TIERS Information Technology Journal

Efficient Rice Leaf Disease Classification Using Enhanced CAE-CNN Architecture

Destia Suhada — 2025-12-30

This study introduces an enhanced Convolutional Autoencoder–Convolutional Neural Network (CAE–CNN) model designed for efficient and accurate classification of rice leaf diseases. This study aims to develop an architecture that achieves high accuracy while maintaining computational efficiency, serving as an integrative and applicative technical innovation for rice disease detection. The proposed architecture integrates a Squeeze and Excitation Block (SE-Block), Global Max Pooling (GMP), and Separable Convolution to improve feature extraction while reducing the number of parameters and inference time. A total of 7,430 labeled images from five rice disease classes were used for model training and evaluation. The model was optimized using Optuna-based hyperparameter tuning and validated through an ablation and comparative analysis to assess the impact of each component. Experimental results show that the proposed model achieves 99.39% accuracy with only 85,859 parameters, a compact size of 0.28 MB, and inference time at 0.06657 ms/image with 15,213 FPS. These findings demonstrate that the proposed CAE–CNN effectively combines high accuracy and low computational cost, making it highly suitable for real-time and edge-based rice disease classification systems.

Operating Room Scheduling Optimization Under Surgeon and Nurse Constraints Using Genetic Algorithm

Ayu Swilugar — 2025-12-30

Operating room scheduling is a complex problem due to the limited availability of surgeons, nurses, and operating rooms, as well as the variability in surgery durations. Inaccurate predictions or scheduling may cause conflicts such as overlapping surgeon schedules, violations of contamination level restrictions, and unavailability of nurses or rooms, ultimately reducing the quality of hospital services. This study integrates multiprocedure surgery duration prediction using machine learning with scheduling optimization based on genetic algorithms. The prediction model considers the American Society of Anesthesiologists (ASA) physical status classification, patient profiles, and sets of surgical procedures variables. Scheduling optimization employs a lexicographic approach with three main objectives: minimizing patient waiting time, nurse overtime, and operating room idle time, while ensuring surgeon presence during critical phases and nurse availability according to shifts. The results show that the Catboost algorithm achieves the best prediction performance. Incorporating the ASA variable reduces prediction errors by 33.880 minutes in MAE and 55.575 minutes in RMSE compared to model without the ASA feature. The optimization model successfully eliminates all scheduling conflicts, ensuring full compliance with medical procedure constraints. Recovery bed utilization remains efficient, with a maximum of five units used, representing less than 50% of the total capacity.

YOLOv8-Based Quality Detection of Bali MSMEs Staple Food

Ni Putu Dita Ariani Sukma Dewi — 2025-12-30

Ensuring the quality of staple foods such as rice, cooking oil, milk, and meat is crucial for consumer safety and health. In Indonesian Micro, Small and Medium Enterprises (MSMEs), quality assessment often depends on subjective and time-consuming visual inspection. This study develops an automatic quality detection system using YOLOv8, applied to food MSMEs in Bali, to detect 14 quality categories across the four commodities based on image data. The methodology includes dataset collection from MSMEs, image annotation, preprocessing, training YOLOv8s and YOLOv8m models, and evaluating performance using mAP50, accuracy, precision, recall, and F1-score. Results show that YOLOv8m achieved a mAP50 of 96.5%, indicating high detection accuracy. The system, implemented as a web-based application, has strong potential to improve efficiency, ensure consistent product quality, and support Sustainable Development Goals (SDGs) 2, 3, 8, and 9.

Hybrid Machine Learning Approach for Nutrient Deficiency Detection in Lettuce

Zuriati Zuriati — 2025-12-30

Early detection of nutrient deficiencies in lettuce is essential for precision agriculture. However, this task remains challenging due to limited data availability and class imbalance, which reduce model sensitivity toward minority classes and hinder generalization. This study introduces a hybrid machine learning approach integrating SMOTE, Optuna, and SVM to enhance the accuracy of nutrient deficiency classification using digital leaf image analysis. The dataset, obtained from Kaggle, includes four categories: Nitrogen Deficiency (-N), Phosphorus Deficiency (-P), Potassium Deficiency (-K), and Fully Nutritional (FN). Image features were extracted using MobileNetV2 pretrained on ImageNet and classified with a Support Vector Machine. Three scenarios were tested: (1) SVM before SMOTE, (2) SVM after SMOTE, and (3) Optuna-SVM after SMOTE, evaluated using accuracy, precision, recall, and f1-score. The hybrid model achieved the best performance with accuracy 0.929, precision 0.946, recall 0.835, and f1-score 0.869, outperforming the other scenarios. This hybrid framework effectively addressed class imbalance and improved classification margin stability through adaptive hyperparameter tuning using the Tree Structured Parzen Estimator within Optuna. The novelty of this study lies in combining MobileNetV2 based feature extraction with SMOTE and Optuna-SVM for small agricultural datasets. The proposed approach offers an efficient, accurate, and practical solution for automated nutrient deficiency diagnosis and contributes to the development of AI-driven smart agriculture systems.