Green Intelligent Systems and Applications

Integration of Naïve Bayes-Based Stunting Status Classification and GIS Hotspot Mapping for the Identification of Priority Areas in Tomohon City, Indonesia

2026-06-19T07:44:33+00:00

Stunting remained a public health problem that required data- and area-based monitoring so that interventions could be implemented in a targeted manner. This study aimed to develop an integrated system for classifying stunting status and identifying priority areas in Tomohon City through the combination of WHO Z-Score standards, the Naïve Bayes algorithm, prevalence calculation, and hotspot mapping based on a Geographic Information System (GIS). This study employed a Research and Development (R&D) approach consisting of needs analysis, design, implementation, testing, and evaluation stages. Toddler data were obtained from the Tomohon City Health Office, including age, sex, height or body length, weight, residential area, urban village, district, and community health center. The system was developed using MySQL, Python, PHP Framework CodeIgniter 3, and GIS. The results showed that the system was able to classify toddlers’ nutritional status into normal, stunted, and severely stunted categories, calculate prevalence by urban village, and display the distribution of cases in the form of a digital map. Gaussian Naïve Bayes modeling using 970 training data points and 243 testing data points produced an accuracy of 94.7%, precision of 31.6%, recall of 33.3%, and F1-score of 32.4%. GIS hotspot visualization helped identify priority areas, although data coverage still needed to be expanded to make the results more representative.

Comparison of Convolutional Neural Network Model for Brain Tumor Disease Gliome Detection

2026-03-05T04:26:31+00:00

Glioma represented one of the most aggressive forms of malignant brain tumors, necessitating early detection to optimize therapeutic intervention outcomes. Manual tumor identification through Magnetic Resonance Imaging (MRI) was labor-intensive and was susceptible to subjective interpretation errors. This study aimed to compare the performance of two Convolutional Neural Network (CNN) architectures, specifically Residual Network (ResNet) and U-Net, for glioma tumor detection in T2-weighted MRI sequences. The datasets employed were obtained from the BraTS and Kaggle repositories and underwent comprehensive preprocessing procedures, including normalization, augmentation, and conversion to Portable Network Graphics (PNG) format. The evaluation metrics demonstrated that the U-Net architecture exhibited superior performance compared to ResNet-18, achieving an accuracy of 88.16%, sensitivity of 80.00%, specificity of 88.43%, and F1-score of 68.97%. Conversely, ResNet-18 yielded an accuracy of 71.43%, sensitivity of 73.52%, specificity of 81.54%, and an F1-score of 70.14%. These findings indicated that U-Net demonstrated greater efficacy in recognizing tumor morphology within MRI data and preserving spatial information through its inherent skip connection mechanism. This investigation demonstrated the potential of the U-Net architecture to facilitate automated and enhanced accuracy in glioma detection, although further refinement was required to improve segmentation precision and clinical applicability.

Comparison of Tea Leaf Disease Classification Using SVM with MobileNetV2 and MobileNetV3-Small Feature Extractors

2026-05-31T01:16:51+00:00

Tea is a strategic plantation commodity that serves as a major source of income for millions of rural families. However, its production is often threatened by devastating pests and diseases. Accurate and timely classification of diseases such as brown blight, gray blight, and tea algal leaf spot is crucial for maintaining crop quality. Traditional identification methods often involve observer subjectivity and require significant time. Although Convolutional Neural Networks (CNNs) have demonstrated effectiveness in automatic recognition, their application on mobile devices is often limited by high computational demands. Previous studies in the tea domain that use MobileNet as a feature extractor combined with an SVM classifier are still limited. Therefore, this study evaluates the implementation of this hybrid model for tea leaf disease classification. This study compares two models: MobileNetV2-SVM and MobileNetV3-Small-SVM, using the TeaLeafBD dataset. Empirical testing shows that both architectures achieve very comparable classification performance, with accuracy rates of 75.3% for MobileNetV2 and 75.1% for MobileNetV3-Small. Despite marginal differences in accuracy, the MobileNetV3-Small-SVM hybrid offers a lower computational footprint, reducing computational load by approximately fivefold and model size by more than half. These findings indicate that the MobileNetV3-Small-SVM architecture provides a favorable balance between recognition stability and resource efficiency. Consequently, this hybrid approach is a viable candidate for the development of on-site tea leaf disease diagnostic tools on resource-constrained mobile devices.

Design and Implementation of a Multi-Node Gas Sensor-Based Indoor Air Quality Monitoring and Control System

2026-03-05T04:26:33+00:00

: Air quality monitoring was a crucial aspect of maintaining occupational health and safety, particularly in industrial environments. This study proposed the design and implementation of an Internet of Things (IoT)-based indoor air quality monitoring system capable of measuring environmental parameters in real time. The system integrated an ENS160 gas sensor and an AHT21 temperature–humidity sensor with a Wemos D1 Mini microcontroller. Sensor data were transmitted via the MQTT protocol to an Orange Pi 4A server and visualized using a Node-RED dashboard. The monitored parameters included Total Volatile Organic Compounds (TVOC), equivalent CO₂ (eCO₂), temperature, and humidity. Experimental evaluation demonstrated that the system responded proportionally to different pollutant exposure levels. Under high NH₃ exposure (100%), TVOC values reached a maximum of 12,697 ppb with an average of 5,037 ppb, clearly exceeding the hazardous threshold (>200 ppb). At moderate exposure (50%), the average TVOC decreased to 2,106 ppb, while at low exposure (10%), the average value remained within the safe range at 84 ppb. For eCO₂ testing, cigarette smoke exposure produced a peak value of 11,524 ppm with an average of 1,663 ppm, indicating hazardous conditions (>1000 ppm). Statistical analysis using mean and standard deviation confirmed that sensor stability improved at lower pollutant concentrations. The proposed system successfully provided stable real-time monitoring, threshold-based classification, and automatic mitigation control, demonstrating its feasibility for intelligent indoor air quality management in industrial workspaces.

Lightweight Rice Leaf Disease Classification Using MobileNetV2: A Comprehensive Performance Evaluation

2026-06-08T00:21:32+00:00

Rice leaf diseases pose a significant threat to agricultural productivity, and accurate automated detection is essential for timely intervention. This study presents a comparative evaluation of lightweight convolutional neural network architectures for the classification of six rice leaf disease categories: Bacterial Leaf Blight, Brown Spot, Leaf Blast, Leaf Scald, Narrow Brown Leaf Spot, and Rice Healthy. MobileNetV2 is proposed as the primary model and benchmarked against EfficientNetB0 and NASNetMobile. All three architectures were trained under an identical experimental setup comprising a two-stage transfer learning strategy, a unified custom classification head consisting of Global Average Pooling, Batch Normalization, two dense layers with dropout and L2 regularization, and a Softmax output layer. The dataset comprised 1,920 images across six classes obtained from Roboflow Universe, with no pre-augmentation applied by the original source. Training-time augmentation including rotation, shifting, shearing, zooming, and horizontal flipping was applied exclusively to the training subset. Experiments were conducted on a stratified split of 1,536 training, 192 validation, and 192 test images with a fixed random seed of 42 to ensure reproducibility. MobileNetV2 achieved the highest test accuracy of 96.35% and macro F1-score of 96.35%, outperforming EfficientNetB0 at 94.27% and NASNetMobile at 89.06%. In terms of computational efficiency, MobileNetV2 also demonstrated the most favorable deployment profile with a TensorFlow Lite model size of 2.75 MB and inference latency of 3.22 ms per image, indicating potential suitability for resource-constrained deployment scenarios. These results suggest that MobileNetV2 offers a competitive balance between classification accuracy and computational efficiency for rice leaf disease identification.

Preliminary Study of ResNet-Based Facial Identification for Access Control Systems

2026-06-14T23:54:40+00:00

Facial recognition is an important application of artificial intelligence (AI) and computer vision in modern security systems, particularly for automated access control. This study aimed to implement and evaluate a Residual Network (ResNet)-based facial identification system for door access control applications. A publicly available facial image dataset was used and was divided into training (70%), validation (20%), and testing (10%) subsets. The proposed methodology consisted of data preprocessing, ResNet-based model training, and performance evaluation using accuracy and loss metrics. The model was trained for 10 epochs to assess its initial learning capability. The experimental results showed relatively low performance, with training accuracy ranging from 3.6% to 3.8% and validation accuracy of approximately 3.6%, while loss values remained high throughout the training process. These findings indicated that the model was unable to effectively learn discriminative facial features from the dataset and exhibited signs of underfitting. The limited performance was likely associated with insufficient dataset diversity, suboptimal preprocessing procedures, and non-optimized training parameters. The study highlighted the challenges of implementing ResNet-based facial recognition systems under constrained training conditions. Future work should focus on expanding the dataset, applying data augmentation techniques, optimizing hyperparameters, and utilizing pretrained models to improve recognition performance and system reliability.

Recency, Frequency, and Monetary-Based Customer Segmentation Using K-Means for Analysing Transactional Behaviour in a Service-Based Micro, Small, and Medium Enterprises

2026-03-14T05:29:20+00:00

Micro, Small, and Medium Enterprises (MSMEs) often faced challenges in designing effective promotional initiatives due to the limited use of systematic customer behavior analysis. This study examined the application of (Recency, Frequency, Monetary) RFM analysis combined with K-Means clustering to explore customer segmentation in a service-based MSME context. Transaction data from a local laundry service operating in Palu, Indonesia, consisting of 2,220 digital transaction records collected between 2022 and 2025, were processed and transformed into RFM variables using min–max normalization. The optimal number of clusters was determined using the Elbow method, resulting in four customer segments. Cluster quality was evaluated using internal validation metrics, yielding a Davies–Bouldin Index (DBI) of 0.61 and a Sum of Squared Errors (SSE) value of 1.73, indicating reasonably compact and well-separated clusters. The resulting segments exhibited distinct transactional profiles across recency, transaction frequency, and monetary contribution, reflecting heterogeneity in customer engagement within the studied MSME. Rather than prescribing specific marketing actions, the findings provided an interpretable analytical basis for considering differentiated promotional strategies aligned with observed customer behavior patterns. Overall, this study demonstrated that RFM-based segmentation offered a feasible and data-driven approach to supporting evidence-informed promotional planning in service-oriented MSMEs operating under data and resource constraints.

Application of Transfer Learning Using Inception-Resnet-V2 for Image-Based Classification of Apple Leaf Diseases

2026-05-31T01:16:52+00:00

Apple leaf diseases posed a major challenge to agricultural productivity due to their similar visual appearance and the limitations of manual classification methods. This study aimed to develop an accurate and efficient image-based classification system for apple leaf diseases using the Inception-ResNet-V2 architecture and a transfer learning approach. The dataset consisted of 3,171 images from the PlantVillage dataset, categorized into four classes: Apple Scab, Cedar Apple Rust, Black Rot, and Healthy. The data were divided into training, validation, and test sets in a 70:15:15 ratio using stratified sampling. Image preprocessing included resizing, normalization, and data augmentation, while class balancing was applied to address class imbalance. The model was trained using the Adam optimizer through a two-stage process consisting of feature extraction and refinement. Experimental results showed that the proposed model achieved a test accuracy of 98.74%, with high precision, recall, and F1-scores across all classes, demonstrating strong classification performance and generalization ability. This study demonstrated that Inception-ResNet-V2 was effective in capturing complex visual features of apple leaf diseases. In conclusion, the proposed approach offered an effective solution for classifying apple leaf diseases and had the potential to support more efficient and accurate agricultural decision-making.

Potato Leaf Disease Classification Using MobileNetV3 Architecture With Adam and Stochastic Gradient Descent Optimizers

2026-05-03T05:09:47+00:00

Potato leaf diseases such as Early Blight and Late Blight reduced productivity and could cause crop failure if they were not detected early. This study analyzed the comparative performance of the Adam and Stochastic Gradient Descent (SGD) optimizers using the MobileNetV3-Large architecture for potato leaf disease classification. The dataset consisted of three categories: healthy leaves, Early Blight, and Late Blight, with a total of 4,072 images. All images were processed through preprocessing stages, including resizing to 224 × 224 pixels and pixel value normalization. The data were divided into training, validation, and testing sets with a ratio of 70:20:10. Random undersampling and data augmentation techniques were applied to the training data to address class imbalance and improve the model’s generalization capability. The model training process was conducted using a transfer learning approach with the MobileNetV3-Large architecture through two stages: feature extraction and fine-tuning. Model performance evaluation was based on accuracy, precision, recall, and F1-score metrics. The results showed that the Adam optimizer achieved a test accuracy of 98.75% with an F1-score of 0.9875, while the SGD optimizer achieved a test accuracy of 96.56% with an F1-score of 0.9635. The Adam optimizer also demonstrated faster and more stable convergence during the training process. This study was expected to serve as a reference for determining an appropriate optimizer for deep learning applications in image classification, particularly in plant disease detection.

An Image Processing-Based Fire Detection System Using Orange Pi 4A with Internet of Things Integration in Indoor Environments

2026-03-05T04:26:34+00:00

Fire hazards in indoor industrial environments require fast and reliable detection systems, as conventional sensor-based methods often suffer from delayed responses and high false-alarm rates. This study proposes a low-cost, Internet of Things-integrated visual fire detection system based on the YOLOv11 deep learning model implemented on an Orange Pi 4A. The system integrates an IP camera for visual acquisition, real-time detection, and automatic data logging through a MySQL-based monitoring platform. Experiments were conducted in a 3 × 3 m indoor environment using candle, stove, and burning fires at various camera distances. System performance was evaluated using confidence score, bounding box pixel area, and recall based on True Positive and False Negative classifications. Candle flames were reliably detected up to 100 cm with recall values of 90.24%–100% and pixel areas below 5,000 px, while stove flames achieved recall above 93% at 50–100 cm with pixel areas of 11,144–42,525 px. Burning fires maintained high performance up to 300 cm, reaching confidence values above 0.70 and recall rates of 78.94%–100% with pixel areas exceeding 44,000 px. The results indicate that detection reliability is primarily influenced by apparent flame size rather than camera distance. Overall, the proposed system demonstrates strong feasibility as an embedded, IoT-integrated fire detection solution for early warning in indoor industrial environments, although limitations remain in detecting small flames under low-resolution and low-light conditions.

Implementation of a Centralized Cloud-Based Hotspot Voucher Management System and Network Traffic Monitoring Using MikroTik Virtual

2026-06-12T02:55:31+00:00

Hotspot voucher services became one of the most widely adopted solutions among local Internet Service Providers (ISPs) because they facilitated user access control, usage duration management, and bandwidth allocation. In the existing network environment, hotspot voucher sales were operated at two different locations. However, the management of these locations was still performed independently using separate MikroTik routers. This condition created several challenges, including decentralized voucher data, standalone user authentication processes, inefficient voucher administration, and the inability to perform comprehensive network traffic monitoring. This study implemented a centralized hotspot voucher management system using a MikroTik Cloud Hosted Router (CHR) to integrate two hotspot voucher sales locations into a unified management platform. The MikroTik CHR was deployed in a cloud computing environment and functioned as the central server for voucher management, user authentication, and network traffic monitoring. Each MikroTik router at the voucher sales locations was connected to the MikroTik CHR through a VPN tunnel. Furthermore, a web-based application was developed as a management interface and was integrated with the MikroTik CHR through an Application Programming Interface (API). The application enabled administrators to generate vouchers, monitor voucher status, manage users, and observe network traffic through a centralized dashboard. The research adopted the Network Development Life Cycle (NDLC) methodology, which consisted of the stages of analysis, design, simulation, implementation, monitoring, and management. The implementation results demonstrated that the centralized system successfully integrated voucher management across two different locations, simplified administrative processes, supported centralized user authentication, and provided unified network traffic monitoring through a single platform. Therefore, the implementation of MikroTik CHR in a centralized hotspot voucher system improved management efficiency and supported the expansion and sustainability of hotspot services across multiple locations.

Integrating TOGAF 10 and ISO 20000-1:2018 for Digital Multi-finance Service Level Agreement/Mean Time to Repair improvements

2026-03-05T04:26:35+00:00

Digital transformation in the multi-finance sector demands service architectures that are flexible, reliable, and scalable; however, misalignment between architectural design and operational execution often leads to weak service performance. This study proposes an integrated framework that combines TOGAF 10 artifacts with ISO/IEC 20000-1:2018 processes to systematically estimate Service Level Agreement (SLA) targets and reduce Mean Time to Repair (MTTR). Using a Design Science Research approach, the framework was implemented in a 14-month case study at PT XYZ Multi-finance. The resulting artifacts include a bidirectional traceability model linking business objectives to SLA and MTTR indicators, as well as an operability pattern catalog to support “design for operability.” The implementation delivered measurable operational improvements: MTTR decreased from a peak of 775 minutes to below 60 minutes, Mean Time to Detect (MTTD) was reduced by approximately 90%, SLA compliance increased to 99.7%, and incidents caused by manual configuration errors declined. These results demonstrate that integrating enterprise architecture design with service management processes can significantly improve service reliability and overall operational performance.