Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Detection of Sugarcane Leaf Disease Using Pre-Trained Feature Extraction and SVM Method
No full textSugarcane (Saccharum officinarum) is an important commodity in the sugar industry, but it is vulnerable to leaf diseases such as Red Rot, Rust, Yellow Leaf, and Mosaic, which can significantly reduce the quality and quantity of yields. Manual identification is time-consuming and prone to subjective errors, therefore an automatic detection method based on digital images is required. This study proposes a combination of VGG16 pre-trained as a feature extractor with Support Vector Machine (SVM) as a classifier. The dataset used is the Sugarcane Leaf Disease Dataset from Kaggle, consisting of 2,521 images of five classes, which were then balanced through augmentation in the form of rotation, zoom, and flipping to a total of 3,000 images (600 per class). The preprocessing stage includes resizing the images to 224×224 pixels and normalization using the preprocess_input function. Three model scenarios were tested, namely SVM, VGG16, and VGG16+SVM. Evaluation was carried out using two methods, namely an 80:20 train–test split and 10-fold cross-validation, with metrics of accuracy, precision, recall, F1-score, G-Mean, and AUC. The experimental results show that VGG16+SVM provides the best performance with an accuracy of 99.60% on the 80:20 scheme, while on 10-fold cross-validation the average accuracy is 80.76%. This value surpasses the baseline SVM and VGG16 + Softmax, proving that the integration of VGG16 feature extraction with SVM classification can produce stable and accurate performance. This research contributes to the development of image-based plant disease detection systems to support precision agriculture and fast decision-making.Sugarcane (Saccharum officinarum) is an important commodity in the sugar industry, but it is vulnerable to leaf diseases such as Red Rot, Rust, Yellow Leaf, and Mosaic, which can significantly reduce the quality and quantity of yields. Manual identification is time-consuming and prone to subjective errors, therefore an automatic detection method based on digital images is required. This study proposes a combination of VGG16 pre-trained as a feature extractor with Support Vector Machine (SVM) as a classifier. The dataset used is the Sugarcane Leaf Disease Dataset from Kaggle, consisting of 2,521 images of five classes, which were then balanced through augmentation in the form of rotation, zoom, and flipping to a total of 3,000 images (600 per class). The preprocessing stage includes resizing the images to 224×224 pixels and normalization using the preprocess_input function. Three model scenarios were tested, namely SVM, VGG16, and VGG16+SVM. Evaluation was carried out using two methods, namely an 80:20 train–test split and 10-fold cross-validation, with metrics of accuracy, precision, recall, F1-score, G-Mean, and AUC. The experimental results show that VGG16+SVM provides the best performance with an accuracy of 99.60% on the 80:20 scheme, while on 10-fold cross-validation the average accuracy is 80.76%. This value surpasses the baseline SVM and VGG16 + Softmax, proving that the integration of VGG16 feature extraction with SVM classification can produce stable and accurate performance. This research contributes to the development of image-based plant disease detection systems to support precision agriculture and fast decision-making
Comparing Different KNN Parameters Based on Woman Risk Factors to Predict the Cervical Cancer
No full textCervical cancer remains a major cause of mortality among women, particularly in low-resource regions where access to conventional screening is limited. Early detection through predictive modeling offers a low-cost and non-invasive alternative to clinical diagnostics. This study aims to evaluate the effectiveness of the k-Nearest Neighbors algorithm for predicting cervical cancer risk using behavioral and psychosocial attributes. The research utilized the publicly available Sobar cervical cancer behavioral dataset comprising 72 instances with 18 input features and a binary target label. Data preprocessing included removal of incomplete records, encoding of categorical variables, and normalization. The algorithm was tested across varying numbers of neighbors and distance metrics, with performance evaluated using 10-fold cross-validation and multiple classification metrics. The optimal configuration was achieved with three neighbors and the Manhattan distance metric, yielding an accuracy of 93.06%, sensitivity of 93.10%, specificity of 85.90%, precision of 93.10%, F1-score of 92.90%, and an area under the curve of 0.8952. This performance surpassed the reported baseline of a probabilistic classifier and demonstrated the algorithm’s capability to capture complex behavioral patterns associated with cervical cancer risk. These findings confirm the feasibility of applying optimized instance-based learning to behavioral data for early cancer risk assessment. The approach offers potential for integration into community health programs to support early detection and prevention strategies
Pengujian YOLOv8 dan Centroid Tracking pada Sistem Deteksi, Klasifikasi, dan Penghitungan Jumlah Kendaraan
No full textAn automatic vehicle detection and counting system is essential for Intelligent Transportation Systems (ITS) to monitor and manage traffic effectively. This study evaluates the performance of the lightweight YOLOv8n (nano) model for vehicle detection and classification, combined with a Centroid Tracking algorithm to improve vehicle counting accuracy. YOLOv8n was selected for its balance between computational efficiency and detection accuracy, making it suitable for devices with limited resources. The research involved collecting a dataset of seven vehicle classes (bus_l, bus_s, car, truck_l, truck_m, truck_s, truck_xl), followed by data preprocessing and training the YOLOv8n model for 40 epochs. Data augmentation techniques were applied to enhance data variability and improve model robustness. The Centroid Tracking algorithm was integrated to maintain vehicle identity across frames and prevent double counting. Model evaluation used precision, recall, F1-score, and mean Average Precision (mAP). Results show YOLOv8n achieved an overall [email protected] of 0.820. The “car” class attained the highest mAP of 0.963, while “truck_s” had the lowest at 0.665, mainly due to imbalanced data distribution. The Centroid Tracking effectively maintained object identities and provided consistent vehicle counts during testing. This combination offers a reliable and efficient system suitable for real-time traffic monitoring, parking management, and enhancing road safety. The YOLOv8n and Centroid Tracking-based system demonstrates strong potential for practical ITS applications, especially on devices with limited computational resources. Future work should focus on expanding the dataset and improving class balance to further enhance detection accuracy and system robustness.Sistem deteksi dan penghitungan kendaraan otomatis merupakan komponen penting dalam penggunaan Intelligent Transportation System (ITS) untuk keperluan pemantauan dan manajemen lalu lintas. Terdapat berbagai model yang dapat digunakan dalam sistem ini, salah satunya adalah YOLO (You Only Look Once). Pada penelitian ini, kami memfokuskan pada pengujian performa model YOLOv8, khususnya varian ringan yaitu YOLOv8n (nano), dalam mendeteksi dan mengklasifikasikan kendaraan melalui beberapa skenario pengujian. YOLOv8n dipilih karena efisien secara komputasi dan dapat berjalan pada perangkat dengan keterbatasan sumber daya, namun tetap mempertahankan akurasi deteksi yang kompetitif. Selain itu, kami juga mengombinasikan algoritma Centroid Tracking guna meningkatkan kapabilitas sistem dalam melakukan pelacakan dan penghitungan kendaraan secara otomatis. Metode yang digunakan dalam penelitian ini meliputi pengumpulan dataset kendaraan dengan 7 kelas (bus_l, bus_s, car, truck_l, truck_m, truck_s, truck_xl), pengolahan data, pelatihan model YOLOv8n, penggunaan sistem pelacakan dengan algoritma centroid, serta evaluasi model. Model dilatih sebanyak 40 epoch dengan menerapkan teknik augmentasi untuk menambahkan variasi data dan meningkatkan hasil model. Evaluasi dilakukan menggunakan metrik precision, recall, F1-score, dan mean Average Precision (mAP). Hasil penelitian menunjukkan bahwa model YOLOv8n mampu mencapai performa baik dengan [email protected] sebesar 0.820 untuk seluruh kelas kendaraan. Kelas \u27car\u27 menunjukkan performa terbaik dengan mAP sebesar 0.963, sementara kelas \u27truck_s\u27 mencatatkan performa terendah dengan mAP sebesar 0.665. Sistem Centroid Tracking juga berhasil melakukan pelacakan objek secara akurat dan menghasilkan penghitungan kendaraan yang konsisten. Pengujian sistem ini menunjukkan potensi besar untuk diaplikasikan dalam pengembangan sistem pemantauan lalu lintas real-time, manajemen parkir, serta sistem keamanan jalan raya yang lebih cerdas dan responsif
Deep Learning-Based Detection of Online Gambling Promotion Spam in Indonesian YouTube Comments
No full textOnline gambling promotion has increasingly penetrated social media platforms, with YouTube comments becoming a frequent target for spam-based advertising. Such activities not only violate platform policies but also expose users to harmful content. Addressing this issue requires automated detection systems capable of handling noisy, informal, and highly imbalanced text data. This study investigates the effectiveness of four recurrent neural architectures LSTM, GRU, BiLSTM, and BiGRU for detecting gambling promotion comments in Indonesian YouTube data. To address class imbalance, multiple experimental scenarios were explored, including the original distribution, undersampling, oversampling, and class weighting. Model performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and confusion matrix analysis. The results show that bidirectional models outperformed their unidirectional counterparts, with BiGRU achieving the best overall performance. When combined with class weighting, BiGRU reached 98% accuracy, 0.83 F1-score, and 0.971 ROC-AUC, demonstrating a superior ability to detect minority-class instances. Oversampling improved recall substantially but increased false positives, while undersampling reduced accuracy; class weighting provided the most balanced performance across metrics. These findings confirm that BiGRU with class weighting offers the most practical balance between accuracy, recall, and computational efficiency, making it well-suited for real-time moderation systems. The study provides a strong foundation for future research on transformer-based architectures and cross-platform spam detection in Indonesian social media environments
Comparative Performance of SVM and BERT-Base Using Hybrid Preprocessing for Fast Fashion Sentiment Analysis
No full textFast fashion poses major environmental and social challenges, yet public awareness in Indonesia remains insufficiently understood. This study compares Support Vector Machine and BERT-Base for sentiment analysis of 3,513 TikTok comments on fast fashion sustainability using a hybrid preprocessing pipeline that incorporates a 404-entry slang dictionary and IndoNLP utilities to address informal language, code-mixing, and character elongation. Sentiment labels generated using VADER were validated against 1,747 manually annotated samples, achieving Cohen\u27s Kappa of 0.7155, indicating substantial agreement. BERT-Base achieves 92.7% accuracy with F1-scores of 0.86, 0.94, and 0.93 for negative, neutral, and positive classes, while SVM attains competitive 90.4% accuracy with F1-scores of 0.84, 0.93, and 0.91. BERT demonstrates superior negative sentiment detection with recall of 0.87 compared to SVM at 0.82, critical for identifying sustainability concerns. Computational analysis reveals significant trade-offs as BERT requires 230.2 seconds of GPU training and 3.449 seconds of inference, whereas SVM operates efficiently on CPU with 25.9 seconds of training and 0.051 seconds of inference, representing 8.9× and 67.6× efficiency advantages. The sentiment distribution comprising 46.9% neutral, 34.5% positive, and 18.6% negative comments indicates limited critical awareness among Indonesian users. These findings demonstrate that systematic preprocessing bridges the performance gap between classical and transformer models while enabling deployment decisions based on resource constraints, providing methodological insights for low-resource informal text analysis and practical guidance for scalable social listening, greenwashing detection, and evidence-based sustainability communication strategies
Comparative No-Reference Evaluation of Classical Image Sharpening Techniques under Varying Degradation Conditions
No full textThis research conducts a comparative evaluation of four image sharpening methods: Unsharp Masking, Laplacian of Gaussian, High-Boost Filtering, and Adaptive High-Boost Filtering. These methods are tested on low-contrast, blurred, normal, and high-contrast images. The assessment uses No Reference Image Quality Assessment metrics, specifically BRISQUE and NIQE, along with intensity histogram analysis and visual inspection. Results show that High-Boost Filtering improves global contrast, reducing BRISQUE scores to 26.28 for low-contrast images and 27.56 for high-contrast images, although it can cause halo artifacts. Unsharp Masking performs best on blurred images, lowering BRISQUE to 26.65, but it is more sensitive to noise. The Laplacian of Gaussian yields relatively low NIQE scores, such as 3.04 in low-contrast and 3.10 in high-contrast images; however, its output often appears coarse in texture. Adaptive High-Boost Filtering performs best on normal images, achieving a BRISQUE score of 11.89, but shows limited improvement in other cases. Notably, alignment between NIQE scores and perceptual evaluation is only observed in high-contrast images. These results confirm that no single technique is universally optimal, emphasizing the importance of selecting sharpening methods based on specific image degradation characteristics. Additionally, this observation highlights that BRISQUE more reliably reflects perceived image quality, whereas NIQE occasionally diverges from subjective judgments.This research conducts a comparative evaluation of four image sharpening methods: Unsharp Masking, Laplacian of Gaussian, High-Boost Filtering, and Adaptive High-Boost Filtering. These methods are tested on low-contrast, blurred, normal, and high-contrast images. The assessment uses No Reference Image Quality Assessment metrics, specifically BRISQUE and NIQE, along with intensity histogram analysis and visual inspection. Results show that High-Boost Filtering improves global contrast, reducing BRISQUE scores to 26.28 for low-contrast images and 27.56 for high-contrast images, although it can cause halo artifacts. Unsharp Masking performs best on blurred images, lowering BRISQUE to 26.65, but it is more sensitive to noise. The Laplacian of Gaussian yields relatively low NIQE scores, such as 3.04 in low-contrast and 3.10 in high-contrast images; however, its output often appears coarse in texture. Adaptive High-Boost Filtering performs best on normal images, achieving a BRISQUE score of 11.89, but shows limited improvement in other cases. Notably, alignment between NIQE scores and perceptual evaluation is only observed in high-contrast images. These results confirm that no single technique is universally optimal, emphasizing the importance of selecting sharpening methods based on specific image degradation characteristics. Additionally, this observation highlights that BRISQUE more reliably reflects perceived image quality, whereas NIQE occasionally diverges from subjective judgments
Two-Stage Maritime Anomaly Detection: Unsupervised Outlier Filtering and Optimized Bidirectional LSTM on Southeast Asian AIS Data
No full textThis paper presents a two-stage framework for detecting anomalous vessel trajectories in Automatic Identification System (AIS) data from Southeast Asian waters, addressing challenges of high traffic density, diverse vessel behaviors, and severe class imbalance. The primary objective is to minimize missed threats while maintaining manageable false alarm rates in security-critical maritime surveillance systems. The research employs a hybrid approach combining unsupervised and supervised learning methods. In the first stage, DBSCAN and Isolation Forest algorithms filter noise and generate high-confidence outlier labels from 15,542 real-world vessel trajectories. Comparative analysis demonstrates substantial agreement between methods with Cohen\u27s Kappa of 0.688 and 55.3% anomaly overlap, indicating complementary detection capabilities that enhance filtering robustness. In the second stage, a Bidirectional Long Short-Term Memory model is optimized through systematic hyperparameter tuning across 48 configurations, covering sequence length, network architecture, dropout rate, learning rate, and sampling strategies. Comprehensive baseline evaluation validates BiLSTM\u27s suitability for security applications, achieving 15.41% F1-score improvement over unidirectional LSTM and 33% fewer false negatives compared to Bidirectional GRU alternative. The optimized BiLSTM attains F1-score of 0.5709 with precision 0.5444 and recall 0.6000, exhibiting 90.03% specificity for normal vessels and 76.17% sensitivity for anomalies. The model misses only 23.8% of threats while maintaining 9.97% false alarm rate, providing balanced performance suitable for human-verified security-critical maritime surveillance in Southeast Asian waters
Optimization of Early Diagnosis Prediction Models for Acute Respiratory Infections (ARI) in Children Using Decision Tree, Random Forest, and Resampling Techniques
No full textAcute Respiratory Tract Infections (ARI) are the leading cause of childhood morbidity in Indonesia, with challenges in early detection due to limited medical personnel and diagnostic data imbalance, where LRTI cases are far fewer than URTI cases. This study developed and optimized an ARI classification prediction model (URTI and LRTI) based on machine learning with resampling techniques to address imbalance. An explanatory quantitative design was used with secondary data from the Mijen Community Health Center, Semarang (2020–2025, 12.177 valid data), with preprocessing including outlier handling (Winsorizing, IQR), stratified split (70:30), and RobustScaler on the training data. Three resampling techniques (SMOTE, ADASYN, SMOTE-ENN) were applied, then tested using Decision Tree and Random Forest with GridSearchCV and 5-fold cross-validation, focusing on Recall and AUC-PR evaluation for minority classes. The results showed that Random Forest with SMOTE-ENN provided the best performance, increasing the LRTI recall from 0.02 to 0.37 and F1-macro to 0.54, while Decision Tree with SMOTE-ENN produced the highest AUC-PR of 0.31. Despite this significant improvement, a recall of 0.37 is still low for clinical applications because the risk of false negatives remains high, potentially delaying patient treatment Future implementation requires the integration of clinical symptom data (e.g., respiratory rate) to achieve clinically acceptable sensitivity. These findings confirm that resampling can improve model capabilities, but additional feature exploration is needed to achieve adequate diagnostic sensitivity in the context of healthcare analytics
Bigger Matters? Exploring the Mediating Effect of Firm Size on ESG Disclosure and Firm Value
No full textThis research investigates the effect of Environmental, Social, and Governance (ESG) disclosure on firm value with firm size as a mediating variable. The study is motivated by the growing importance of ESG initiatives in attracting investors, strengthening stakeholder trust, and promoting sustainable growth, particularly in developing markets such as Indonesia. The research adopts a quantitative approach using secondary data from manufacturing companies listed on the Indonesia Stock Exchange (IDX) during the 2021–2024 period. ESG disclosure was measured through content analysis of sustainability reports, firm value was proxied by Tobin’s Q, and firm size was assessed using the natural logarithm of total assets. Data analysis employed Partial Least Squares Structural Equation Modeling (PLS-SEM) to evaluate both direct and mediated relationships among variables. The results indicate that ESG disclosure significantly enhances firm value, supporting the argument that transparent sustainability practices reduce information asymmetry and strengthen market confidence. Furthermore, ESG disclosure was found to have a positive and significant effect on firm size, suggesting that firms with robust ESG practices tend to expand more rapidly and attract greater resources. Firm size also positively influences firm value, confirming its role as a strategic advantage. Mediation analysis reveals that firm size partially mediates the relationship between ESG disclosure and firm value, indicating that ESG disclosure affects valuation both directly and indirectly through organizational growth. In conclusion, the findings underscore the dual role of ESG disclosure as a direct driver of firm value and an indirect enhancer through firm size. For managers, this emphasizes the strategic importance of ESG initiatives not only as compliance measures but as growth enablers and value-creation mechanisms. The study provides important insights for regulators, investors, and corporate leaders in Indonesia, highlighting the need to strengthen ESG reporting standards and encourage firms to integrate sustainability into their core strategies to enhance competitiveness and long-term value
The Impact of External Auditor Forensic Accounting Competencies on Financial Performance of Listed Companies on the Indonesia Stock Exchange
No full textAs business complexity and financial fraud risks grow, external auditors play an increasingly vital role in maintaining the integrity of financial statements. Competencies such as forensic accounting are believed to enhance audit quality by strengthening fraud detection and prevention. Therefore, this study examines the impact of External Auditor Forensic Accounting Competencies (EAFAC) on the financial performance of companies listed on the Indonesian Stock Exchange in 2024. Using a quantitative approach, the research analyzes data from 282 companies selected through simple random sampling. Company size, leverage, and company age are included as control variables, and the data are analyzed using multiple linear regression in SPSS. The results show that, both partially and simultaneously, all variables have a significant effect on financial performance. Specifically, EAFAC and company size have a positive effect, while leverage and company age have a negative effect. Overall, these findings confirm that implementing forensic accounting competencies within external audit practices is a strategic investment to improve company financial performance, especially through more effective risk mitigation and fraud prevention. However, the research also identified challenges in implementing these competencies, especially within the Big Four public accounting firms, indicating room for further improvement and standardization in Indonesia’s audit profession