique (SVMSMOTE), achieves superior performance in distinguishing positive and negative reviews. This outcome reaches Receiver Operating Characteristic - Area Under the Curve (ROC AUC) score of 92.48%, outperforming baseline and conventional machine learning models. These results also show good generalization ability, characterized by consistent values with a very low standard deviation of 0.0009 for each fold.
Natural disasters in archipelagic regions often disrupt communication networks, particularly in geographically isolated islands where terrestrial infrastructure is limited and highly vulnerable. Hence, adaptive, infrastructure-independent solutions are required to maintain connectivity during emergencies. The research proposes an adaptive routing protocol for Delay Tolerant Network (DTN), named Q-learning-based Forwarding Routing (QFR), designed to enhance data delivery performance in disaster scenarios characterized by intermittent connectivity and constrained resources. QFR employs a lightweight, tabular Q-learning framework to make intelligent forwarding decisions based on real-time state information, including buffer occupancy, encounter history, and local node density. The protocol further integrates adaptive replica control and prioritybased scheduling mechanisms to regulate congestion and optimize bandwidth and buffer utilization. Performance evaluation is conducted using the ONE Simulator with realistic maritime mobility traces derived from vessel movement patterns around Madura Island, Indonesia, representing inter-island emergency communication conditions. The results indicate that QFR consistently outperforms benchmark protocols such as Epidemic and PRoPHETv2, particularly in maintaining a high delivery ratio under heavy traffic loads while keeping routing overhead moderate and latency stable. Time-series analysis further demonstrates QFR’s ability to improve its performance over time as the agent learns. The key finding is that a lightweight, adaptive algorithm based on a tabular Q-learning framework provides a practical and effective solution for reliable communication in resource-constrained emergency networks, avoiding the computational complexity of deep reinforcement learning approaches.
The research presents a novel electric bike rental system that operates efficiently without an internet connection. The system integrates a logistic map algorithm into an embedded microcontroller and a web-based rental information system. The logistic map has pseudo-random code generation properties, allowing identical activation tokens to be generated on both the server and the embedded system, enabling robust token validation on e-bike units. Users rent e-bikes by scanning a barcode on the e-bike unit. The barcode then directs the user to a web interface, where they can select the rental duration via a secure payment gateway. The rental system is pay-per-use, eliminating the need for deposits or account registration and enhancing user convenience. Once payment is verified, a message is sent containing the receipt details and an activation token that can be used on the e-bike. The obtained token generation and verification times are 0.68 ms and 0.72 ms, respectively. This test demonstrates the ability to generate tokens in real time without significant latency. However, this approach faces practical issues of token collisions and hardware limitations. Token collisions occur when a 9-digit token appears twice across five different ID tokens, with counters ranging from 1 to 10,000 for the same e-bike ID and rental duration. Rounding to significant digits causes this problem due to the microcontroller’s inherent limitations. However, these duplications occur at widely separated counters, making them difficult to exploit. This system effectively balances simplicity, security, and scalability for local e-bike sharing services.
Sentiment and emotion analysis of mobile application reviews has attracted significant attention as a means to understand users’ perceptions and experiences. The research proposes a novel Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model for multi-task sentiment and emotion classification on Indonesian TikTok application reviews. A large-scale corpus consisting of 500,000 reviews is collected from the Google Play Store and preprocessed through cleaning, normalization, tokenization, stopword removal, and stemming. Sentiment labels (positive, negative, and neutral) are assigned using a lexicon-based approach, while emotion labels are annotated through emoji analysis and word matching based on five basic emotions: anger, fear, happiness, love, and sadness. The proposed CNN-LSTM model is evaluated against a hybrid Bidirectional Encoder Representations from Transformers – Convolutional Neural Network (BERT-CNN) architecture. Experimental results show that the CNN-LSTM model outperforms the BERT-CNN model, achieving an accuracy of 91.30% for sentiment classification and 99.15% for emotion classification, compared to 42.43% and 72.85%, respectively, obtained by the BERT-CNN model. These findings indicate that the CNN-LSTM architecture is more effective in capturing sequential patterns and contextual features in Indonesian review texts, particularly in a multi-task learning setting. Despite its strong performance, the research is limited by its focus on a single platform and the use of lexicon-based automatic labeling, suggesting future work on cross-domain evaluation and manual annotation refinement.
The Sundanese language, as one of Indonesia’s regional tongues, holds deep cultural value but is still underrepresented in computational linguistics. The research addresses this gap by developing a translation model between Sundanese and Indonesian using a transformer-based sequence-to-sequence (Seq2Seq) architecture. With a parallel dataset of 3,616 sentence pairs, the model is fine-tuned to capture linguistic and contextual subtleties. The evaluation yields strong results: Bilingual Evaluation Understudy (BLEU) score of 44.12, Recall - Oriented Understudy for Gisting Evaluation (ROUGE)-1 F1-Score of 0.72, and ROUGE-L F1-Score of 0.71. Those demonstrate high translation quality despite limited data. Unlike earlier Sundanese translation studies that rely on Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), or standard transformer models, this research uniquely leverages the multilingual pretrained M2M100 Transformer, enabling transfer learning from high-resource languages to improve low-resource performance. These outcomes highlight the model’s potential for real-world applications, such as translation tools for education and cultural exchange. The research emphasizes the importance of improving access to Sundanese texts and promoting its digital presence to aid in language preservation. Overall, the research not only advances Natural Language Processing (NLP) research for low-resource languages but also reinforces the importance of integrating regional languages like Sundanese into modern technology. Building upon prior studies on Indonesian–Sundanese translation, the research novelty lies in fine-tuning a multilingual Seq2Seq Transformer that captures both linguistic and contextual nuances, thereby setting a new benchmark for lowresource language processing.
Soil nutrient assessment is essential for optimizing crop yield. Still, existing machine learningbased soil nutrient recommendation systems face several challenges, including limited real-time adaptability, inconsistent integration with IoT frameworks, and a lack of scalability for smallholder use. Many of these systems rely heavily on static, pre-collected datasets and lack the capability to respond dynamically to field conditions. These limitations reduce the practical effectiveness of such models in achieving precision agriculture goals, particularly in resource-constrained environments. These limitations also hinder efficient soil fertility management, leading to ineffective fertilizer application, nutrient imbalances, and reduced crop productivity. To address these issues, the researchers develop an Internet of Things (IoT)-enabled K-Nearest Neighbors (KNN)-based soil nutrient recommendation system specifically for rice (Oryza Sativa L.) cultivation. The system integrates an RS485 Integrated Soil Nitrogen, Phosphorus, and Potassium (NPK) Sensor with an Arduino-based IoT framework to continuously monitor essential nutrients: nitrogen (N), phosphorus (P), and potassium (K). The collected data are processed using Naive Bayes, Support Vector Machine (SVM), Logistic Regression, Linear Regression, and KNN. After rigorous model training on Google Colab, KNN outperforms other models with an accuracy of 98%, making it the optimal choice for predictive soil fertility assessment. This system provides accurate and automated fertilizer recommendations to improve soil management efficiency and sustainability by combining real-time IoT monitoring with machine learning. The research contributes to precision agriculture by offering a scalable data-driven approach that enhances crop yield, reduces fertilizer waste, and minimizes environmental impact.
Tobacco is one of the leading agricultural commodities in Indonesia, making a significant contribution to the local economy, particularly in tobaccoproducing regions. However, tobacco cultivation, which generally takes place during the dry season, faces challenges such as limited water availability and the high labor intensity required for irrigation. The research aims to develop an Internet of Things (IoT)-based drip irrigation system to enhance water-use efficiency and simplify the irrigation process for tobacco farmers. The system integrates a DHT11 temperature sensor, a soil moisture sensor, and a soil pH sensor. An Arduino Uno and an ESP8266 microcontroller are used to process sensor data and transmit it in real-time to Firebase. Moreover, Mamdani fuzzy logic method is applied to determine irrigation duration based on temperature and soil moisture readings. Experimental results indicate that the system can reduce water usage by up to 36.67% compared to conventional manual watering methods, with an average water consumption of 297.32 mL per automated irrigation cycle. Moreover, the system demonstrates high accuracy, with an average deviation of only 0.33 between fuzzy logic results generated by the Arduino Uno and MATLAB simulations. The novelty of the research lies in the integration of an IoT-based drip irrigation system utilizing Mamdani fuzzy logic, specifically designed for tobacco cultivation, which enables real-time monitoring by farmers. This system is expected to offer an innovative solution to support precision agriculture and promote efficient water resource management.
