Jurnal Nasional Teknik Elektro dan Teknologi Informasi

Student Behavior Detection Using YOLOv10 for Classroom Engagement Analysis

2026-05-13T08:47:31+07:00

Student engagement is a critical determinant of learning effectiveness, yet manual observation in classroom environments remains labor-intensive, subjective, and difficult to scale. This study examined a student behavior detection framework built on You Only Look Once (YOLO) version 10 or YOLOv10, the latest generation of real-time object detection models. A dataset of 2,600 annotated classroom images covering eight behavioral categories was collected under diverse conditions, including variations in lighting, camera perspectives, and occlusion. Five YOLOv10 variants (n, s, m, l, x) were trained and evaluated using precision, recall, F1 score, and mean average precision (mAP). The best-performing configuration achieved an overall mAP@0.5 of 0.821 and mAP@0.5:0.95 of 0.640, with strong performance on upright (AP = 0.967), bow head (AP = 0.958), and sleep (AP = 0.943), while more subtle behaviors such as writing (AP = 0.519) and hand-raising (AP = 0.650) proved challenging. Importantly, the system maintained real-time inference speeds ranging from 40 to 88 FPS depending on the YOLOv10 variant, when evaluated on an RTX 2060 GPU, thereby demonstrating its robustness for deployment in classroom settings. To ensure usability, the optimized YOLOv10 model was integrated into a Streamlit-based interactive dashboard, enabling educators to monitor engagement levels and respond with timely interventions. By combining state-of-the-art YOLOv10 architecture with real-time behavioral analytics, this work establishes a scalable foundation for intelligent classroom monitoring and contributes to advancing technology-enhanced education.

Optimizing YOLOv8 Architecture and Augmentation for Efficient License Plate Detection

2026-05-13T08:47:32+07:00

Automatic Number Plate Recognition (ANPR) is crucial for intelligent transportation systems but often falters in real-world conditions due to environmental variations. This study constructed a robust and computationally efficient vehicle license plate detection system that achieved high accuracy under diverse real-world challenges and was deployable on resource-constrained edge hardware for real-time operation. The proposed holistic framework integrated three key components: (a) the creation of the Dynamic Vehicle License Plate Dataset (DVLPD) v1.0, containing 866 annotated images with variations in lighting, weather, and camera angles; (b) the implementation of a targeted data augmentation pipeline employing geometric and photometric transformations to enhance model robustness; and (3) the architectural optimization of a You Only Look Once (YOLO) version 8 or YOLOv8 model through pruning, quantization, and hyperparameter tuning specifically for edge deployment. The optimized model achieved a mean average precision (mAP) of 91% on the test set. When deployed on a Raspberry Pi 4 in a prototype parking system, it demonstrated practical viability with an inference latency of 0.4 seconds per frame and an error rate of 4.2%. The results validate that the integration of a diverse dataset, strategic augmentation, and model optimization can yield an accurate and efficient ANPR solution suitable for real-time edge applications. Future work will focus on expanding the dataset to include more extreme conditions for greater generalization.

A Comparison of SR and CBAM for Optimized Thermal Drone Object Detection

2026-05-13T08:47:32+07:00

Human detection using thermal cameras is very useful in certain conditions, such as detecting people lost in mountainous areas that are difficult to explore. Rescue operations are usually conducted by deploying a search and rescue (SAR) team to the location, which is not always effective because this operation can only be carried out under certain conditions and may pose a risk to the SAR team itself. Therefore, one alternative approach is the use of drones equipped with human detection and recognition capabilities. In this context, thermal cameras are used because they can penetrate challenging environments, making them suitable for SAR operations. The object detection method used in this study was You Only Look Once (YOLO) version 8 or YOLOv8. This study aimed to compare the effectiveness of integrating enhanced super-resolution generative adversarial networks (ESRGAN) with YOLOv8 and incorporating a convolutional block attention module (CBAM) into the neck architecture of YOLOv8. The performance of ESRGAN with YOLOv8 and CBAM with YOLOv8 was evaluated using precision, mean average precision (mAP), and training loss. Based on the experimental results, the combination of ESRGAN with YOLOv8 outperformed the CBAM-based modification. This is indicated by higher precision and mAP values, as well as lower training loss in the ESRGAN-enhanced YOLOv8 detection framework. The experimental findings highlight that image enhancement using ESRGAN is more effective than CBAM-based modification in improving thermal image-based human detection performance for SAR applications.

Text Classification of Public Complaint Validity With Deep Learning Approaches

2026-05-22T13:45:22+07:00

Mobile apps Jakarta Kini (JAKI) recorded 173,327 public complaints in 2023, accounted for 91.37% of all public complaints to the DKI Jakarta Provincial Government. Reports submitted to the Cepat Respon Masyarakat system must be handled in accordance with the service level agreement (SLA) regarding handling time. Currently, the process of validating incoming reports is still done manually by officers, taking more than 30 minutes per report. In the same year, 15,634 complaints were recorded as unclear or invalid. This led to a decrease in the performance of local government agency, impacting 13% of them did not achieve a 100% SLA in 2023. This study aimed to automate the validity classification of public complaints to distinguish between valid and invalid reports. The study utilized a dataset of 2,000 reports and employed deep learning models, including Indonesian version of bidirectional encoder representations from transformers (IndoBERT) and multilingual BERT (mBERT), and to compare their performance against traditional machine learning baselines, including term frequency-inverse document frequency (TF-IDF) + extreme gradient boosting (XGBoost), naïve Bayes, and support vector machine (SVM) using a 5-fold cross-validation scheme. The results showed that the IndoBERT model could classify valid or invalid reports with an average accuracy of 88.8%, which was higher than other models. The implementation of this method has proven to increase the efficiency of report validation time with computation time of 6 minutes for 300 reports, thus helping government agencies achieve their SLA targets and contributing to research on the effectiveness of BERT in public complaint classification.

CNN-Based Transfer Learning Model for Early Detection of Diseases in Corn Plants

2026-05-22T14:34:34+07:00

Corn (Zea mays L.) is one of Indonesia’s primary food commodities, playing a crucial role in food security and the agricultural industry. However, its productivity is often compromised due to foliar diseases such as leaf blight, gray leaf spot, and common rust, significantly reducing crop yield. Conventional methods for disease detection rely on visual observation, which can be subjective, limited by the availability of agricultural experts, and delayed in disease control. To address these challenges, this study proposed a transfer learning-based approach utilizing convolutional neural networks (CNN) for early detection of maize diseases through digital images. The research implemented two widely used CNN architectures, Visual Geometry Group1 6 (VGG16) and residual network 101 (ResNet101), which were initialized with pretrained weights from ImageNet and fine-tuned to classify four maize leaf categories. The dataset consisted of 4,188 images, with 80% allocated for training and 20% for validation. Experimental results demonstrated that ResNet101 achieved the highest validation accuracy of 93.78%, with a validation loss of 0.2521, while VGG16 achieved a validation accuracy of 89.36% and a validation loss of 0.8905. These findings underscore the superiority of ResNet101 in terms of stability and generalization, whereas VGG16 is more efficient in computational resources. This study highlights the potential of transfer learning to facilitate rapid, accurate, and cost-effective disease detection, providing an essential tool for innovative farming applications in Indonesia, where limited data availability is often a barrier to implementing advanced artificial intelligence (AI) solutions.

Performance of PSO and GWO Optimized Triple Exponential Smoothing for Rice Forecasting

2026-05-25T13:52:02+07:00

Rice production forecasting is crucial for supporting food security policies, where Indonesia’s food security remains a key goal of the current government, particularly in a developing city like Malang. Triple exponential smoothing (TES) is a suitable and reliable forecasting method for limited, univariate, and seasonal data. TES is a statistical method whose accuracy depends on parameter selection, which is typically determined through trial and error. This study aimed to evaluate the results of TES forecasting optimized using well-known metaheuristic algorithms, particle swarm optimization (PSO) and grey wolf optimization (GWO). The novelty of this study lies in comparing the performance of PSO and GWO for TES optimization on limited rice production data in Malang City. The training data was rice production data for 2022–2024, and the testing data is rice production data for 2025. The study found that the accuracy of rice production forecasting using the pure TES method varied, with mean absolute percentage error (MAPE) ranging from 31.11% to 89.69%. Meanwhile, optimization significantly reduced the MAPE to 19.96% for PSO and 20.90% for GWO. The results showed that PSO produced a smaller standard deviation than GWO, indicating that PSO produces more stable forecasting results. However, at 100 iterations, GWO had a computation time of 0.078 s, shorter than PSO’s 0.136 s. The research findings recommend the use of metaheuristic algorithms to optimize rice production forecasting with limited and univariate data. Combining forecasting methods such as TES with metaheuristic algorithms has been shown to reduce MAPE, thereby improving forecasting accuracy.

Smart Greenhouse Data Integration Using Machine Learning Approaches

2026-05-29T10:46:24+07:00

In modern agriculture, smart greenhouse systems play an important role in improving agricultural productivity by integrating internet of things (IoT) technology and data-driven decision making. This approach is particularly relevant in hydroponic lettuce cultivation, where rapidly changing greenhouse conditions directly influence plant growth and yield, making accurate yield prediction a challenging task. This study aimed to compare the performance of three machine learning algorithms, namely random forest, decision tree, and naïve Bayes, in predicting hydroponic lettuce yield using data obtained from an IoT-based smart greenhouse system. A total of 15,492 data samples were collected from two greenhouse locations using sensors that measured air temperature, water temperature, humidity, light intensity, pH, and nutrient levels. Prior to model development, the dataset was preprocessed through data cleaning, normalization, and integration stages, and then divided into training and testing sets using an 80:20 ratio. Model performance was evaluated using mean absolute error (MAE) and the coefficient of determination (R²). The experimental results showed that the random forest model achieved the best performance with an MAE of 3.51 and an R² value of 0.8511, followed by the decision tree model with an R² of 0.851 and the naïve Bayes model with an R² of 0.7245. These findings indicate that integrating IoT-based smart greenhouse monitoring with machine learning models, particularly random forest, enables accurate crop yield prediction and supports effective decision making for sustainable hydroponic agriculture.

User-Centered Development of a Non-Halal Food Service Locator App

2026-05-29T14:42:00+07:00

The rapid growth of culinary tourism has increased demand for location-based applications that support specific dietary preferences. In Indonesia, particularly in Batam City, non-Muslim users often have trouble identifying non-halal food services due to the generalized nature of existing similar platforms, which rarely provide explicit non-halal categorization. This research aimed to develop a mobile-based non-halal food service locator application with accurate, location-aware restaurant recommendations tailored to non-halal culinary needs. The system was developed using a geographic information system (GIS) approach integrating design thinking and agile scrum methodologies. User requirements were identified through an empathy-oriented requirements-gathering process using structured questionnaires with 30 non-Muslim respondents. Agile scrum supported iterative development and feature prioritization. The application architecture consisted of a react native mobile client and a Laravel backend, supported by PostgreSQL with PostGIS for spatial data processing and Nominatim for geocoding. A GIS-based multi-factor recommendation algorithm was applied to rank nearby non-halal food services based on distance and contextual parameters. The system was evaluated through functional black-box testing covering 11 core use cases, all of which produced successful results. Usability evaluation using the system usability scale (SUS) involved 34 respondents and yielded an average score of 85.29, indicating acceptable system usability. Prototype validation using Figma Mirror further confirmed alignment between user expectations and interface design. In conclusion, this study demonstrates that integrating GIS technology with user-centered and agile development approaches can support non-halal food discovery in Batam City and provide a foundation for future system enhancements.

Game-Based Learning to Teach Body Safety for Early Childhood

2026-05-29T15:47:45+07:00

Sexual violence against children remains a serious and increasing concern in Indonesia, particularly among early childhood populations who often lack the cognitive and communicative skills necessary to recognize, resist, or report unsafe situations. Preventive education at an early age is therefore essential; however, conventional instructional approaches frequently face limitations in addressing sensitive topics such as body safety and personal boundaries. This study aimed to develop and evaluate a 2D platformer-based educational game designed to enhance body safety awareness among children aged 5–7 years in early childhood education settings. The game was developed using the waterfall model, encompassing requirement analysis, design, implementation, testing, and evaluation stages. Construct 3 was utilized as the development platform to create interactive modules covering body autonomy, good and bad touch, trusted adults, and assertive response strategies. A quasi-experimental pretest and posttest design was employed involving 20 preschool learners from two institutions in Malang. Learning outcomes were measured using visual scenario-based assessments, while usability was evaluated through a modified system usability scale (SUS) administered with teacher assistance. The results demonstrated a statistically significant improvement in children’s understanding, with average scores increasing from 52.5 in the pretest to 86.3 in the posttest (p < 0.01). Usability findings indicated high levels of engagement, clarity, and motivation to replay the game. These findings suggest that culturally adapted digital game-based learning can serve as an effective and developmentally appropriate medium for early childhood sexual violence prevention education.

Development of Noncontact Thermometer based Artificial Neural Network for Cattle Health Monitoring

2026-06-02T07:54:32+07:00

Manual measurement of animal body temperature is done by inserting a thermometer into the rectum or anus. This causes discomfort, stress, and the risk of disease transmission. Therefore, a method of measuring body temperature with a noncontact thermometer is needed. Noncontact sensors for human skin have different emissivity levels than cowhide surfaces. This study aimed to design and test a noncontact temperature measuring instrument based on an artificial neural network (ANN) model for cows. Observation parameters include performance tests, coefficient of determination (R²), root mean square error (RMSE), and relative RMSE (RRMSE). The development process of the ANN used 2 input layers, 3 hidden layers 1, 2 hidden layers 2, and 1 output layer. The type of training used was trainlm, and a learning rate of 0.001, producing the best prediction with the activation function tansig-logsig-tagsig. The sensor design produced R², RMSE and RRMSE values that were in accordance with the research parameters, namely obtaining an R² value of 0.9909 or 99.09%, an RMSE value of 0.6361 and an RRMSE value of 0.13% on cowhide. The design of the sensor device tested on live cows produced an R² value of 0.6429 or 64.29%. The most accurate distance was in the range of 4 cm to 6 cm with an error value of 0.606. The noncontact thermometer is able to measure the body temperature of cows at close range with high accuracy.

Front Pages

2026-06-09T16:00:26+07:00

Back Pages

2026-06-09T15:58:41+07:00