Journal of Soft Computing Exploration

Squeeze-and-Excitation networks and attention mechanism in automatic detection of coffee leaf diseases based on images

Muhammad Izza Iqbal — 2024-12-04

This research examines the effectiveness of Squeeze-and-Excitation Networks (SENet) combined with Attention Mechanism for automated detection of coffee leaf diseases. The integration of SENet and Attention Mechanism presents a promising technological opportunity as SENet has proven effective in improving CNN performance by modeling channel interdependencies, while Attention Mechanism enables focused feature extraction on crucial leaf areas - a combination that remains underexplored in coffee leaf disease detection. Using a combination of three datasets: Coffee Leaf Diseases, Disease and Pest in Coffee Leaves, and RoCoLe.Original, comprising 3,177 coffee leaf images divided into four classes (Healthy, Miner, Phoma, and Rust), this study compares the performance of SENet against other deep learning architectures such as InceptionV3, ResNet101V2, and MobileNet. Experiments were conducted with variations in epochs (15 and 30), three data split ratios, and three optimizer types. Results demonstrate that SENet with Attention mechanism performs, achieving a peak accuracy of 96% at 30 epochs with an 80:20 data ratio and RMSprop optimizer. InceptionV3 and MobileNet showed competitive performance with 93% accuracy, while ResNet101V2 achieved 81%. Class-wise analysis reveals SENet's proficiency in detecting various coffee leaf diseases, with F1-scores 91% for all classes.

Eye disease classification using deep learning convolutional neural networks

Eko Hari Rachmawanto — 2024-12-10

This study begins with the analysis of the growing challenge of accurately diagnosing eye diseases, which can lead to severe visual impairment if not identified early. To address this issue, we propose a solution using Deep Learning Convolutional Neural Networks (CNNs) enhanced by transfer learning techniques. The dataset utilized in this study comprises 4,217 images of eye diseases, categorized into four classes: Normal (1,074 images), Glaucoma (1,007 images), Cataract (1,038 images), and Diabetic Retinopathy (1,098 images). We implemented a CNN model using TensorFlow to effectively learn and classify these diseases. The evaluation results demonstrate a high accuracy of 95%, with precision and recall rates significantly varying across classes, particularly achieving 100% for Diabetic Retinopathy. These findings highlight the potential of CNNs to improve diagnostic accuracy in ophthalmology, facilitating timely interventions and enhancing patient outcomes. For future research, expanding the dataset to include a wider variety of ocular diseases and employing more sophisticated deep learning techniques could further enhance the model's performance. Integrating this model into clinical practice could significantly aid ophthalmologists in the early detection and management of eye diseases, ultimately improving patient care and reducing the burden of ocular disorders.

Implementation of integrated temperature, humidity, and dust monitoring system on building electrical panel

Agus Khumaidi — 2024-12-04

This research aims to develop and implement an electrical power monitoring system at the Sub Sub Distribution Panel (SSDP) in the Building. The system is designed to monitor power usage in real-time, provide accurate information on energy consumption, and detect potential energy waste. The methodology used includes hardware and software design. The hardware consists of current and voltage sensors connected to a microcontroller. The data collected by the sensors is then transmitted via Wi-Fi network to the server for analysis. The software uses an Internet of Things (IoT) platform that displays the data in the form of graphs and tables. The implementation shows that the system is capable of monitoring power usage with a high degree of accuracy. The sensors used, namely PM2100 for voltage, SHT20 for temperature and humidity, and GP2Y101AU0F for dust concentration, proved effective in generating accurate real-time data. Based on the test results, the voltage measurement error with the PM2100 was only 0.035%, while the current measurement resulted in an error of 0.48%. The SHT20 sensor recorded an error of 2.4% for temperature and 1.0% for humidity. Dust measurements with the GP2Y101AU0F sensor had a very small error of 0.02%. These results indicate that the tested device has a sufficient level of precision for electrical power and environmental monitoring applications.

Development of an IoT-based temperature and humidity prediction system for baby incubators using solar panels

Radian Indra Mukromin — 2024-12-05

Baby incubators are crucial medical devices to maintain environmental stability for babies born prematurely or have health problems. This research aims to develop a prediction system for temperature and humidity variables in baby incubators by utilizing Internet of Things (IoT) technology and solar panels as the main energy source. Despite advancements in IoT-based incubator systems, existing solutions often rely on reactive approaches, making them less effective in preventing harmful environmental fluctuations. Addressing this gap, this study focuses on optimizing temperature and humidity predictions using artificial intelligence (AI) for proactive control. Using a DHT22 sensor to measure temperature and humidity, as well as a 1 Wp solar panel, the system is designed to operate autonomously in areas with limited access to electricity. The methods used include data collection, data processing to calculate correlation coefficients, and selection of linear regression models for the analysis of relationships between variables. The results showed that the linear regression model applied had a good temperature and humidity prediction with a Mean Squared Error (MSE) value of 0.45 for the training data and 7.32 for the test data.

Valid and practical integrated monitoring instrument of tahfidz qur'an

Efan Efan — 2024-12-06

In implementing tahfidz qur'an learning in Islamic boarding schools, students must face many activities, and they are usually given up to five times a day. Almost all of these activities must be recorded by the teacher in a logbook so that there is the potential for slow and invalid reporting. This study aims to create an integrated monitoring instrument of tahfidz qur'an and reveal its validity and practical values. This study was conducted using a research and development (R&D) approach. The instrument was created by combining the Analysis, Design, Development, Implementation, and Evaluation (ADDIE) development procedure and the Rapid Application Development (RAD) development procedure. Furthermore, the application of the Object-Oriented Programming (OOP) paradigm into the application creation process aims to produce a monitoring instrument that is integrated into various types of devices and can provide data and information on the student's achievement of tahfidz qur'an learning to all interested parties. The results of the validity test revealed an Aiken's V value of 0.81 so it was worthy of being tested at the implementation stage. The implementation resulted in a practicality value of 80.65% from teachers, 79.84 from parents of students, and 78.28% from the management of the boarding school. Overall, both teachers, parents of students, and management stated that this integrated monitoring instrument of tahfidz qur'an was practical during use.

Comparison of supervised machine learning methods in predicting the prevalence of stunting in north sumatra province

Vinny Ramayani Saragih — 2024-12-20

Stunting is a growth and development disorder in children caused by chronic malnutrition and repeated infections. Stunting has significant short- and long-term impacts and is one of the major health issues currently faced by Indonesia. The prevalence of stunting in North Sumatra Province is 18.9%, and the provincial government aims to reduce this prevalence to 14% by 2024. This study aims to compare the performance of several supervised machine learning methods in predicting stunting prevalence in North Sumatra Province. The data used is secondary data from 2021 to 2023, covering 33 districts/cities in the province. This study evaluates three machine learning models: Support Vector Regression (SVR), Decision Tree, and Random Forest, using evaluation metrics such as Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE). The analysis results show that Random Forest provides the most accurate and consistent predictions, with lower MSE, MAE, RMSE, and MAPE values compared to the other models in most areas. Decision Tree yields good results in some regions but tends to produce higher errors in certain cases. SVR exhibits a more varied performance, with some regions showing higher prediction errors. Overall, Random Forest is the superior model for predicting district/city-level data, although model selection should be tailored to the data characteristics and application needs