Please use this identifier to cite or link to this item: http://repository.ipb.ac.id/handle/123456789/164483
Title: Analisis Klasifikasi Citra Tingkat Kematangan Tandan Buah Segar (TBS) Kelapa Sawit Menggunakan EfficientNetV2B1
Other Titles: Image Classification Analysis of Oil Palm Fresh Fruit Bunch (FFB) Maturity Level Using EfficientNetV2B1
Authors: Oktarina, Sachnaz Desta
Soleh, Agus Mohamad
Adeva, Muhammad Farhan
Issue Date: 2025
Publisher: IPB University
Abstract: Indonesia, sebagai produsen utama minyak kelapa sawit, sangat bergantung pada kualitas Crude Palm Oil (CPO), yang dipengaruhi oleh tingkat kematangan buah. Penilaian manual kematangan tandan buah segar (TBS) sering kali tidak konsisten dan tidak efisien. Penelitian ini menerapkan transfer learning dengan EfficientNetV2B1 untuk mengklasifikasikan kematangan TBS ke dalam tiga kategori yaitu mentah, matang, dan terlalu matang. Gugus data berisi 690 gambar dari dua sumber smartphone dan kamera mirrorless. Data dibagi menjadi 60% pelatihan, 20% validasi, dan 20% pengujian. Sebanyak 12 model diuji dengan variasi batch size (16 dan 32), tiga optimizer (Adam, AdamW, Nadam), serta dua jenis input (gambar saja dan gambar dengan jenis kamera). Fine tuning dilakukan menggunakan fungsi aktivasi Swish dan Softmax. Hasil menunjukkan bahwa jenis kamera tidak signifikan terhadap performa. Model terbaik yaitu model 3 dengan menggunakan batch size 16 dan optimizer Nadam, dengan balanced accuracy 85,17%, F1-score 82,34%, recall 85,30%, dan presisi 80,60%. Pendekatan ini terbukti efektif dalam meningkatkan efisiensi dan akurasi klasifikasi kematangan TBS.
Indonesia, as a major producer of palm oil, relies heavily on the quality of Crude Palm Oil (CPO), which is influenced by fruit ripeness. Manual assessment of fresh fruit bunch (FFB) ripeness is often inconsistent and inefficient. This research applies transfer learning with EfficientNetV2B1 to classify FFB maturity into three categories: unripe, ripe, and overripe. The dataset contains 690 images from two sources of smartphone and mirrorless camera. The data is divided into 60% training, 20% validation, and 20% testing. A total of 12 models were tested with various batch sizes (16 and 32), three optimizers (Adam, AdamW, Nadam), and two types of inputs (image only and image with camera type). Fine tuning was performed using Swish and Softmax activation functions. Results showed that camera type is not significant to performance. The best model used a batch size of 16 and the Nadam optimizer, and achieved a balanced accuracy of 85.17%, F1-score of 82.34%, recall of 85.30%, and precision of 80.60%. This approach proved effective in improving the efficiency and accuracy of FFB ripeness classification.
URI: http://repository.ipb.ac.id/handle/123456789/164483
Appears in Collections:UT - Statistics and Data Sciences

Files in This Item:
File Description SizeFormat 
cover_G1401211062_553ab77cdeac4df39bb228bd508e7e40.pdfCover552.8 kBAdobe PDFView/Open
fulltext_G1401211062_f21a5971c6bd4346bf67cc54c2a230d2.pdf
  Restricted Access
Fulltext843.7 kBAdobe PDFView/Open
lampiran_G1401211062_0e24aecf2dd444d79b08a97b16a048e6.pdf
  Restricted Access
Lampiran267.76 kBAdobe PDFView/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.