Jurnal Politeknik Negeri Batam (PoliBatam)
Not a member yet
3001 research outputs found
Sort by
Implementation of ResNet-50-Based Convolutional Neural Network For Mobile Skin Cancer Classification
Get PDFThe skin is one of the most important parts of the human body, serving vital functions such as protecting internal organs from injury, shielding against direct bacterial exposure, regulating body temperature, and more. However, the skin is also susceptible to diseases, one of which is skin cancer. Skin cancer can be extremely dangerous if not treated promptly, as it can lead to death. Therefore, early detection is crucial. This study proposes a technology-based solution by classifying skin cancer using a convolutional neural network (CNN) with a ResNet50 architecture implemented into a mobile application via a REST API using Flask. The HAM10000 dataset, consisting of 10,015 skin lesion images across seven classes, was used for model training. Various testing scenarios were conducted to determine the optimal parameter combination. The best results were achieved with an accuracy of 83.84%, precision and recall of 83%, and an F1-score of 83%, using a training data configuration of 70%, dropout of 0.4, and a batch size of 64. The model implemented in this Android application can perform early detection of skin cancer quickly, practically, and easily accessible to the general public, though healthcare professionals must still supervise it. However, although this model can assist users in making early predictions, the prediction results from this model are only a tool for early detection and do not replace clinical diagnosis by professional medical personnel.2) Figure 8 shows the display for taking pictures through the gallery or camera. Users can choose the image they want to upload from the gallery or the camera to be analysed and predicted by the model
Implementation of Text Mining for Evaluating the Relevance Between News Headlines and Content on a Web-Based Platform
Get PDFTechnological advancements in the era of the Industrial Revolution 4.0 have significantly transformed how society accesses and consumes information, particularly through online news portals. This study aims to analyze the relevance between news headlines and article content on Indonesian online news platforms by employing text mining techniques and similarity checking methods. To enhance the accuracy of relevance assessment, this research utilizes two deep learning-based modeling algorithms: Long Short-Term Memory (LSTM) and IndoBERT. The data was collected from three leading Indonesian news portals detik.com, kompas.com, and suara.com with a total of 52,242 articles from the entertainment and national news categories, gathered between July 1 and September 30, 2024. The dataset includes attributes such as headline, category, publication date, author, article URL, and news content. The research process consists of several stages, including data collection through web scraping, data pre-processing (which involves cleaning the category, author, and content columns), content summarization, text similarity calculation, and data labeling into three classes (relevan, berlebihan, and nonrelevan). Evaluation results show that the IndoBERT model outperforms LSTM, achieving the best performance with a training accuracy of 0.9048 and a training loss of 0.2514, as well as a validation accuracy of 0.8604 and a validation loss of 0.4039. These findings demonstrate that IndoBERT is effective in assessing the coherence between news headlines and content in today’s digital age
Spatial Distribution Patterns of Urban Growth, Dasymetric Mapping, and Air Pollution Standard Index (ISPU) in Bogor Regency (2018-2023)
Get PDFBogor Regency has undergone rapid industrialization and urban growth, raising concerns about deteriorating air quality and its health impacts. This study aims to analyze the spatial distribution of urban expansion and its relationship with air pollution levels between 2018 and 2023. Using Sentinel-5P and Sentinel-2A satellite imagery processed via Google Earth Engine (GEE), the research maps changes in land use and monitors key pollutants such as nitrogen dioxide (NO₂) and sulfur dioxide (SO₂). The Air Pollution Standard Index (ISPU) is used to classify air quality conditions. The results indicate a notable increase (16%) in urban areas, with population density rising from 178 to 180 inhabitants per hectare, and a shift toward higher pollution categories, with significant portions of the region now classified as "Unhealthy" increased from 17% in 2018 to 31% in 2023. The urban growth model demonstrated high accuracy (92.5%) and strong alignment with local monitoring data. Model evaluation showed a Mean Absolute Error (MAE) of 1.295 µg/m³ and Root Mean Square Error (RMSE) of 1.478 µg/m³, demonstrating strong agreement between predicted and observed data from the Bogor Regency Environmental Agency. These findings highlight the need for integrated urban planning and effective air quality management to reduce future health risks in Bogor Regency
Application of the PSI-VIKOR Method in Determining Priorities for Poor Areas Based on Poverty Indicators in Central Java
Get PDFPoverty remains a significant challenge in developing countries, including Indonesia. Although the national poverty rate has declined, Central Java still shows relatively high rates. This study aims to identify priority areas in Central Java requiring government intervention to support effective poverty alleviation planning. Data were sourced from the Central Statistics Agency (BPS) of Central Java Province in 2023. A Decision Support System (DSS) approach was applied using the integrated Preference Selection Index (PSI) and VIKOR methods. PSI was used to determine objective criteria weights based on preference variations, while VIKOR ranked regions based on compromise solutions closest to ideal conditions.The ranking results were visualized spatially through a digitization process using QGIS to produce thematic maps. Analysis showed that Purworejo, Wonogiri, and Batang are high-priority regencies, whereas Semarang City, Banyumas, and Kendal have relatively stable socio-economic conditions. Validation using the Normalized Discounted Cumulative Gain (NDCG) method yielded a score of 0.9268, indicating strong alignment with historical data. These findings confirm the effectiveness of the PSI–VIKOR approach in supporting data-driven poverty alleviation strategies. The novelty of this study lies in the integrated application of PSI–VIKOR for spatial poverty prioritization, which has not previously been implemented in the Indonesian context.Poverty remains a significant challenge in developing countries, including Indonesia. Although the national poverty rate has declined, Central Java still shows relatively high rates. This study aims to identify priority areas in Central Java requiring government intervention to support effective poverty alleviation planning. Data were sourced from the Central Statistics Agency (BPS) of Central Java Province in 2023. A Decision Support System (DSS) approach was applied using the integrated Preference Selection Index (PSI) and VIKOR methods. PSI was used to determine objective criteria weights based on preference variations, while VIKOR ranked regions based on compromise solutions closest to ideal conditions.The ranking results were visualized spatially through a digitization process using QGIS to produce thematic maps. Analysis showed that Purworejo, Wonogiri, and Batang are high-priority regencies, whereas Semarang City, Banyumas, and Kendal have relatively stable socio-economic conditions. Validation using the Normalized Discounted Cumulative Gain (NDCG) method yielded a score of 0.9268, indicating strong alignment with historical data. These findings confirm the effectiveness of the PSI–VIKOR approach in supporting data-driven poverty alleviation strategies. The novelty of this study lies in the integrated application of PSI–VIKOR for spatial poverty prioritization, which has not previously been implemented in the Indonesian context
SMOTE and Weighted Random Forest for Classification of Areas Based on Health Problems in Java
Get PDFRandom Forest (RF) is a popular Machine Learning (ML) approach extensively employed for addressing classification issues. Nevertheless, the RF method for classification problems demonstrates suboptimal performance in cases of data imbalance. There are several approaches to enhance RF performance when coping with data imbalance issues, such as using weighting and oversampling. This research explores the intervention of RF in addressing data imbalances, focusing on case studies of health problem classification in Java This study aims to develop models to analyze the health status of regions using RF, WRF, SMOTE-RF, and SMOTE-WRF methods. The objective is to compare the performance of these models and identify the best model for classifying DBK and Non-DBK categories in Java. The research results show that SMOTE-WRF is the most effective model in classifying DBK, achieving an accuracy level of 93.62%, sensitivity of 85.71%, precision of 75%, F-score of 80%, and AUC of 93.57%. The three key variables in the SMOTE-WRF model entail access to adequate sanitation, egg and milk consumption, and the number of doctor
Passenger Density Prediction at the Airport Using LSTM and SARIMA: A Case Study at Radin Inten Airport, Lampung
Get PDFPassenger density prediction at airports is a critical aspect of operational planning and strategic decision-making. This study aims to develop a passenger count prediction model for Radin Inten Airport in Lampung using a combination of Seasonal AutoRegressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM), and to compare it with Random Forest and XGBoost models. The dataset consists of daily passenger counts from January 2023 to December 2024. The research includes data exploration, preprocessing, separate modeling with SARIMA and LSTM, and their integration through a residual learning approach. Evaluation results show that SARIMA achieved the best performance in capturing seasonal patterns with a Mean Absolute Percentage Error (MAPE) of 3.81%, followed by Random Forest with 5.81% and XGBoost with 5.84%. The LSTM model performed less effectively with a MAPE of 6.81%. Although the SARIMA–LSTM combination is theoretically promising, it produced a worse result with a MAPE of 14.27% due to error accumulation in the residual learning stage. This study highlights that the choice of prediction model strongly depends on data characteristics and forecasting objectives, as well as the importance of multi-model integration to improve prediction accuracy in airport passenger density forecasting applications.Passenger density prediction at airports is a critical aspect of operational planning and strategic decision-making. This study aims to develop a passenger count prediction model for Radin Inten Airport in Lampung using a combination of Seasonal AutoRegressive Integrated Moving Average (SARIMA) and Long Short-Term Memory (LSTM), and to compare it with Random Forest and XGBoost models. The dataset consists of daily passenger counts from January 2023 to December 2024. The research includes data exploration, preprocessing, separate modeling with SARIMA and LSTM, and their integration through a residual learning approach. Evaluation results show that SARIMA achieved the best performance in capturing seasonal patterns with a Mean Absolute Percentage Error (MAPE) of 3.81%, followed by Random Forest with 5.81% and XGBoost with 5.84%. The LSTM model performed less effectively with a MAPE of 6.81%. Although the SARIMA–LSTM combination is theoretically promising, it produced a worse result with a MAPE of 14.27% due to error accumulation in the residual learning stage. This study highlights that the choice of prediction model strongly depends on data characteristics and forecasting objectives, as well as the importance of multi-model integration to improve prediction accuracy in airport passenger density forecasting applications
Enhancing The Security of E-Invoicing for Distribution Companies Through Image-Based PDF Conversion and QR Verification
Get PDFDistribution companies face significant challenges in securing electronic invoices, as PDF files are susceptible to unauthorized text extraction and manipulation. Prior solutions include SHA-256 digital signatures, and QR code-based verification. There are often require specialized tools, stable internet, or user intervention posing barriers for general trade customers with limited digital access. To address these limitations, this study proposes a hybrid e-invoicing method by converting invoices into image-based PDFs embedded with QR codes. This approach enhances document security, increases resistance to text manipulation, and ensures file sizes remain under 1 MB for smooth distribution via WhatsApp. A dataset of 1000 invoices was tested using OCR and FuzzyWuzzy string similarity to compare extractability between text-based and image-based formats. A composite score was calculated by combining file size and manipulation resistance metrics. Results show that image-based PDFs achieve a significantly higher score (0.595) compared to text-based PDFs (0.005), confirming their superiority in terms of size efficiency and data security. The findings demonstrate that this method provides a robust, low-cost, and scalable solution for secure invoice distribution in environments with limited infrastructure and technical literacy
Mental Health Classification Using Naïve Bayes and Random Forest Algorithms
Get PDFMental health is a crucial issue affecting individual and societal well-being. This study aims to investigate and compare the performance of Machine Learning algorithms, namely Naïve Bayes and Random Forest, for text-based mental health classification. The dataset used is the Mental Health Corpus from Kaggle, consisting of 27,977 English text messages from online forums, with binary labels (0: no indication of mental disorder, 1: indication of mental disorder) pre-annotated by the dataset creators. Text preprocessing involved lowercasing, negation handling, stopword removal, slang normalization, tokenization, and stemming. Data transformation was performed using TF-IDF. Model evaluation utilized accuracy, precision, recall, and F1-score metrics, along with 5-Fold Cross Validation. Evaluation results indicate high performance for both algorithms. Naïve Bayes achieved 88.7 % accuracy, 84.2 % precision, 95.2 % recall, and 89.3 % F1-score on the test data. Random Forest demonstrated more balanced performance with 89.3 % accuracy, 88.1 % precision, 90.5 % recall, and 89.3 % F1-score. The 5-Fold Cross Validation for Naïve Bayes yielded average scores of 88.8 % accuracy, 84.4 % precision, 94.9 % recall, and 89.3 % F1-score. In contrast, Random Forest showed averages of 89.2 % accuracy, 88.8 % precision, 89.5 % recall, and 89.3 % F1-score. While Naïve Bayes had higher recall, Random Forest exhibited the best overall performance, considering the combination of accuracy, precision, and stable generalization, making it more effective for mental health text classification.Mental health is a crucial issue affecting individual and societal well-being. This study aims to investigate and compare the performance of Machine Learning algorithms, namely Naïve Bayes and Random Forest, for text-based mental health classification. The dataset used is the Mental Health Corpus from Kaggle, consisting of 27,977 English text messages from online forums, with binary labels (0: no indication of mental disorder, 1: indication of mental disorder) pre-annotated by the dataset creators. Text preprocessing involved lowercasing, negation handling, stopword removal, slang normalization, tokenization, and stemming. Data transformation was performed using TF-IDF. Model evaluation utilized accuracy, precision, recall, and F1-score metrics, along with 5-Fold Cross Validation. Evaluation results indicate high performance for both algorithms. Naïve Bayes achieved 88.7 % accuracy, 84.2 % precision, 95.2 % recall, and 89.3 % F1-score on the test data. Random Forest demonstrated more balanced performance with 89.3 % accuracy, 88.1 % precision, 90.5 % recall, and 89.3 % F1-score. The 5-Fold Cross Validation for Naïve Bayes yielded average scores of 88.8 % accuracy, 84.4 % precision, 94.9 % recall, and 89.3 % F1-score. In contrast, Random Forest showed averages of 89.2 % accuracy, 88.8 % precision, 89.5 % recall, and 89.3 % F1-score. While Naïve Bayes had higher recall, Random Forest exhibited the best overall performance, considering the combination of accuracy, precision, and stable generalization, making it more effective for mental health text classification
Mushroom Classification Using Convolutional Neural Network MobileNetV2 Architecture for Overfitting Mitigation and Enhanced Model Generalization
Get PDFFungal identification is a significant challenge due to the morphological similarities among different species. Previous studies using Convolutional Neural Networks (CNNs) for mushroom classification still face overfitting issues, which lead to poor performance on new data. Therefore, this research develops a MobileNetV2-based Convolutional Neural Network (CNN) model capable of classifying three mushroom species (Amanita, Boletus, and Lactarius) with a primary focus on mitigating overfitting. The dataset consists of 3,210 RGB images, divided into 1,979 training data, 493 validation data, and 738 testing data. The model is developed using transfer learning with MobileNetV2, combined with additional layers such as Conv2D, pooling, and Dense, along with Dropout for regularization. The training process employs the Adam optimizer with a learning rate of 1.0×10⁻⁵ and is monitored with EarlyStopping and ModelCheckpoint. The model successfully addresses overfitting, achieving a minimal generalization gap of 1.33%, compared to 7% in previous studies. The evaluation results show a training accuracy of 77.35%, validation accuracy of 78.79%, and testing accuracy of 76.02%, with precision of 80.6% and recall of 68.1%. The consistent performance, with a maximum difference of only 2.77% across the three datasets, demonstrates superior generalization ability and provides a strong foundation for the implementation of a reliable automatic mushroom identification system.Fungal identification is a significant challenge due to the morphological similarities among different species. Previous studies using Convolutional Neural Networks (CNNs) for mushroom classification still face overfitting issues, which lead to poor performance on new data. Therefore, this research develops a MobileNetV2-based Convolutional Neural Network (CNN) model capable of classifying three mushroom species (Amanita, Boletus, and Lactarius) with a primary focus on mitigating overfitting. The dataset consists of 3,210 RGB images, divided into 1,979 training data, 493 validation data, and 738 testing data. The model is developed using transfer learning with MobileNetV2, combined with additional layers such as Conv2D, pooling, and Dense, along with Dropout for regularization. The training process employs the Adam optimizer with a learning rate of 1.0×10⁻⁵ and is monitored with EarlyStopping and ModelCheckpoint. The model successfully addresses overfitting, achieving a minimal generalization gap of 1.33%, compared to 7% in previous studies. The evaluation results show a training accuracy of 77.35%, validation accuracy of 78.79%, and testing accuracy of 76.02%, with precision of 80.6% and recall of 68.1%. The consistent performance, with a maximum difference of only 2.77% across the three datasets, demonstrates superior generalization ability and provides a strong foundation for the implementation of a reliable automatic mushroom identification system
Hybrid Decomposition ICEEMDAN-EWT Deep Learning Framework for Wind Speed Forecasting
Get PDFAccurate wind speed forecasting plays a crucial role in supporting early warning systems for extreme wind events. However, the inherent non-linearity and non-stationarity of wind speed data pose significant challenges. This study addresses these issues by evaluating the effectiveness of targeted Empirical Wavelet Transform (EWT) denoising applied to specific Intrinsic Mode Functions (IMFs) derived from Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN). Daily wind speed data from 2000 to 2023 were decomposed using ICEEMDAN, and denoising was selectively applied to IMF1, IMF2, and IMF3. Each IMF was then modeled using a Bidirectional Long Short-Term Memory (BiLSTM) network under a time-series cross-validation framework. Among all model configurations, the ICEEMDAN+EWT(IMF1 & IMF2)+BiLSTM model achieved the highest predictive accuracy, with an R² of 0.8885, RMSE of 0.501, and MAPE of 7.64%. This result outperformed both the baseline BiLSTM model (R² = 0.0501) and the ICEEMDAN+BiLSTM model without EWT denoising (R² = 0.6433). Moreover, denoising on IMF1 alone also yielded a strong performance (R² = 0.8879), emphasizing the importance of early component selection. Conversely, applying EWT to IMF2 or IMF3 individually resulted in lower R² values of 0.6639 and 0.6327, respectively, indicating limited individual contribution. These findings confirm that selective denoising, especially on the high-frequency IMFs, substantially enhances forecasting accuracy. The proposed approach holds significant potential to improve the timeliness and reliability of wind-related early warning systems, thus contributing to more effective disaster risk reduction strategies.Accurate wind speed forecasting plays a crucial role in supporting early warning systems for extreme wind events. However, the inherent non-linearity and non-stationarity of wind speed data pose significant challenges. This study addresses these issues by evaluating the effectiveness of targeted Empirical Wavelet Transform (EWT) denoising applied to specific Intrinsic Mode Functions (IMFs) derived from Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN). Daily wind speed data from 2000 to 2023 were decomposed using ICEEMDAN, and denoising was selectively applied to IMF1, IMF2, and IMF3. Each IMF was then modeled using a Bidirectional Long Short-Term Memory (BiLSTM) network under a time-series cross-validation framework. Among all model configurations, the ICEEMDAN+EWT(IMF1 & IMF2)+BiLSTM model achieved the highest predictive accuracy, with an R² of 0.8885, RMSE of 0.501, and MAPE of 7.64%. This result outperformed both the baseline BiLSTM model (R² = 0.0501) and the ICEEMDAN+BiLSTM model without EWT denoising (R² = 0.6433). Moreover, denoising on IMF1 alone also yielded a strong performance (R² = 0.8879), emphasizing the importance of early component selection. Conversely, applying EWT to IMF2 or IMF3 individually resulted in lower R² values of 0.6639 and 0.6327, respectively, indicating limited individual contribution. These findings confirm that selective denoising, especially on the high-frequency IMFs, substantially enhances forecasting accuracy. The proposed approach holds significant potential to improve the timeliness and reliability of wind-related early warning systems, thus contributing to more effective disaster risk reduction strategies