Journal Software, Hardware and Information Technology

Evaluasi Efisiensi Model Deep Learning pada Deteksi Penyakit Cabai Dengan DenseNet201 dan EfficientNetB0

2026-06-07T17:05:48+08:00

This study compares the performance of DenseNet201 and EfficientNetB0 architectures in detecting five types of conditions on chili plant leaves, namely healthy, leaf curl, leaf spot, whitefly, and yellowish. Manual disease identification by farmers has accuracy limitations and is time-consuming, hence a transfer learning-based deep learning approach is proposed as a solution. A secondary dataset of 500 digital images was divided into 80% training data, 10% validation data, and 10% testing data. The training process utilized the Adam optimization function, an initial learning rate of 0.001, and data augmentation techniques to prevent overfitting. Both models were evaluated based on standard classification metrics and memory size efficiency. The results showed that DenseNet201 achieved the highest accuracy of 92.00% with a storage model size of 141.25 MB and a training time of 5.4 minutes. On the other hand, EfficientNetB0 produced an accuracy of 88.00% with a model size 2.2 times smaller at 62.25 MB and a shorter training time of 4.3 minutes. There is a significant trade-off in both architectures; DenseNet201 is recommended for cloud computing systems that prioritize high accuracy (92.00%) with a precision of 92.32%, while EfficientNetB0 is more prospective for development on mobile devices because the number of parameters is much smaller (8,068,200 compared to 24,347,653).

Identifikasi Daya Tarik Wisata Benteng Rotterdam Berdasarkan Ulasan Wisatawan Menggunakan Topic Modeling

2026-05-31T13:59:07+08:00

The development of online review platforms provides opportunities to explore information regarding tourists’ perceptions of tourism destinations. This study aims to identify the main attractions of Benteng Rotterdam based on tourist reviews using the Latent Dirichlet Allocation (LDA) topic modeling method. The research data were obtained from TripAdvisor through a web scraping technique using Selenium and BeautifulSoup libraries. Data scraping was performed in December 2023, resulting in the collection of 679 reviews written in Indonesian. The collected reviews were then subjected to a series of preprocessing steps, namely cleaning, tokenization, stemming, and stopword removal. Topic modeling was conducted using the Gensim library in the Python programming language, while the optimal number of topics was determined using the coherence score. Visualization using word clouds and intertopic distance maps was applied to assist in interpreting the relationships among topics. The results show that the LDA model successfully generated six topics which, based on the interpretation of dominant keywords and visualization results, can be identified as the main tourist attractions of Benteng Rotterdam, including the historical value of the building, building atmosphere and condition, strategic location, social and cultural experiences, and affordable entrance fees. The findings of this study can serve as input for tourism managers in developing promotional strategies and improving tourism services based on visitor reviews.

Analisis Pengaruh Augmentasi Data pada Model MobileNetV3-Large untuk Klasifikasi Penyakit Daun Labu Manis

2026-06-02T12:58:50+08:00

Sweet pumpkin (Cucurbita moschata) is a horticultural commodity with high economic value, but its productivity has frequently decreased due to various leaf diseases. Manual disease identification still has several limitations because it is subjective, time-consuming, and inefficient for large-scale agricultural areas. This study aimed to analyze the effect of data augmentation on sweet pumpkin leaf disease classification performance using the MobileNetV3-Large model supported by Early Stopping and Learning Rate Scheduler optimization techniques. The dataset consisted of 1,000 images representing five classes, namely Healthy Leaf, Downy Mildew, Leaf Curl, Mosaic Disease, and Red Beetle. The dataset was divided into training, validation, and testing subsets using an 80:10:10 ratio. Two experimental scenarios were conducted, namely using the original dataset and the augmented dataset. The experimental results showed that the first scenario achieved an accuracy of 85.00%, while the second scenario improved the accuracy to 88.00%. The precision, recall, and F1-score values also increased from 86%, 85%, and 85% to 87%, 87%, and 87%, respectively. In addition, the model size remained lightweight at 15 MB in both scenarios. The results indicated that the combination of the MobileNetV3-Large model, data data augmentation, Early Stopping, and Learning Rate Scheduler successfully improved model generalization while maintaining computational efficiency.

Analisis Kompleksitas Waktu Eksekusi Algoritma Insertion Sort dan Selection Sort pada Berbagai Skenario Distribusi Data

2026-05-18T10:31:59+08:00

Understanding the efficiency of sorting algorithms cannot be reduced merely to theoretical analysis; rather, it depends heavily on interpreter execution within modern computing environments. Although Insertion Sort and Selection Sort share the same theoretical complexity of , their empirical performance exhibits significant divergence due to initial data characteristics and memory allocation mechanisms. This research evaluates the interaction between data structure distributions and the CPython interpreter's execution overhead on both algorithms. The experiments were conducted in an isolated Google Colab environment using a high-resolution timing instrument on dataset sizes ranging from up to numerical elements. The tests were replicated through 10 independent iterations across three distribution scenarios: random, ascending, and descending. Quantitative analysis demonstrates that the O(n^2) label does not apply uniformly in the empirical realm. At , Insertion Sort achieved its highest efficiency at 1.24 ms in the ascending case, yet suffered a drastic performance collapse, reaching 3,842.60 ms in the descending case. Conversely, Selection Sort exhibited a rigid latency resistance, remaining constant in the range of 1,950–2,150 ms across all scenarios, completely unaffected by the advantages of pre-sorted data. This extreme performance disparity in the worst-case scenario is proven to not stem purely from algorithmic computation, but rather from the severe overhead of reference counting and Python's dynamic memory management when Insertion Sort executes massive index shifts (memory writes), as opposed to Selection Sort which minimizes memory mutations.

Implementasi dan Evaluasi Sistem Absensi RFID-IoT Berbasis Cloud Menggunakan Firebase Realtime Database: Uji Ketahanan pada Berbagai Kondisi Jaringan

2026-05-18T22:56:29+08:00

Manual attendance recording in educational environments still faces various problems, such as low time efficiency, high potential for recording errors, and the potential for fraud in the attendance process. Although several attendance systems based on Radio Frequency Identification (RFID) and Internet of Things (IoT) have been previously developed to address these problems, most still use local data storage and therefore do not support real-time attendance monitoring. This study aims to develop and evaluate an RFID and IoT-based smart attendance system integrated with real-time cloud storage to improve recording efficiency and ease of attendance data monitoring. The proposed system uses a Wemos D1 Mini microcontroller as the main controller, an RC522 module as the user identification medium, and Firebase Realtime Database as cloud-based attendance data storage. Functional testing was conducted 30 times using registered and unregistered RFID cards to evaluate identification accuracy, and 30 times for each network condition (stable, unstable, weak) using NetLimiter and Clumsy software to simulate network conditions, measure response time, and assess data recording reliability to the database. The test results showed 100% identification accuracy and 100% success rate of data recording to the cloud across all network conditions, with average latencies of 1.05 seconds (stable), 1.06 seconds (unstable), and 1.50 seconds (weak). This study contributes evidence that real-time cloud-based attendance systems are feasible for adoption in educational environments with varying network quality, noting that latency increases by approximately 43% under weak network conditions.

Hybrid Deep Feature Extraction and Machine Learning for Potato Disease Classification

2026-05-14T19:01:11+08:00

Accurate classification of potato leaf diseases is crucial for improving agricultural productivity and reducing crop losses. This study proposes a hybrid classification framework that combines EfficientNetB3 as a deep feature extractor with classical machine learning classifiers, namely Support Vector Machine (SVM), Random Forest (RF), and XGBoost, to overcome the limitations of end-to-end convolutional neural networks in terms of computational cost and training time. Deep features extracted by EfficientNetB3 are used as inputs for the respective classifiers, and the proposed models are evaluated using accuracy, precision, recall, F1-score, and training time. Experimental results demonstrate that the hybrid models achieve superior performance and efficiency compared to the baseline EfficientNetB3 with Softmax classifier. Among all methods, EfficientNetB3 + XGBoost attains the highest accuracy of 0.966 with reduced training time, while EfficientNetB3 + SVM offers competitive accuracy with the fastest training process. These findings indicate that integrating deep feature extraction with traditional machine learning classifiers provides an effective and computationally efficient approach for potato leaf disease classification, making it suitable for practical implementation in precision agriculture.

Implementasi Yolov11 pada Sistem Deteksi Cacat Biji Kopi Pasca-Roasting

2026-05-15T13:56:00+08:00

Coffee bean quality after roasting is one of the important factors in determining the quality of coffee products. The process of identifying and sorting defective coffee beans is generally still performed manually, which may lead to inconsistencies and observation errors. This study aims to implement the You Only Look Once version 11 (YOLOv11) method based on Convolutional Neural Network (CNN) in an automatic post-roasting coffee bean defect detection and classification system. The research stages include image data collection, preprocessing, object labeling, dataset splitting into training, validation, and testing data, as well as model training using several hyperparameter configurations. Testing was conducted using variations in data splitting, optimizer, learning rate, and confidence threshold to obtain optimal model performance. The dataset used consisted of 694 images, including 290 primary images and 404 augmented images. The results showed that the 90:5:5 dataset splitting scenario achieved the highest mAP@50 value of 0.948. However, the 70:20:10 configuration was selected as a more stable and representative configuration because it provided a better balance between training, validation, and testing data. The use of the SGD optimizer with a learning rate of 0.01 and a confidence threshold of 0.5 produced more stable detection performance. The resulting model achieved a Mean Average Precision (mAP@50) value of 0.946 and was able to automatically detect and classify coffee bean defects effectively under various testing conditions.

Klasifikasi Sentimen Ulasan Buku Pada Platform Shopee: Studi Komparatif Naive Bayes dan SVM Dengan Pembobotan TF-IDF pada Dataset Seimbang

2026-05-08T19:37:03+08:00

The surge in informal and unstructured product review data in e-commerce has become a major obstacle in manual business evaluation. This study aims to conduct a comparative study of book product review sentiment classification on the Shopee platform using the Multinomial Naive Bayes algorithm and Support Vector Machine (SVM) with a linear kernel. The dataset used consists of 1,101 reviews obtained through web scraping techniques. The sentiment class labeling procedure was carried out manually through cross-validation by three annotators based on Shopee star rating curation and textual content analysis to ensure data validity. After going through the text pre-processing (Sastrawi) stage and Term Frequency-Inverse Document Frequency (TF-IDF) feature extraction, the model was tested and its strength was evaluated using the Stratified 10-Fold Cross Validation technique. The test results showed that the SVM algorithm produced slightly superior performance with an accuracy rate of 86.47%, precision of 88.37%, and recall of 84.04%. Meanwhile, Multinomial Naive Bayes achieved 86.29% accuracy, 87.16% precision, and 85.12% recall. In conclusion, although both models demonstrated similar robustness on a balanced dataset, SVM offered a higher level of precision in addressing the specific characteristics of the book review domain, which is characterized by lengthy and informal descriptions in the digital ecosystem.