Electronic Theses and Dissertation

Tomat merupakan salah satu komoditas hortikultura yang memiliki nilai ekonomi tinggi. Namun, proses pemilahan tingkat kematangan tomat masih banyak dilakukan secara manual sehingga hasilnya sering kali kurang konsisten. Penelitian ini bertujuan membandingkan performa MobileNetV2 dan Stacking Ensemble Learning dalam mengklasifikasikan citra tomat dalam tiga kelas, yaitu matang, setengah matang, dan rusak. Dataset yang digunakan terdiri dari 2.282 citra tomat yang berasal dari dataset publik Kaggle dan data tambahan yang dikumpulkan secara mandiri. Tahapan penelitian meliputi pra-pemrosesan citra, pengubahan ukuran menjadi 224x224 piksel, konversi citra ke ruang warna HSV, segmentasi objek, pembagian data dengan rasio 80:10:10, pelatihan model, serta evaluasi menggunakan metrik akurasi, presisi, recall, dan f1-score. MobileNetV2 dikembangkan menggunakan pendekatan transfer learning dan fine-tuning, sedangkan Stacking Ensemble Learning memanfaatkan fitur warna dan tekstur yang diekstraksi menggunakan HSV, GLCM, dan LBP dengan base models berupa Random Forest, Gradient Boosting, XGBoost, dan LightGBM, serta menggunakan Random Forest sebagaimeta model. Hasil pengujian menunjukkan bahwa MobileNetV2 memperoleh akurasi sebesar 87% sedangkan Stacking Ensemble Learning mencapai akurasi sebesar 91%. Selain itu, model Stacking Ensemble Learning juga menghasilkan nilai presisi, recall, dan f1-score yang lebih tinggi. Berdasarkan hasil tersebut, dapat disimpulkan bahwa Stacking Ensemble Learning lebih efektif dibandingkan MobileNetV2 dalam mengklasifikasikan tomat pada dataset yang digunakan.

Kata kunci : klasifikasi tomat, MobileNetV2, Stacking Ensemble Learning, pengolahan citra, machine learning, deep learning.

Tomato is one of the horticultural commodities with high economic value. However, the sorting process based on tomato maturity levels is still widely carried out manually, which often results in inconsistent classification outcomes. This study aims to compare the performance of MobileNetV2 and Stacking Ensemble Learning in classifying tomato images into three classes: ripe, half-ripe, and reject. The dataset used in this study consists of 2,282 tomato images obtained from a public Kaggle dataset and additional data collected independently. The research stages include image preprocessing, resizing images to 224 × 224 pixels, converting images into the HSV color space, object segmentation, data splitting with an 80:10:10 ratio, model training, and evaluation using accuracy, precision, recall, and F1-score metrics. MobileNetV2 was developed using transfer learning and fine-tuning approaches, while Stacking Ensemble Learning utilized color and texture features extracted using HSV, GLCM, and LBP, with Random Forest, Gradient Boosting, XGBoost, and LightGBM as the base models, and Random Forest as the meta model. The testing results show that MobileNetV2 achieved an accuracy of 87%, while Stacking Ensemble Learning achieved an accuracy of 91%. In addition, the Stacking Ensemble Learning model also produced higher precision, recall, and F1-score values. Based on these results, it can be concluded that Stacking Ensemble Learning is more effective than MobileNetV2 in classifying tomatoes using the dataset in this study. Keywords: tomato classification, MobileNetV2, Stacking Ensemble Learning, image processing, machine learning, deep learning.