Implementasi Model MobileNetV2 dalam Pengembangan Sistem Prediksi Kematangan Pisang Berbasis Mobile

Abstract

Bananas are a highly nutritious tropical fruit widely consumed by the public. However, the manual process of determining ripeness levels often leads to inconsistencies. Pak Sanali’s Banana Plantation in Krembung Subdistrict, Sidoarjo Regency, faces challenges in accurately determining fruit ripeness, which can potentially cause sorting errors that impact the quality and market value of the harvest. This study aims to design an image-based banana ripeness prediction system using a Deep Learning approach. The model used is a Convolutional Neural Network (CNN) with a MobileNetV2 architecture that employs transfer learning. The process begins with banana image inputs that undergo preprocessing stages of augmentation and normalization, followed by feature extraction through convolutional layers to capture visual characteristics such as the color and texture of the banana peel. These features are then processed in the classification head layer, consisting of Global Average Pooling and a fully connected layer, to generate predictions for four ripeness classes: unripe, ripe, overripe, and rotten. Test results show that the model using the Adam optimizer delivers the best performance, with an accuracy of 99.47% and a test loss of 0.39%. The model was developed using Python and TensorFlow on Google Colaboratory and implemented in a Kotlin-based application. Evaluation using a confusion matrix demonstrates excellent classification performance based on the metrics of accuracy, precision, recall, and F1-score.