Klasifikasi Jenis Sampah Menggunakan Transfer Learning Berbasis EfficientNetV2-B0 dengan Strategi Fine-Tuning

Abstract

Manual waste sorting often leads to inefficient waste management and classification errors. This study implements the EfficientNetV2-B0 architecture using transfer learning and a fine-tuning strategy to classify organic and recyclable waste images from the Waste Classification Data dataset, which contains two classes. The dataset was split into 80% training, 10% validation, and 10% testing data. The methodology included image preprocessing through resizing, float32 data type conversion, and the use of EfficientNetV2-B0 internal preprocessing, followed by data augmentation, transfer learning, and fine-tuning of the last 30% of model layers using the Adam optimizer. Model performance was evaluated using accuracy, precision, recall, and F1-score metrics under two batch size settings, 16 and 32. The results showed that a batch size of 32 achieved the best performance, with 96.45% accuracy, 96.20% precision, 95.77% recall, and 95.99% F1-score. These findings demonstrate the effectiveness of EfficientNetV2-B0 for waste image classification and its potential for automated waste sorting systems in various real-world environmental management applications.

Keywords: Efficientnetv2-B0; Waste Classification; Transfer Learning; Fine-Tuning; Deep Learning

Abstrak

Pemilahan sampah yang masih dilakukan secara manual sering menyebabkan proses pengelolaan sampah menjadi tidak efisien dan rentan terhadap kesalahan klasifikasi. Penelitian ini bertujuan mengimplementasikan arsitektur EfficientNetV2-B0 berbasis transfer learning dengan strategi fine-tuning untuk mengklasifikasikan citra sampah organik dan recyclable menggunakan dataset Waste Classification Data yang terdiri atas dua kelas. Dataset dibagi menjadi 80% data pelatihan, 10% data validasi, dan 10% data pengujian. Tahapan penelitian dilakukan melalui preprocessing berupa resizing, konversi tipe data, dan pemanfaatan preprocessing internal EfficientNetV2-B0, augmentasi data, transfer learning, serta fine-tuning pada 30% lapisan terakhir model dengan menggunakan optimizer Adam. Evaluasi kinerja dilakukan menggunakan metrik accuracy, precision, recall, dan F1-score dengan membandingkan dua skenario batch size, yaitu 16 dan 32. Hasil penelitian menunjukkan bahwa batch size 32 mampu menghasilkan performa terbaik dengan accuracy 96,45%, precision 96,20%, recall 95,77%, dan F1-score 95,99%. Hasil ini menunjukkan bahwa EfficientNetV2-B0 dengan strategi fine-tuning efektif dalam mengklasifikasikan citra sampah serta berpotensi mendukung pengembangan sistem pemilahan sampah otomatis.