Jurnal Teknik Informatika (Jutif)

Comparative Analysis of RGB and Grayscale Pixel-Based Similarity Methods for Lung X-ray Image Retrieval in Clinical Decision Support Systems

2025-12-13T15:34:31+00:00

Chest X-ray imaging is widely used to support the diagnosis of lung diseases, yet many automated similarity techniques still rely on RGB formats, which differ from the grayscale images commonly used in clinical systems. This discrepancy raises the question of whether color information is necessary for effective similarity assessment. This study aims to evaluate the performance of RGB and grayscale pixel-based similarity methods for lung X-ray analysis and determine whether grayscale images can provide comparable similarity performance with lower computational demands. A total of 300 chest X-ray images representing normal, pneumonia, and COVID-19 categories were processed in both formats. Pixel-level similarity was calculated across 30,000 image pairings, followed by statistical testing to assess differences between methods. The results show that grayscale similarity scores closely match those of RGB, with variations generally below 0.3%. A meaningful difference was observed only in the comparison between normal and COVID-19 images, indicating that RGB may capture subtle visual variations not present in grayscale. Overall, this study demonstrates that grayscale pixel-based similarity analysis provides a reliable and computationally efficient approach, contributing to the development of lightweight medical image retrieval and clinical decision support systems in the field of health informatics.

Improving Detection Accuracy of Network Intrusions Using a Hybrid Network Intrusion Detection System Based on Isolation Forest and Random Forest Algorithms

2025-05-26T04:45:11+00:00

The growing sophistication of cyberattacks has increased the urgency of securing organizational networks, especially those handling sensitive and large-scale data. Traditional intrusion detection systems (IDS) such as Suricata rely on signature-based methods and often fail to detect zero-day or evolving threats. To address this gap, this research proposes a hybrid intrusion detection model that integrates Suricata with machine learning algorithms—Isolation Forest and Random Forest. Suricata performs real-time packet inspection and anomaly filtering, while the machine learning component enhances detection of novel threats and reduces false positives. The methodology involves capturing real-time network traffic, pre-processing data, training models on both CICIDS2017 and simulated attack data, and evaluating performance using accuracy, precision, recall, and F1-score. Experimental results show that the hybrid model achieves high detection accuracy—99.86% on simulated data and 96.33% on the CICIDS2017 dataset. Compared to standalone Suricata, the hybrid model detects more unknown threats and reduces alert fatigue by minimizing false positives. This study contributes a scalable and adaptive IDS framework that combines anomaly- and signature-based detection techniques. The proposed system enhances threat detection capabilities in enterprise-level networks and offers practical implications for intelligent cybersecurity defences. The findings advance research in computer science, particularly in the domains of machine learning applications and network security systems.

Improving the Accuracy of Stunting Prediction in Children in Pagar Alam City Using XGBoost Feature Selection and K-Nearest Neighbor Classification

2025-11-26T07:54:39+00:00

Stunting remains a major public health concern in Indonesia, including in Pagar Alam City. Early identification of at-risk children is essential to enable timely interventions and reduce long-term developmental consequences. However, predictive models such as K-Nearest Neighbor (K-NN) often experience reduced accuracy when faced with irrelevant features and imbalanced class distributions. This study integrates feature selection using Extreme Gradient Boosting (XGBoost) to enhance the predictive performance of K-NN in assessing stunting risk. Child growth data obtained from local health facilities were analyzed to build an initial baseline model, which exhibited limited accuracy due to excessive attributes and class imbalance. Through feature-importance analysis, XGBoost identified key predictors including sex, age, weight, and height. The optimized dataset was then used to retrain the K-NN model. Evaluation using accuracy, precision, recall, and F1-score demonstrated an improvement in accuracy from 85.63% to 93.72%. Beyond the computational results, this research provides significant contributions to the field of health informatics. The integration of XGBoost and K-NN offers an efficient analytical mechanism suitable for clinical decision support systems, particularly for data-driven screening in primary healthcare settings. The optimized, lightweight model can be embedded into health information systems to support child growth monitoring, strengthen evidence-based policymaking, and assist healthcare workers in targeting interventions more effectively. This approach can be replicated across other regions, supporting nationwide efforts to reduce stunting prevalence.

Proximal Policy Optimization for Adaptive Resource Allocation in Mobile OS Kernels: Enhancing Multitasking Efficiency

2025-11-23T22:55:06+00:00

Traditional mobile operating system (OS) schedulers struggle to maintain optimal performance amidst the increasing complexity of user multitasking, often resulting in significant latency and energy waste. This study aims to integrate a Proximal Policy Optimization (PPO) based Reinforcement Learning (RL) framework for predictive and adaptive resource allocation. Methodologically, we formulate the scheduling problem as a Markov Decision Process (MDP) where States (S) encompass CPU load, memory usage, and workload patterns; Actions (A) involve dynamic core affinity, frequency scaling, and cgroup adjustments; and Rewards (R) are calculated based on a weighted trade-off between performance maximization and energy conservation. A PPO actor-critic network is implemented and trained on a modified Android kernel (discount factor γ=0.99) under simulated high-load scenarios, including simultaneous video conferencing, data downloading, and web browsing. Experimental results demonstrate that the proposed RL mechanism reduces average task latency by 18% and boosts system responsiveness by 25%, while simultaneously achieving a 12% reduction in CPU power consumption compared to the baseline scheduler. These findings pioneer intelligent OS informatics, offering a robust foundation for sustainable multitasking for over a billion Android users through scalable, on-device fine-tuning.

Sentinel-2 NDVI Analysis Using GEE and QGIS for Green Open Space Sustainability Assessment in Kendari City

2025-10-29T12:48:56+00:00

Rapid urbanization has profoundly transformed land cover in many growing cities, leading to a substantial decline in Green Open Space (GOS) and a progressive deterioration of ecological functions. The continuous conversion of vegetated zones into impervious and built-up surfaces has reduced the city’s ability to absorb carbon, regulate local microclimates, and maintain overall ecological resilience. Consequently, assessing the sustainability and spatial distribution of GOS is crucial for ensuring environmentally balanced urban development and resilience to future land-use pressures. This study aims to evaluate the sustainability of urban green spaces in Kendari City through an integrated geospatial approach that combines remote sensing and open-source cloud computing technologies. Sentinel-2 Level-2A imagery was analyzed in Google Earth Engine (GEE) using the QA60 band for cloud masking and spatial clipping to accurately define the study boundaries. Normalized Difference Vegetation Index (NDVI) values were subsequently processed and classified in QGIS using a reclassification technique to distinguish vegetation density categories. The results indicate that 56.7% of the total land area, equivalent to 15,213 hectares, exhibits high greenness, reflecting dense and healthy vegetation, whereas 32.3% consists of low or non-vegetated surfaces dominated by built-up and barren lands. These findings reveal substantial spatial disparities in vegetation coverage and underscore the importance of sustainable land management and green infrastructure policies. Furthermore, this research contributes to the advancement of geospatial informatics by developing an open, reproducible workflow that integrates cloud-based computation and open-source GIS for urban ecological monitoring and sustainability assessment.

Implementation and Evaluation of Static Code Analysis to Identify Security and Code Quality Issues in Academic Information Systems

2025-10-23T09:05:56+00:00

In today's digital era, websites have become a key component of various digital services, from government and education to business. However, many security incidents occur due to undetected source code vulnerabilities, such as vulnerabilities, bugs, and code smells, which can degrade system performance and reliability. Therefore, a systematic approach is needed to detect and prevent these issues as early as possible. This study aims to implement and evaluate the effectiveness of the Static Code Analysis (SCA) method in identifying security and code quality issues in web applications. The tool used was SonarQube, which was then implemented in the SIMAK Universitas Siliwangi. Evaluation and testing were conducted on the tool's ability to detect various types of problems, its level of accuracy, and its ease of integration into the software development process. In this study, the evaluated aspects were bugs, code smells, and vulnerabilities. The results of this study found 23,241 issues, consisting of 2,356 bugs and 20,885 code smells, without any vulnerabilities found. With a problem ratio of 3.84% of the total code lines of 605,130, and a severity classification dominated by issues at the Critical and Major levels, these results provide an overview of the technical condition of the code used in SIMAK Universitas Siliwangi. This research is expected to provide practical contributions for software developers and security teams in continuously improving the quality and security of web applications. The outcomes of this study are expected to offer substantial and actionable contributions toward advancing the overall quality, robustness, and security of software systems. By strengthening these foundational aspects, the research is projected to positively influence the reliability, continuity, and long-term sustainability of academic service delivery within higher-education environments.

Comparative Analysis of Classification Models for Sales Prediction in E-commerce: Decision Tree, Random Forest, SVM, Naive Bayes, and KNN

2025-08-11T07:09:06+00:00

The swift expansion of e-commerce has markedly heightened the necessity for precise sales forecasting, essential for efficient marketing tactics and inventory control. This research evaluates five classification models—Decision Tree, Random Forest, Support Vector Machine (SVM), Naive Bayes, and K-Nearest Neighbors (KNN)—to predict sales outcomes using e-commerce transaction data. The models were assessed utilizing criteria including accuracy, precision, recall, F1-score, AUC, and Log Loss. The findings indicate that Random Forest exceeds the performance of the other models, with an accuracy of 97.5% and an AUC of 0.991, markedly outperforming the alternatives. This study presents a unique contribution by contrasting these classification models in the realm of e-commerce in Indonesia, yielding significant insights for the advancement of more effective predictive algorithms in informatics. The results not only enhance the optimization of marketing strategies but also enrich the comprehension of machine learning applications in sales forecasting. This study underscores the necessity of choosing the appropriate model for enhanced sales forecasting, with considerable ramifications for data-driven decision-making in the e-commerce sector.

Optimization Strategy for Electric Vehicle Charging Station Development at Gas Stations Using GIS-AHP-SAW Framework

2025-05-27T04:31:51+00:00

The rapid adoption of electric vehicles (EVs) requires the acceleration of electric vehicle charging station (EVCS) development. However, selecting optimal locations for EVCS development remains challenging. The EVCS Infrastructure Standard emphasizes that technical factors are essential, which aligns with earlier studies that point out the need to consider technical requirements along with sustainability criteria. This study aims to identify a novel optimization strategy for the EVCS development at gas stations, utilizing both technical and sustainability factors. We identified the gas station as an alternative site, conforming to regulatory guidelines and prior studies. This Framework integrates GIS, AHP, and SAW methods to achieve the research objectives. We evaluated the framework using suitability analysis, mathematical optimization techniques and conducted empirical study in a designated region of Indonesia to assess the practical applicability. The study's revealed substantial findings and efficient optimization strategies. The power network subcriterion ranking as the most critical in the hierarchy of criteria. The GS05 and GS22 locations attain an optimal level across all optimization scenarios. The improved accessibility of power network facilities can augment the total alternative weight by 22.5% and improve the coverage demand from 6% to 47%. The results indicated the optimization strategy focused on improving electricity network facilities at the gas station is the best strategy for EVCS Development. This framework demonstrated a replicable model for decision support systems within the domain of spatial informatics and smart infrastructure planning, specifically spatial decision support systems for EV infrastructure planning, and offers valuable insights for investor decision-making.

Complex Word Identification in Indonesian Children’s Texts: An IndoBERT Baseline and Error Analysis

2025-12-13T10:22:57+00:00

Complex Word Identification (CWI) is a crucial step for building text simplification systems, especially for Indonesian children’s reading materials where unfamiliar vocabulary can hinder comprehension. This study formulates token-level CWI for Indonesian children’s texts and establishes two baselines: an interpretable rule-based model using linguistic features e.g., length, syllable heuristics, and affix patterns, and an IndoBERT model fine-tuned for token classification. This study construct and annotate a children’s text corpus and evaluate both approaches using standard classification metrics. On the test set (22.584 tokens), IndoBERT achieves an F1-score of 0.9972 for the CWI class, substantially outperforming the rule-based baseline (F1 = 0.8607). The IndoBERT system makes only 39 errors (23 false positives and 16 false negatives), indicating near-perfect performance under the evaluated setting. Furthermore, this study provides an error analysis to highlight remaining failure patterns and borderline cases that are difficult even for contextual models. The resulting benchmark and findings contribute to Informatics/Computer Science by providing a strong baseline and analysis for educational NLP in a low-resource language setting, supporting the development of Indonesian child-oriented NLP resources and downstream text simplification tools.

Aligning Software Architecture with Cost Structure: A Comparative Study Using ATAM and Lean Canvas in Early Startup Development

2025-02-12T07:20:04+00:00

Startups in the early phase often face challenges in balancing operational efficiency with resource constraints. This research find how startups can choose software architecture to align with cost structures with the Lean Canvas framework and the Architecture Trade-off Analysis Method (ATAM). Lean canvas allows for startups to identify cost structures at an early stage and align with market demands efficiently and ATAM helps to evaluate software architecture systematically by analysing trade-offs and quality attributes. Although microservice architecture offers modularity and scalability, its implementation can lead to higher operational costs making it unsuitable for startups with limited budgets. On the other hand, monolithic architecture is more cost-effective, easy to manage and suitable for the needs of early-stage startups. This research emphasizes that systematic evaluation of software architecture based on business goals and resource limitations is essential for startup growth for sustainability. By combining Lean Canvas for business validation and ATAM for architectural decision making, startups can optimize operational and technical strategies, analyse risks, and identify trade-offs that are implemented according to business development.

MAnTra: A Transformer-Based Approach for Malware Anomaly Detection in Network Traffic Classification

2025-11-28T07:34:19+00:00

Cybersecurity is a critical priority in the ever-evolving digital era, particularly with the emergence of increasingly sophisticated and difficult to detect malware. Traditional detection techniques, such as static and dynamic analysis, are often limited in their ability to recognize novel and concealed malware that poses a threat to security systems. Consequently, this study investigates the potential of Transformer models for network traffic classification to detect anomalies associated with malware activity. The proposed approach emphasizes retrospective analysis, wherein the model is evaluated across various platforms and datasets encompassing different virus variants. By incorporating diverse types of malwares into the training data, the model is better equipped to identify a range of attack patterns. The Transformer model employed in this study was trained over 30 epochs. The evaluation results demonstrated excellent performance, achieving a training accuracy of 99.16% and a test accuracy of 99.32%. The very low average loss value of 0.01 indicates that the model effectively reduces classification errors. These findings underscore the potential of Transformer models as an efficient method for malware detection, offering greater accuracy and speed compared to traditional approaches. The results further reveal that the Transformer exhibits strong capabilities in handling sequential data, which is highly relevant to the dynamic nature of network traffic. For future research, it is recommended to explore the scalability of this method in larger network environments and assess its effectiveness in real-time detection scenarios. Expanding its application could establish the Transformer model as a more reliable and efficient solution for identifying evolving malware threats, thereby enhancing overall network security. This approach presents a robust framework for protecting systems and data against increasingly complex cyber threats.

Predictive Modeling for Underweight Detection in Toddlers Using Support Vector Machine, K-Nearest Neighbors, and Decision Tree C4.5 Algorithms

2025-11-18T12:54:24+00:00

Gizi kurang (underweight) pada balita masih menjadi tantangan utama kesehatan masyarakat di Indonesia, dengan prevalensi mencapai 15,9% berdasarkan Survei Kesehatan Indonesia tahun 2023. Kondisi ini berdampak serius terhadap pertumbuhan fisik, perkembangan kognitif, dan kualitas hidup anak. Penelitian ini bertujuan untuk mengembangkan model prediktif guna mendeteksi dini status gizi balita dengan menggunakan metode supervised machine learning. Tiga algoritma pembelajaran terawasi diterapkan dan dievaluasi, yaitu Support Vector Machine (SVM), K-Nearest Neighbor (KNN), dan Decision Tree C4.5, dengan memanfaatkan dataset berisi 9.284 catatan balita dari Kabupaten Sukoharjo yang mencakup delapan atribut dan satu label kelas status gizi. Hasil analisis menunjukkan bahwa algoritma SVM memberikan performa klasifikasi tertinggi dengan akurasi 98,56%, diikuti KNN dengan akurasi 97,99% dan Decision Tree C4.5 dengan akurasi 96,96%. Temuan ini menegaskan bahwa machine learning dapat menjadi alat yang efektif untuk identifikasi dini risiko gizi kurang pada anak, sehingga memungkinkan intervensi yang lebih cepat, tepat, dan berbasis data. Pendekatan ini berkontribusi pada peningkatan efektivitas program kesehatan anak dan mendukung pencapaian target pembangunan kesehatan nasional.

Comparative Performance Evaluation of Linear, Bagging, and Boosting Models Using BorutaSHAP for Software Defect Prediction on NASA MDP Datasets

2025-11-01T14:08:02+00:00

Software defect prediction aims to identify potentially defective modules early on in order to improve software reliability and reduce maintenance costs. However, challenges such as high feature dimensions, irrelevant metrics, and class imbalance often reduce the performance of prediction models. This research aims to compare the performance of three classification model groups—linear, bagging, and boosting—combined with the BorutaSHAP feature selection method to improve prediction stability and interpretability. A total of twelve datasets from the NASA Metrics Data Program (MDP) were used as test references. The research stages included data preprocessing, class balancing using the Synthetic Minority Oversampling Technique (SMOTE), feature selection with BorutaSHAP, and model training using five algorithms, namely Logistic Regression, Linear SVC, Random Forest, Extra Trees, and XGBoost. The evaluation was conducted with Stratified 5-Fold Cross-Validation using the F1-score and Area Under the Curve (AUC) metrics. The experimental results showed that tree-based ensemble models provided the most consistent performance, with Extra Trees recording the highest average AUC of 0.794 ± 0.05, followed by Random Forest (0.783 ± 0.06). The XGBoost model provided the best results on the PC4 dataset (AUC = 0.937 ± 0.008), demonstrating its ability to handle complex data patterns. These findings prove that BorutaSHAP is effective in filtering relevant features, improving classification reliability, and strengthening transparency and interpretability in the Explainable Artificial Intelligence (XAI) framework for software quality improvement.

Comparative Analysis Of Machine Learning Algorithms For Dengue Fever Prediction Based On Clinical And Laboratory Features

2025-09-22T03:30:36+00:00

Dengue fever (DF) remains a global health problem requiring accurate early detection to prevent severe complications. This study applies machine learning (ML) algorithms to clinical and laboratory data for improving diagnostic accuracy. Six classifiers were compared: Decision Tree (DT), K-Nearest Neighbor (KNN), Logistic Regression (LR), Naïve Bayes (NB), Neural Network (NN), and Support Vector Machine (SVM). The dataset consists of 1,003 patient records with nine feature columns, of which 989 were used after preprocessing. Class distribution was imbalanced, with 67.6% positive and 32.4% negative cases. Model performance was evaluated using 10-fold cross-validation based on accuracy, precision, recall, F1-score, confusion matrix, and ROC curve analysis. The results indicate that DT achieved the highest performance with 99.4% accuracy, 99.4% precision, 99.7% recall, and 99.6% F1-score, slightly outperforming NN. KNN, LR, and SVM produced comparable results, while NB showed substantially lower accuracy (44.3%) and limited discriminatory power. ROC analysis confirmed these findings, with DT, NN, SVM, and LR achieving AUC values between 0.992 and 0.999, whereas NB performed poorly. These findings highlight the strong potential of ML algorithms, particularly DT, to support medical decision systems, strengthen informatics-based decision support applications, and enhance the accuracy and speed of dengue diagnosis in clinical practice.

Labeling Optimization and Hybrid CNN Model in Sentiment Analysis of Movie Reviews with Slang Handling

2025-05-31T05:37:23+00:00

This research focuses on the development of a hybrid Convolutional Neural Network (CNN) model for sentiment analysis of movie comments, specifically designed to overcome the challenges of handling nonstandard language and slang. Slang is often an obstacle in sentiment analysis due to its non-standard nature and is difficult to recognize by traditional algorithms. By utilizing an kamusalay as a data preprocessing step, this research successfully converts slang words into standardized forms, thus improving the quality of data used in modeling. The data was collected through YouTube Data API on the comments of the movie “Pengabdi Setan 2: Communion” and processed using tokenization, stemming, stopwords removal, and TF-IDF feature extraction techniques. The hybrid model combines machine learning algorithms such as Naive Bayes, Logistic Regression, and Random Forest with CNN's ability to extract complex spatial patterns from text data. The evaluation results show that this model is able to achieve up to 95% accuracy, with consistently high precision, recall, and F1-score. This approach not only improves the accuracy of sentiment analysis, but also provides an effective solution for handling non-standard language variations, making it relevant for application in digital opinion analysis on social media.

Evaluating Lexicon Weighting and Machine Learning Models for Sentiment Classification of Indonesian Mangrove Ecotourism Reviews

2025-12-14T00:05:11+00:00

Sentiment analysis on ecotourism reviews presents specific challenges due to descriptive writing styles, the use of ambiguous words, and contextual meaning shifts (contextual polarity shift). These characteristics often cause lexicon-based approaches to produce unstable polarity labels. This study aims to evaluate the influence of two lexicon weighting methods, namely Mean Weighting and Summation Weighting, on the initial sentiment labeling of mangrove ecotourism reviews and to assess the performance of machine learning models trained using these labels. The research method includes text preprocessing, lexicon-based scoring using the InSet lexicon, feature extraction with Term Frequency–Inverse Document Frequency (TF–IDF), and the training of two classification algorithms, Support Vector Machine (SVM) and Logistic Regression (LR). The results show that the Mean Weighting method produces more stable polarity scores and higher model performance. The combination of SVM with Mean Weighting achieves the best results with an accuracy of 0.902, macro precision of 0.876, macro recall of 0.819, a macro F1-score of 0.841, and a weighted F1-score of 0.899. Meanwhile, LR with Mean Weighting reaches an accuracy of 0.891 with a similar performance pattern. In contrast, the Summation Weighting method results in lower performance for both algorithms. Error analysis indicates that neutral sentences and ambiguous words such as “bagus” and “ramai” frequently lead to misclassification. These findings highlight that the choice of lexicon weighting method plays a crucial role in improving sentiment classification accuracy and contributes to the development of hybrid approaches in text mining and sentiment analysis for the Indonesian language.

Optimizing Early Network Intrusion Detection: A Comparison of LSTM and LinearSVC with SMOTE on Imbalanced Data

2025-06-03T22:29:47+00:00

This study aims to improve network intrusion detection systems (IDS) by addressing class imbalance in the CICIDS 2017 dataset. It compares the effectiveness of Long Short-Term Memory (LSTM) networks and Linear Support Vector Classifier (LinearSVC) in detecting intrusions, with a focus on the impact of Synthetic Minority Over-sampling Technique (SMOTE) for balancing the dataset. The dataset was preprocessed by removing irrelevant features, handling missing values, and applying Min-Max normalization. SMOTE was applied to balance the training dataset. Results showed that LSTM outperformed LinearSVC, especially in recall and F1-score, after applying SMOTE. This research highlights the benefits of combining LSTM with SMOTE to address class imbalance in IDS and emphasizes the importance of temporal sequence models like LSTM for detecting network intrusions. Future work could involve using the full dataset, exploring advanced feature engineering, and implementing more complex architectures to further enhance performance. This research underscores the critical need for improving network security by addressing the challenges of class imbalance in intrusion detection systems, which is vital for ensuring the real-time identification and mitigation of sophisticated cyber threats in the ever-evolving landscape of network security.

Deep Learning Rnn-Lstm Model For Forecasting Tourist Visits In Yogyakarta Using Bps Time-Series Data

2025-11-24T22:10:11+00:00

Tourism is a crucial sector in Indonesia's economic growth, particularly in Yogyakarta, contributing significantly to revenue, job creation, and infrastructure development. However, the COVID-19 pandemic has significantly impacted the tourism industry, making tourist arrival forecasting crucial for effective government policy decision-making. This study aims to predict tourist arrivals in Yogyakarta using deep learning models, specifically the Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM) algorithms, chosen for their ability to process time series data and address non-linearity issues. Tourist arrival data from the Yogyakarta Central Statistics Agency (BPS) was used to train and test the model. Model evaluation was conducted using the Root Mean Squared Error (RMSE) metric to measure prediction accuracy. The results show that this model can accurately predict tourist arrival patterns, which can support strategic decision-making regarding the procurement of tourism facilities in Yogyakarta. The impact of this research is to provide practical benefits for local governments and tourism industry players in planning tourism promotion and management strategies. With more accurate predictions, relevant parties can prepare necessary resources and optimize tourism services according to projected visitor numbers.

Incremental CNN-k-NN Hybrid Facial Recognition for Helmeted Facial Recognition in IoT-Enabled Smart Parking: A Case Study at Universitas Mataram

2025-11-23T22:53:45+00:00

Helmeted rider identification challenges traditional facial recognition, especially in Indonesian campuses like UNRAM, where motorbike use is prevalent and theft risks are high. This study develops a hybrid CNN-k-NN system for secure parking access. The dataset contains 2,800 augmented images (Haar Cascade crop, 224x224 grayscale), with features extracted via VGG16/ResNet and classified using k-NN (k=1, Euclidean/Cosine). The system achieves 95.62% accuracy, with precision, recall, and F1 scores of 0.96. Incremental retraining reduces processing time to under 1 second, compared to 30 minutes for full retraining. The use of cosine similarity improves accuracy slightly over Euclidean distance. This solution enhances IoT-based smart campuses by enabling efficient, real-time identification and reducing theft by improving access control. It is adaptable to low-resource environments, supporting scalable deployments in smart parking and campus security systems.

An Interpretable Deep Learning Framework for Multi-Class Lung Disease Diagnosis Using ConvNeXt Architecture

2025-11-04T22:37:50+00:00

Lung diseases remain a major global health challenge, requiring accurate and interpretable diagnostic systems to support timely detection and treatment. This study proposes a high-fidelity deep learning approach using the ConvNeXt architecture for automated multi-class classification of chest X-ray (CXR) images into five categories: Bacterial Pneumonia, Viral Pneumonia, COVID-19, Tuberculosis, and Normal. The methodology involved preprocessing 10.095 Kaggle-sourced images (normalization, CLAHE, augmentation, resizing) and training a ConvNeXt model for 70 epochs with the Adam optimizer. The model achieved strong performance with 92.66% validation accuracy, 86.32% test accuracy, a macro-average F1-score of 0.86, and a macro-average AUC of 0.99. Grad-CAM visualizations demonstrated the model's consistent focus on clinically relevant lung regions, significantly improving interpretability and clinical applicability. This study contributes to advancing interpretable AI methods for clinical decision support in medical imaging, offering a reliable and transparent framework for automated lung disease diagnosis.

Analyzing Marketplace Reviews Using Word2Vec, CNN, and Deep K-Means with Sociolinguistic Approaches

2025-10-07T22:49:24+00:00

This study investigates the effectiveness of deep learning methods in analyzing linguistically diverse customer reviews on Shopee to generate actionable product insights. By integrating Word2Vec, Convolutional Neural Networks (CNN), and Deep K-Means clustering, the proposed workflow moves beyond simple polarity detection toward aspect-based sentiment analysis. Customer reviews were preprocessed and represented using Word2Vec (skip-gram) to capture semantic proximity across informal registers, slang, abbreviations, and code-switching. A one-dimensional CNN then classified reviews into positive and negative sentiments, achieving 93–94% accuracy with balanced F1-scores across both classes. To extract aspect-level insights, reviews were projected into a latent space via an autoencoder and clustered using K-Means, with evaluation metrics (Silhouette ≈ 0.6; DBI ≈ 0.5) confirming adequate cohesion and separation. Positive clusters highlighted product design, durability, and ease of use, while negative clusters emphasized material quality, packaging, and delivery issues. These findings demonstrate that deep learning can adapt to sociolinguistic variation in Indonesian e-commerce discourse while providing structured, socially meaningful insights. This research is significant for the field of Informatics as it advances Natural Language Processing techniques for multilingual and code-switched data, addressing a key challenge in real-world text mining applications. The approach offers practical value for sellers in improving product quality, enhancing customer satisfaction, and refining marketing strategies.

An Intelligent IoT-Based Hydroponic Irrigation System for Strawberry Cultivation Using Extreme Gradient Boosting Decision Model

2025-09-03T02:19:42+00:00

Most existing implementations rely on static rule-based or fuzzy logic control, which lack adaptability to dynamic environmental changes and often require manual tuning by experts. These limitations are particularly challenging for small-scale farmers who face constraints in technical knowledge, infrastructure, and operational flexibility. To address these issues, this study proposes an intelligent hydroponic irrigation system that embeds the Extreme Gradient Boosting (XGBoost) algorithm as a decision-making model. The system collects real-time sensor data including temperature, humidity, and light intensity, and uses the trained XGBoost classifier to determine irrigation needs with binary output (FLUSH or NO). The system was implemented on a vertical hydroponic setup for strawberry cultivation, and evaluated over a 21-day observation period. The results show that the XGBoost-based model was effective in maintaining consistent vegetative growth, with plants in upper-tier pipes achieving an average height above 25 cm by the end of the cycle. This demonstrates that the model could support responsive and resource-efficient irrigation control. Beyond technical performance, the research highlights the urgency of adopting data-driven smart farming systems to ensure sustainable food production, optimize limited resources, and empower small-scale farmers with accessible and scalable solutions. However, the proposed XGBoost model is still limited to local crops; therefore, when introducing new plant types or additional sensor inputs, parameter adjustments and retraining are required to maintain accuracy. Future improvements may include dynamic model retraining and integration with real-time feedback systems to enhance system autonomy and resilience in broader agricultural settings.

Adaptive Heuristic-Based Ant Colony Optimization for Multi-Constraint University Course Timetabling with Morning Slot Preference for Energy Efficiency

2025-12-22T23:30:46+00:00

University course timetabling is a well-known NP-hard combinatorial optimization problem that involves multiple interacting constraints, including lecturer availability, classroom capacity, time-slot allocation, and course duration. Most existing metaheuristic-based approaches primarily focus on eliminating academic conflicts, while contextual and operational aspects, such as energy efficiency, are rarely considered explicitly. In addition, standard Ant Colony Optimization (ACO) methods often suffer from premature convergence and limited adaptability during the solution search process. This study proposes an Adaptive Heuristic-Based Ant Colony Optimization (AHB-ACO) approach for multi-constraint university course timetabling with a particular emphasis on morning slot preference as an energy efficiency proxy. The proposed method extends the conventional ACO framework by integrating an adaptive heuristic mechanism that dynamically guides the solution construction process toward compact and conflict-free schedules, while simultaneously favoring morning time slots to support reduced classroom cooling demand. Hard constraints, including lecturer and room conflicts, are strictly enforced, whereas the temporal preference is modeled as a soft constraint. The performance of AHB-ACO is evaluated through extensive scheduling simulations using academic datasets under various parameter settings. Experimental results demonstrate that the proposed approach consistently produces conflict-free timetables, achieving a conflict function value of C(S)=0 with stable convergence behavior. Furthermore, parameter sensitivity analysis indicates that AHB-ACO exhibits good robustness with respect to variations in the number of ants and iterations, showing a reasonable trade-off between solution quality and computational time. Additional analysis reveals an increased utilization of morning time slots compared to non-optimized schedules, indicating the effectiveness of the proposed energy-aware preference. Overall, the results suggest that AHB-ACO provides an effective and adaptive solution for university course timetabling that not only satisfies academic constraints but also addresses operational considerations related to energy efficiency.

Validation of Question Classification Using Support Vector Machine and Intraclass Correlation Coefficient Based on the Revised Bloom’s Taxonomy

2025-06-13T07:29:23+00:00

The assessment process must be carried out accurately as it is a crucial aspect of identifying cognitive abilities in students. Cognitive ability identification needs to be done by providing exam questions that refer to the Revised Bloom's Taxonomy for difficulty-level classification to ensure students' understanding of what has been taught. The traditional manual classification process carried out by educators often requires significant time and is susceptible to subjective variability. The classification of questions from levels C1 to C6 based on the Revised Bloom's Taxonomy shows an imbalance in the data distribution for each level, leading to inaccurate classification results. The automatic classification technique using the SVM algorithm allows educators to quickly classify questions based on their difficulty levels. The automated classification technique needs to be validated to what extent the difficulty levels classified by the machine align with the perceptions of educators and students. This research will validate the results of question classification generated from the SVM algorithm, supplemented by the oversampling technique to address data imbalance. The validation method used is ICC. Applying the SMOTE oversampling technique to handle a class imbalance in the training data shows improvement, with an accuracy rate of 91% when using SMOTE compared to 83% without it. Results of the classification suitability test with the SVM algorithm by educators and students indicate a high level of agreement. The ICC Average Measures values are as follows: SVM classification is 0,979, assessment by non-science subject educators is 0,956, assessment by science subject educators is 0,991, assessment by non-science subject students is 0,982, and assessment by science subject students is 0,984. ICC testing consistently yields excellent results in non-science and science subjects, indicating that the assessments conducted by educators and students have a very high level of agreement.

Enhancing Classification of Self-Reported Monkeypox Symptoms on Social Media Using Term Frequency-Inverse Document Frequency Features and Graph Attention Networks

2025-11-27T01:02:49+00:00

Early detection of infectious diseases plays a crucial role in minimizing their spread and enabling timely intervention. In the digital era, social media has emerged as a valuable source of real-time health information, where individuals often share self-reported symptoms that can serve as early warning signals for disease outbreaks. However, textual data from social media is typically unstructured, noisy, and contextually diverse, posing challenges for conventional text classification methods. This study proposes a hybrid model combining Term Frequency–Inverse Document Frequency (TF-IDF) feature representation with a Graph Attention Network (GAT) to enhance the early detection of Monkeypox-related self-reported symptoms on Indonesian social media. A dataset of 3,200 tweets was collected through Tweet-Harvest and subsequently preprocessed and manually labeled, producing a balanced distribution between positive (51%) and negative (49%) samples. TF-IDF vectors were used to construct a document similarity graph via the k-Nearest Neighbors (k-NN) method with cosine similarity, enabling GAT to leverage both textual and relational information across posts. The model’s performance was evaluated using accuracy, precision, recall, and macro-F1, with macro-F1 serving as the primary indicator. The proposed TF-IDF + GAT model achieved 93.07% accuracy and a macro-F1 score of 93.06%, outperforming baseline classifiers such as CNN (92.16% macro-F1), SVM (85.73%), Logistic Regression (84.89%). These findings demonstrate the effectiveness of integrating classical text representations with graph-based neural architectures for improving social media based disease surveillance and supporting early epidemic response strategies.

Analyzing User Needs and Recommending Targeted Features for Bi’ih Village Tourism Website Using Text Mining and K-Means Clustering

2025-11-20T02:18:30+00:00

Tourism village websites often do not fully reflect user needs, resulting in digital services that cannot be optimally utilized by residents and potential tourists. This situation limits access to information and reduces the effectiveness of tourism promotion efforts, especially in villages that are undergoing digital transformation. This study was conducted to identify the overall needs of users and compile data-based feature recommendations for the development of the Bi'ih Village website as a durian tourism village. The research method used a quantitative approach through the distribution of an online questionnaire to 110 respondents consisting of visitors and residents, with five open-ended questions and several structured questions. The data was analyzed using text mining to find dominant words and themes, as well as the K-Means Clustering technique determined through the Elbow method to group user characteristics. The analysis results showed that there were 2,702 tokens and 677 meaningful words, with the highest demand for government information and visual tourism content. The segmentation process produced three main groups, namely Active Supporters (61.4%), Tech Enthusiasts (27.3%), and Moderate Users (11.4%). This study contributes a data-driven approach to designing more relevant and measurable features for tourism village websites. The impact is expected to increase the adoption of village digital services, strengthen tourism competitiveness, and support the acceleration of the Smart Village concept implementation. The novelty of this study lies in the integration of text mining and clustering as the basis for developing user-oriented feature recommendations.

Interpretable Hybrid YOLOv8s-GWO Framework for Bounding-Box Viral Pneumonia Detection on Kaggle Chest X-ray Images

2025-11-14T01:46:26+00:00

Viral pneumonia continues to impose a substantial global health burden, making rapid and reliable radiographic detection essential for early clinical management. This study proposes a hybrid framework integrating the YOLOv8s detection model with the Grey Wolf Optimizer (GWO) to enhance hyperparameter tuning for Viral Pneumonia identification in chest X-ray images. A curated set of Normal and Viral Pneumonia samples was manually annotated and preprocessed before training. The optimization process involved multi-stage refinement of learning rate, momentum, weight decay, and loss-gain parameters to improve convergence stability and detection accuracy. The optimized YOLOv8s + GWO model demonstrated notable performance gains, achieving 0.965 recall, 0.983 mAP@50, and 0.827 mAP@50–95 on internal evaluations. External testing further validated its robustness, delivering 98.80% accuracy, 99.48% specificity, and 97.46% sensitivity. These results highlight not only enhanced clinical diagnostic reliability but also contributions to Informatics and Computer Science, demonstrating the effectiveness of metaheuristic-guided optimization in improving deep-learning model performance, generalization, and computational efficiency for AI-driven image detection tasks.

Random Forest and Artificial Neural Network Data Mining for Environmental and Public Health Risk Modeling in Flood-Prone Urban Areas of Indonesia

2025-10-29T01:40:36+00:00

Floods in urban Indonesia pose severe environmental and public health challenges, exacerbating water contamination, vector proliferation, and disease outbreaks. Rapid urbanization, inadequate drainage systems, and climate change have intensified these impacts, emphasizing the need for integrated predictive frameworks. This study aims to develop a Data Mining (DM)-based modeling approach that combines environmental and health indicators to predict flood-related disease risks. Random Forest (RF) and Artificial Neural Network (ANN) algorithms were applied to multi-domain datasets from 30 flood-prone urban sub-districts between 2018 and 2023, encompassing rainfall, drainage density, land use, and water quality variables, integrated with disease incidence data such as diarrhea, dengue, and leptospirosis. The ANN model achieved superior predictive performance (93% accuracy, AUC 0.93) compared to RF (90% accuracy, AUC 0.90), identifying rainfall intensity, drainage density, and coliform contamination as the most influential predictors. These results demonstrate the capability of AI-driven DM techniques to capture complex interdependencies between environmental and health systems. The developed framework contributes to the field of informatics by providing a scalable, data-driven early warning tool for flood-related health risks, supporting evidence-based decision-making in disaster risk management and enhancing public health resilience in rapidly urbanizing regions.

Early Fusion of CNN Features for Multimodal Biometric Authentication from ECG and Fingerprint Using MLP, LSTM, GCN, and GAT

2025-09-08T22:27:21+00:00

Traditional authentication methods such as PINs and passwords remain vulnerable to theft and hacking, demanding more secure alternatives. Biometric approaches address these weaknesses, yet unimodal systems like fingerprints or facial recognition are still prone to spoofing and environmental disturbances. This study aims to enhance biometric reliability through a multimodal framework integrating electrocardiogram (ECG) signals and fingerprint images. Fingerprint features were extracted using three deep convolutional networks—VGG16, ResNet50, and DenseNet121—while ECG signals were segmented around the first R-peak to produce feature vectors of varying dimensions. Both modalities were fused at the feature level using early fusion and classified with four deep learning algorithms: Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Graph Convolutional Network (GCN), and Graph Attention Network (GAT). Experimental results demonstrated that the combination of VGG16 + LSTM and ResNet50 + LSTM achieved the highest identification accuracy of 98.75 %, while DenseNet121 + MLP yielded comparable performance. MLP and LSTM consistently outperformed GCN and GAT, confirming the suitability of sequential and feed-forward models for fused feature embeddings. By employing R-peak-based ECG segmentation and CNN-driven fingerprint features, the proposed system significantly improves classification stability and robustness. This multimodal biometric design strengthens protection against spoofing and impersonation, providing a scalable and secure authentication solution for high-security applications such as digital payments, healthcare, and IoT devices.

Stroke Risk Prediction using Winsorizing Interquartile Range and Tree-Based Classification with Explainable Artificial Intelligence

2025-05-26T02:22:09+00:00

According to the Global Burden of Disease (GBD) Study, stroke is the third leading cause of death globally. Recognizing its signs early is crucial for both prevention and effective treatment. Although machine learning has made significant progress in predicting strokes, many current models operate like "black boxes", making them hard to interpret and often resulting in high error rates. This study aims to enhance prediction accuracy and interpretability in stroke risk detection by integrating Winsorizing Interquartile Range (IQR) for outlier management, a tree-based classification method, and Explainable Artificial Intelligence (XAI) techniques. The proposed approach applies Winsorizing Interquartile Range to handle extreme values while employing tree-based methods for prediction due to their superior performance in processing tabular data. Additionally, Explainable Artificial Intelligence techniques are utilized to improve model transparency and interpretability. Testing was conducted using the Cerebral Stroke Prediction-Imbalanced Dataset, comparing results with various existing models. The suggested approach demonstrated the lowest prediction error rates, achieving a False Positive Rate (FPR) of 15.74% and a False Negative Rate (FNR) of 8.56%. Additionally, it attained an accuracy of 84.39%, sensitivity of 91.43%, specificity of 84.26%, Area Under the Receiver Operating Characteristic Curve (AUROC) of 94.74%, and G-Mean of 87.76%, outperforming previous studies in stroke risk prediction. The combination of Winsorizing Interquartile Range, Random Under-Sampling, tree-based classification, and Explainable Artificial Intelligence techniques effectively enhances prediction accuracy and transparency, supporting early stroke detection with improved interpretability. This study contributes to medical informatics by integrating transparent predictive models suitable for decision support systems.

Fault-Tolerant Telegram Bot Architecture for Odoo 14: Validated Production Reporting in Flexible Packaging

2025-12-05T09:49:40+00:00

In flexible-packaging manufacturing, manual reporting dramatically delays synchronization with the ERP — and that means operational latency and traceability issues. The proposed work is the design, implementation, and validation of a fault-tolerant Telegram bot interconnected with Odoo 14 for six production departments. Our bot architecture that combines conversational workflows with schema-based validation and XML-RPC for slow, large payloads, enables accurate and timely reporting. In a four-week pilot with 1,066 production entries, we achieved 98.7% field completeness and lowered reporting latency to less than 2 minutes. Manual baselines received 75% more requests for corrections. At disconnected state, the layered middleware of the system abstracted retry logic and media ingestion. Both SDG 9 (Resilient infrastructure, including ) and SDG 12 (Continue to reduce production waste at source, including consumables) are connected to the work presented here which evidence the feasibility of automatic conversational interfaces with a computer in the manufacturing informatics domain, and provide pathways towards scalable digital transformation and sustainability in the small-to-medium industry sector.

Efficient Waste Classification in Cisadane River Using Vision Transformer and Swin Transformer Architectures

2025-07-23T09:15:33+00:00

The increasing volume of waste in rivers has become a serious environmental problem. This study proposes the implementation of Artificial Intelligence (AI)-based models, specifically Vision Transformer (ViT) and Swin Transformer, for an automatic waste sorting system in the Cisadane River, Tangerang. The dataset used combines public sources and field data, processed through preprocessing and augmentation to improve robustness. Model training was conducted using k-fold cross-validation, pruning, and deployment testing on edge devices to ensure generalization and efficiency. Several architectural innovations were introduced, including Dynamic Patch Size for adapting to various waste shapes and sizes, and Spatial-Aware Attention to enhance focus on waste objects against complex river backgrounds. The evaluation involved a confusion matrix and statistical analysis using a paired t-test to validate the significance of the results. Experimental findings show that Swin Transformer achieved the highest accuracy of 94.2%, surpassing ViT at 91.8%, with precision of 93.5%, recall of 92.7%, and F1-score of 93.1%. Swin Transformer also proved more reliable in dynamic lighting and cluttered environments. This study demonstrates the potential of Transformer-based architectures in automatic waste classification, contributing to smarter and more efficient AI-based environmental management technologies.

Optimization of Software Effort Estimation Using Hybrid Consistent Fuzzy Preference Relation and Least Squares Support Vector Machine

2025-11-28T07:34:59+00:00

The success of software project management hinges on the ability to reliably forecast development effort. However, achieving precise estimates is notoriously difficult, primarily due to inherent project complexities and numerous uncertain variables. While various techniques exist, no single method has proven consistently reliable, leading to inaccurate scheduling and cost overruns. This study aims to develop a more accurate and robust estimation model by hybridizing a multi-criteria decision-making (MCDM) method for handling uncertainty with a machine learning algorithm for predictive modeling. The proposed approach integrates the Consistent Fuzzy Preference Relation (CFPR) method to derive consistent weights for cost drivers from expert judgments. These weights are then used as Effort Adjustment Factors (EAF) to preprocess the COCOMO and NASA datasets, which are subsequently modeled using the Least Squares Support Vector Machine (LSSVM). Evaluation of the hybrid CFPR-LSSVM model confirmed its enhanced predictive accuracy. For the COCOMO dataset, the model yielded an MMRE of 28.463% and an RMSE of 0.4705. Its performance on the NASA dataset was particularly remarkable, with results indicating an MMRE of 1.104% and an RMSE of 0.4593, demonstrating a level of precision that underscores the model's effectiveness. This research contributes a novel hybrid framework that effectively combines consistent fuzzy preference handling with powerful non-linear regression. By providing a more structured and robust methodology for managing uncertainty, this approach offers a substantial advancement in software effort estimation, delivering more reliable predictions for improved project planning.

A Hybrid Deep Learning Architecture for Cost-Effective, Real-Time IV Infusion Anomaly Detection using IoT Sensors

2025-11-24T03:41:52+00:00

Intravenous (IV) infusion therapy is a critical medical procedure, yet manual monitoring increases the risk of complications such as air embolism and irregular infusion flow, particularly in resource-constrained environments. Although several automated infusion monitoring systems have been proposed, their high implementation cost limits practical adoption. This research develops a low-cost IoT-based infusion monitoring system capable of real-time anomaly detection using a multi-architecture machine learning approach. The proposed prototype integrates an ESP32 microcontroller with load cell (HX711) and optical (LM393) sensors to acquire time-series infusion data. Ten models from classical machine learning, deep learning, hybrid, and ensemble categories were evaluated using a dataset of 10,420 records under a unified experimental setup. The results show that XGBoost had a perfect recall (1.0000) and a strong PRAUC, while the LSTM Autoencoder had the highest F1-Score (0.9343) and precision (0.8934). The best overall performance came from hybrid and ensemble methods, with CNN–LSTM having an F1-Score of 0.89, a recall of 0.99, and a precision of 0.80. This means they would be great for clinics where being sensitive is very important. The research shows that using a low-cost IoT infrastructure with carefully chosen deep learning or ensemble models can help find problems in real time. A web dashboard explains how the technology operates and its capabilities. This study examines a cost-effective and easily scalable method to enhance infusion safety in hospitals with limited financial resources.

Development of Mobile Quran App with Screen Time Monitoring Using DRM, Agile, and Sus-Use Testing

2025-10-29T10:31:13+00:00

The rapid growth of mobile applications has changed user behavior in the digital age, including how individuals interact with religious content. However, excessive use of social media has led to behavioral problems such as doom scrolling, zombie scrolling, and digital addiction, phenomena collectively known as “brain rot,” which negatively impact cognitive, emotional, and spiritual well-being. This study aims to develop and evaluate Quran Break, a mobile Quran application that integrates screen time monitoring as a digital behavior intervention to encourage users to stop scrolling and engage in reading the Quran. The methodology applies the Design Research Methodology (DRM) through four iterative stages, supported by an Agile development model with short, adaptive sprints that enable continuous feedback and improvement. 18 participants were involved in usability testing using the System Usability Scale (SUS) and the Usability, Satisfaction, Ease of Learning, and Ease of Use (USE) questionnaire. The results showed that the application achieved an average SUS score of 75 (Good) and a USE score of 87.7% (Very Good), indicating that Quran Break is effective, useful, and easy to use. This discovery contributes to the fields of Religious Informatics and Human-Computer Interaction (HCI) by integrating persuasive technology into faith-based digital systems, supporting digital well-being, and promoting a balanced interaction between technology use and spiritual activities.

Integration of Squeeze-and-Excitation in Densenet-121 for Classifying Real and AI-Generated Images

2025-10-02T22:40:52+00:00

Recent advancements in generative technologies, such as Generative Adversarial Networks (GANs) and Latent Diffusion Models, have enabled the creation of AI-generated synthetic images that are increasingly indistinguishable from real ones, posing significant challenges for verifying the authenticity of visual content. This study develops a DenseNet-121 model with hyperparameter optimization and the integration of Squeeze-and-Excitation (SE) attention mechanisms at Early, Mid, and Late positions. Experiments were conducted using the CIFAKE dataset with a resolution of 32×32 pixels to compare the baseline Plain model with three SE variants. Hyperparameter optimization was applied to maximize model performance. The results demonstrate that the Plain DenseNet-121 with optimized hyperparameters achieved an accuracy of 98.52%, outperforming the standard configurations reported in previous studies. The integration of SE yielded varied outcomes, where Mid SE attained the highest accuracy of 98.56%, while Early SE (98.45%) and Late SE (98.48%) exhibited greater stability with lower standard deviations. These findings highlight that combining hyperparameter optimization with appropriate SE placement can enhance model performance for classifying real and AI-generated images. Moreover, SE placement at different positions (Early, Mid, Late) has a significant impact on feature representation and generalization in synthetic image classification, which is increasingly important given the growing difficulty of distinguishing real from AI-generated images.

Detection of Endangered Indonesian Species Across Multiple Taxonomic Classes Using Faster R-CNN

2025-06-03T17:29:31+00:00

Indonesia’s rich biodiversity includes many endangered species across various taxonomic groups. This study presents a Faster R-CNN deep learning model to detect ten endangered Indonesian species, covering birds, reptiles, mammals, and fishes. A custom dataset with diverse images was annotated and used to train the model with transfer learning on the Detectron2 framework. Evaluation using COCO metrics yielded an average precision (AP) of 54.93%, with the Komodo Dragon achieving the highest AP (82.57%) and Wallace’s Standardwing the lowest (30.82%). The model excels at detecting larger, distinct species but has difficulty with smaller or camouflaged ones in complex environments. Training results confirm that transfer learning aids performance despite limited data. Analysis of misclassifications suggests the need for additional data modalities or context to improve accuracy. This work highlights the potential of Faster R-CNN for automated endangered species monitoring in Indonesia and recommends dataset expansion, data augmentation, and model refinement to enhance detection, particularly for challenging species. This study contributes to computer vision applications in conservation, particularly within low-resource biodiversity contexts.