Underwater fish detection is a key task in marine monitoring and aquaculture, but its performance is often degraded by low illumination, turbidity, scattering, and color distortion in underwater environments. This paper presents an underwater fish detection framework that integrates CLAHEbased image enhancement with YOLOv3 object detection and demonstrates its deployment on an edge AI platform using a Streamlit-based interface. The DeepFish dataset (6,517 images with 15,463 annotated fish) was used for training and evaluation, while a local Indonesian dataset from Sabang, Aceh (3,111 annotated images) was employed to evaluate crossdomain generalization. Several CLAHE variants were quantitatively compared using UIQM, UCIQE, and LOE, with Blending CLAHE combined with Percentile Stretching selected as the optimal pre-processing method. Detection performance was evaluated under multiple scenarios, including training and testing on original and enhanced images, direct cross-domain testing, and fine-tuning on local data. Experimental results show that image enhancement improves detection accuracy on DeepFish (mAP@0.5 from 96.15% to 97.05%), while fine-tuning is necessary to mitigate domain shift and achieve reliable performance on local waters. The proposed end-to-end system was successfully deployed on a Jetson Orin Nano and operates entirely offline, providing real-time detection results through a web-based interface.

Aceh is an exceptional region in Indonesia characterized by significant biodiversity and abundant natural wealth. Despite these numerous benefits, there is limited information regarding the potential of biodiversity in this region. Therefore, this research aimed to design and develop Website and Android-based applications to facilitate access to information regarding biodiversity in Aceh. In the development process, the Waterfall method was used, and the applications consisted of two main components, namely the back-end and the front-end. The users were divided into three categories, including contributors, verificators, and admins. After the development was complete, functionality testing was conducted using the black box method, and usability was examined using the Post-Study System Usability Questionnaire (PSSUQ). The results of functionality testing showed that both applications operate efficiently, thereby providing users with a satisfying experience when accessing information about Aceh biodiversity. The usability score was approximately 7 for Website-based and 6 for Android-based applications, showing a high level of usability.

Hepatitis C is a pressing global health issue that urgently requires the development of effective antiviral medications. In this study, we focus on targeting the Hepatitis C virus non-structural protein 5B (NS5B) polymerase, a key enzyme in viral RNA replication, to hinder the viral life cycle and reduce viral load. We introduce a computational approach that combines multiple LightGBM models to predict the bioactivity of Hepatitis C virus NS5B inhibitors with enhanced performance. By leveraging a voting mechanism, we achieve a higher predictive performance that surpasses individual LightGBM models. Our model achieves an R-squared (R2) value of 0.760, indicating strong predictive capability, along with a root mean squared error (RMSE) of 0.637, a mean absolute error (MAE) of 0.456, and a Pearson correlation coefficient (PCC) of 0.872, demonstrating the model's precision in predicting inhibitor potency and its strong linear correlation with experimental values. To enhance the interpretability of the model, we performed SHAP analysis, which identified critical molecular features influencing bioactivity and facilitated a deeper understanding of the model's decision-making process. Validation through Y-Scrambling tests confirmed that our model's accuracy significantly exceeds what would be expected by random chance alone, ensuring its robustness and reliability. This study demonstrates the power of ensemble machine learning in computational chemistry and drug design, offering a methodologically transparent and interpretable framework for predicting compound potency, which is critical for virtual screening in early-stage drug development. Our approach not only accelerates the discovery of potent NS5B inhibitors but also provides insights into the molecular determinants of bioactivity, underscoring the potential of machine learning in advancing antiviral research. Future research should focus on integrating QSAR modeling with other computational methods and experimental validation to fully realize the potential of this approach in discovering novel HCV therapeutics.

Computational Thinking (CT) is an essential 21st-century skill to prepare students for higher education and future careers. However, comprehensive insights into how CT is effectively implemented in mathematics learning regarding strategies, suitable topics, and integration trends are still limited. This systematic review explores empirical studies on CT in mathematics education from December 2019 to November 2024, sourced from Emerald, EBSCO, and ProQuest databases. Following PRISMA guidelines, 22 articles were selected from an initial 8,518 based on defined inclusion and exclusion criteria. The findings show that CT strongly supports students’ problem-solving skills, particularly through Project-Based Learning (PjBL), which fosters engagement, collaboration, and algorithmic thinking. Geometry and statistics emerged as the most effective topics for developing CT, as they promote decomposition, pattern recognition, and abstraction skills aligned with junior high school cognitive development. Although CT-related research varies in focus, integrating CT into mathematics remains vital, especially with the rise of digital tools and interdisciplinary learning. This review provides insight into current research trends, key strategies, and appropriate mathematical content for CT development. Recommendations include providing CT training for teachers, embedding CT into the curriculum, and encouraging interdisciplinary collaboration to equip students with the digital-age competencies needed for real-world problem-solving and conceptual understanding.

Mushroom poisoning remains a public health concern, often caused by misidentifying toxic species that visually resemble edible ones. This study investigates the feasibility of using a Convolutional Neural Network (CNN) to classify five mushroom species, Amanita caesarea, Amanita phalloides, Cantharellus cibarius, Omphalotus olearius, and Volvariella volvacea into toxic and non-toxic categories based on image data. A dataset of 137 images was collected and preprocessed through resizing, normalization, and data augmentation. A modified AlexNet-based CNN was trained and evaluated using accuracy, precision, recall, and F1-score. The best-performing model achieved a validation accuracy of 0.40, indicating limited discriminative capability. These findings highlight that the dataset size is insufficient for training a CNN from scratch and that the model cannot reliably distinguish species with subtle morphological differences. The study concludes that larger datasets, improved image quality, and transfer learning approaches are essential for achieving practical and deployable mushroom classification performance.

Stunting remains a critical public health concern in developing countries such as Indonesia, where early identification of at-risk populations is vital for effective intervention. However, predictive efforts are often challenged by class imbalanced datasets that bias conventional classifiers toward the majority class. This study proposes a hybrid soft voting ensemble (SVE) model that combines logistic regression (LR), random forest (RF), extreme gradient boosting (XGBoost), and support vector machine (SVM) to improve stunting risk classification. The model employs synthetic minority oversampling and edited nearest neighbor (SMOTEENN) to address class imbalance by oversampling the minority class and removing noisy instances from the majority class. Data from the 2023 West Sumatra Family Data Update were partitioned into training and testing sets using an 80:20 split. Model performance was evaluated using accuracy, precision, recall, and F1-score. Before resampling, XGBoost achieved the best individual performance with 74.56 % accuracy and 70.45 % recall. After applying SMOTEENN, all models improved, with XGBoost reaching 91.82 % accuracy and 91.74 % recall. The best-performing hybrid ensemble, combining RF and XGBoost, achieved 91.95% accuracy and 93.21 % recall, demonstrating the effectiveness of integrating resampling techniques with ensemble learning to enhance prediction reliability in public health analytics.

In the digital era, sentiment analysis represents a significant domain within natural language processing (NLP). Nevertheless, research on Indonesian regional languages, particularly Acehnese, remains limited. The primary challenges involve the unavailability of representative datasets and the absence of BERT-based models optimized through the Masked Language Modeling (MLM) approach. Existing models, such as IndoBERT, are trained on Indonesian corpora and therefore fail to adequately capture the unique linguistic features of Acehnese. This study develops the AcehX Sentiment dataset and introduces the AcehXBERT model through pre-training IndoBERT-base with the MLM approach on the AcehX corpus. The model is then fine-tuned for Acehnese sentiment classification tasks. Experimental results show an F1-macro score of 82.50% on the AcehX Sentiment dataset and 81.89% on the NusaX Sentiment dataset, outperforming NusaBERT. These findings highlight the importance of adapting pretrained models and tokenizers for regional languages, while supporting the preservation and technological integration of the Acehnese language.

The increasing complexity of urban public spaces, such as the Malioboro pedestrian area in Yogyakarta, requires more intelligent and adaptive crowd monitoring systems than conventional manual CCTV supervision. This paper proposes an end-to-end deep learning system for automatic group activity recognition from video. The system integrates YOLOv8 for pedestrian detection and tracking, 3D ResNet-18 for individual human activity recognition, and a 1D CNN with temporal attention for group-level inference based on activity ratio sequences. Experimental results show that the human activity recognition model achieves 93% accuracy with a 92% average F1-score, while the group activity model reaches 99.03% accuracy and a 99.03% Macro F1-score, outperforming GRU, TCN, and LSTM baselines. Evaluation under real-world conditions indicates that the system can adapt to varying lighting and crowd densities, although performance degrades under extreme conditions due to cascading errors from the individual recognition stage. These findings demonstrate the system’s potential for real-time collective behavior understanding, supporting automated crowd monitoring and sustainable urban area management.

Transformer-based language models have achieved remarkable success across diverse natural language processing (NLP) tasks. However, their ability to model sequential dependencies remains limited, particularly in low-resource and morphologically rich languages, such as many of Indonesia's regional languages. This paper introduces a hybrid architecture that integrates Bidirectional Long Short-Term Memory (BiLSTM) and Bidirectional Gated Recurrent Unit (BiGRU) layers on top of the NusaBERT backbone, a Transformer pretrained on Indonesian and regional corpora. We conduct extensive fine-tuning on three benchmark datasets—NusaParagraph, NusaTranslation, and NusaX—targeting multi-class classification tasks. The hybrid variants differ in their pooling strategies (CLS token, last hidden state, mean, and max), and also include a concatenated embedding approach, where the last four hidden layers of NusaBERT are fused to enrich semantic representation, while maintaining identical training configurations. Experimental results demonstrate that the hybrid models outperform strong Transformer baselines, with the NusaBERT+BiGRU with mean pooling variant achieving a macro F1-score of 83.34% on the NusaX sentiment classification task. Further evaluation on previously unseen languages, including Batak Toba, Madurese, and Ngaju, reveals improved generalization and mitigated catastrophic forgetting. These findings highlight the efficacy of combining Transformer-based contextual embeddings with recurrent architectures for robust, low-resource multilingual NLP, and support their potential in fostering language preservation and digital inclusion.

Manual coffee bean sorting in Indonesia is laborintensive, subjective, and limits access to premium international markets. Edge Artificial Intelligence (AI) offers a transformative solution for real-time quality assessment at the source. This work develops an Edge AI system for binary classification (normal vs. defective) of Arabica coffee beans using a novel multi-dataset approach. Unlike prior works focusing primarily on cloud-based accuracy metrics, this research uniquely integrates: (1) rigorous comparative analysis of EfficientNetV2-Small and FocalNet-Tiny deployed on NVIDIA Jetson Orin Nano under FP32, FP16, and INT8 quantization, (2) comprehensive sustainability assessment using Green Software Foundation’s Software Carbon Intensity (SCI) specification (2024) providing transparent carbon accounting measured with CodeCarbon library, and (3) multi-dataset integration combining two independently captured repositories totaling 5,958 harmonized images. EfficientNetV2-Small FP16 achieves optimal performance with 98.83% accuracy, 24.73 images/sec throughput, 40.43 ms latency, and lowest SCI of 0.103 gCO2e per 1,000 images. This work establishes new benchmarks for sustainable, deployment-ready edge-AI in agricultural quality control, positioning Indonesian coffee for "Green-AI sorted" premium market differentiation.

Atopic dermatitis (AD) is a chronic inflammatory skin condition with increasing prevalence in Indonesia, presenting significant challenges in diagnosis and management. This study introduces a novel approach to objectively classify AD severity using advanced machine learning techniques and comprehensive color feature extraction from clinical images. A unique dataset of 3,037 high-resolution photographs from 250 Acehnese patients was collected at RSUD Dr. Zainoel Abidin, captured using a standardized 12-megapixel smartphone protocol. Ninety color features were extracted from multiple color spaces to capture the diverse visual characteristics of AD lesions. Four state-of-the-art machine learning models - XGBoost, CatBoost, LightGBM, and Random Forest - were employed to classify AD severity into four categories: None, Mild, Moderate, and Severe. Models were trained and evaluated using various performance metrics including accuracy, precision, recall, and F1-score. Results showed high performance across all models, with LightGBM achieving the highest accuracy of 95.07%, followed closely by XGBoost at 94.74%. SHAP analysis revealed that standard deviation features, particularly those related to hue and chrominance variability across different color spaces, were most influential in severity classification. This research demonstrates the potential of combining advanced color feature extraction with machine learning to create an objective, consistent, and accessible tool for AD severity assessment. The findings have significant implications for improving AD management in resource-limited settings and standardizing severity assessment in clinical practice and research

Psoriasis is a chronic skin condition with challenges in the accurate assessment of its severity due to subtle differences between severity levels. The aim of this study was to evaluate deep learning models for automated classification of psoriasis severity. A dataset containing 1,546 clinical images was subjected to pre-processing techniques, including cropping and applying noise reduction through median filtering. The dataset was categorized into four severity classes: none, mild, moderate, and severe, based on the Psoriasis Area and Severity Index (PASI). It was split into 1,082 images for training (70%) and 463 images for validation and testing (30%). Five modified deep convolutional neural networks (DCNN) were evaluated, including ResNet50, VGGNet19, MobileNetV3, MnasNet, and EfficientNetB0. The data were validated based on accuracy, precision, sensitivity, specificity, and F1-score, which were weighted to reflect class representation; Pairwise McNemar's test, Cochran's Q test, Cohen's Kappa, and Post-hoc test were performed on the model performance, where overall accuracy and balanced accuracy were determined. Findings revealed that among the five deep learning models, ResNet50 emerged as the optimum model with an accuracy of 92.50% (95%CI: 91.2-93.8%). The precision, sensitivity, specificity, and F1-score of this model were found to be 93.10%, 92.50%, 97.37%, and 92.68%, respectively. In conclusion, ResNet50 has the potential to provide consistent and objective assessments of psoriasis severity, which could aid dermatologists in timely diagnoses and treatment planning. Further clinical validation and model refinement remain required.

The Kovats retention index is a critical parameter in gas chromatography used for the identification of volatile compounds in essential oils. Traditional methods for determining the Kovats retention index are often labor-intensive, time-consuming, and prone to inaccuracies due to variations in experimental conditions. This study presents a novel approach combining Artificial Neural Networks (ANN) with Particle Swarm Optimization (PSO) to predict the Kovats retention index of essential oil compounds more accurately and efficiently. The ANN-PSO hybrid model leverages the strengths of both techniques: the ANN's capacity to model complex nonlinear relationships and PSO's capability to optimize hyperparameters by finding the global optimum. The model was trained using a dataset of 340 essential oil compounds with molecular descriptors, with the performance evaluated based on Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). Results indicate that a simpler ANN configuration with one hidden neuron achieved the lowest RMSE (80.16) and MAPE (5.65%), suggesting that the relationship between the molecular descriptors and the Kovats retention index is not overly complex. This study demonstrates that the ANN-PSO model can serve as an effective tool for predictive modeling of the Kovats retention index, reducing the need for experimental procedures and improving analytical efficiency in essential oil research.

Breast cancer is one of the highest known causes of death in Indonesia, especially for women. Pogostemon cablin Benth is one of the Aceh's endemic herbal plants that has been studied to have potential such as anti-inflammatory, antiproliferative, antioxidant, antimicrobial, and proapoptotic. Network pharmacology approach was conducted to explore and analyze the potential of P. cablin as an anti-breast cancer. P. cablin plant compounds were obtained from GCMS and breast cancer genes data were obtained from OMIM, GeneCard, and DisGeNet databases. Target proteins and pathways involved were identified using STRING-DB and Metascape. Network analysis was performed using Cytoscape. A total of 65 plant compounds with 554 target proteins and 1854 disease genes were obtained. Based on the results of combining protein targets using Venn diagrams, 138 overlapping proteins between drug compounds and breast cancer disease targets were identified. Based on KEGG and GO analysis, P. cablin is known to have potential in breast cancer treatment/therapeutic mechanisms. Based on the "compound-protein target-pathway" multi-target mechanism, pogostol; 1H-Cycloprop[e]azulen-7-ol, decahydro-1,1,7-trimethyl-4-methylene-, [1ar-(1aα,4aα,7β,7aβ,7bα)]-; 3-Hexen-1-ol, 2,5-dimethyl-, acetate, (Z)-, 5β,7βH,10α-Eudesm-11-en-1α-ol; Acetic acid, 3-hydroxy-6-isopropenyl-4,8a-dimethyl 1,2,3,5,6,7,8,8a-octahydronaphthalen-2-yl ester; and Humulenol-II interacts with proteins that play a significant role in breast cancer, which are MAPK1, EGFR, TNF, AKT1, and JAK2. Hence, it can be concluded that P. cablin has good potential to be a source of therapeutic treatment against breast cancer. However, it needs to be tested clinically to further determine the effects of this P. cablin compound.

The 2024 general election in Indonesia has been a major highlight in the development of democracy, triggering diverse emotional reactions amongst society. Various emotions such as anger, anticipation, disgust, fear, joy, sadness, surprise, and trust appeared on various social media platforms, especially X platform. This research uses datasets obtained using crawling techniques on the X platform. The method used is the NRC Lexicon approach to label emotions by mapping words that contain emotions in the text. Anger emotion is the most dominant emotion with a percentage reaching 33.36%. This research involves the use of deep learning models, namely Recurrent Neural Network (RNN) and Long Short-Term Memory (LSTM), to classify emotions from the analyzed texts. In addition, this research also explores the application of the SMOTE technique to handle class imbalance in the data. With the application of SMOTE to the RNN model, the accuracy achieved was 74.47%, and in the LSTM model with the application of SMOTE, the accuracy reached 90.49%. The LSTM model (without SMOTE) achieved an accuracy value of 98.52%, precision 90.48% recall 87.48%, and f1-score 87.48%. While the RNN model (without SMOTE) achieved 92.52% accuracy, 90.48% precision, 85.75% recall, and 87.48% f1-score. The comparison with the RNN model confirms the superiority of LSTM in overcoming complexity and long-term patterns in structured text data.

This research implements an Intelligent Information System for Text Recognition on Indonesian Identity Cards (eKTP) based on Deep Learning architecture. The aim of this research is to develop a system capable of accurately and efficiently detecting and recognizing text on e-KTP images. The first phase of this research involves selecting ideal images from the e-KTP dataset. Subsequently, pre-processing is conducted to crop the image sides and merge them with background images to create a more diverse dataset and obtain the e-KTP image masks as training data. After selecting the ideal images, the total training data amounts to 144 images. The U-Net architecture is chosen as the deep learning method for image segmentation in this research, while for text detection and recognition, two neural networks, CRAFT and TRBA (TPS-ResNet-BiLSTM-Attention), are employed. The testing, based on accuracy values, Dice coefficients, and IoU for the segmentation process, results in a U-Net model accuracy of 99.52%. Text detection and recognition follow confidence values.

Atopic dermatitis, also known as eczema, is a common chronic skin condition that affects millions of people worldwide. It is characterized by inflamed, itchy, and dry skin, often accompanied by redness, swelling, and the formation of small bumps or blisters. This condition can vary in severity, ranging from mild and intermittent to severe and persistent. Precise and dependable decision-making by a dermatologist is essential to ensure improved therapy and risk stratification for the patient. Usually, dermatologists depend on visual assessment to ascertain a severity score, but this approach is subjective and can vary among different doctors. This research aim to classify the atopic dermatitis severity based on advance machine learning model such as XGBoost, CatBoost, LigthGBM, and Random Forest. Among the evaluated models, the CatBoost model showed as the most effective and reliable in classification, It achieved an accuracy of 89%, precision of 90%, recall of 89%, and an F-Score of 89%.

This study investigates the application of the Gradient Boosting machine learning technique to enhance the classification of Atopic Dermatitis (AD) skin disease images, reducing the potential for manual classification errors. AD, also known as eczema, is a common and chronic inflammatory skin condition characterized by pruritus (itching), erythema (redness), and often lichenification (thickening of the skin). AD affects individuals of all ages and significantly impacts their quality of life. Accurate and efficient diagnostic tools are crucial for the timely management of AD. To address this need, our research encompasses a multi-step approach involving data preprocessing, feature extraction using various color spaces and evaluating classification outcomes through Gradient Boosting. The results demonstrate an accuracy of 93.14%. This study contributes to the field of dermatology by providing a robust and reliable tool to support dermatologists in identifying AD skin disease, facilitating timely intervention and improved patient care.

Classification is the process of building a model that can distinguish between different classes of data. The model aims to predict the class of testing data based on patterns or relationships learned from training data. One of the data processing algorithms used to build classification models is Categorical Boosting (CatBoost). However, in general, the resulting models are difficult to interpret. To facilitate the interpretation of complex classification models, methods such as SHAP (SHapley Additive exPlanations) are needed. SHAP is a method to explain individual predictions. SHAP is based on the game theoretically optimal shapley values. In this study, an analysis of important SHAP variables was conducted on the CatBoost classification model to identify variables characterizing occurrences of food insecurity in households. The data used in this study was obtained from the Survei Sosial Ekonomi Nasional (Susenas) in March 2021 in Aceh Province, sourced from the Badan Pusat Statistik (BPS). There are 13,126 observations in the research data. The results from four evaluated classification models on the testing data showed that the best model had accuracy, sensitivity, specificity, and AUC values of 0.703, 0.349, 0.798, and 0.637, respectively. Furthermore, the results of the analysis of important SHAP variables showed that the variables number of household members who smoke ( ), education of the household head ( ), wall types ( ), drinking water source ( ), and decent sanitation ( ) significantly contributed to the occurrences of food insecurity in households in Aceh Province in the year 2021.

expected model scan fosters students' mathematical connection skills, namely through the application of the AIR (Auditory Intellectual Repetition) model assisted by geogebra software. The purpose of this study was to determine differences in students' mathematical connection abilities after obtaining learning by applying the AIR (Auditory Intellectual Repetition) model assisted by geogebra software. This study used a quantitative approach with a pre-test and post-test control group design. The population of this study was class IX students of SMPN 1 Peukan Pidie by taking samples of two classes consisting of an experimental class and a control class. The sample selection was done by random sampling. The instrument used is a mathematical connection ability test. The data analysis technique uses the ANOVA test. Based on the results of the study, it was found that there were differences in the mathematical connection abilities of students who were taught through the application of the AIR (Auditory Intellectual Repetition) model assisted by geogebra software with conventional learning. Furthermore, the results of this study also identified that there was no interaction between learning and student level on students' mathematical connection abilities.