Laporkan Masalah

Klasifikasi dan Deteksi Sampah Multi Kelas Menggunakan Arsitektur YOLO dalam Proses Sortasi

Stefani Kumala Dewi Wijayanti, Ir. Achmad Pratama Rifai, S.T., M.Eng., Ph.D.

2026 | Tesis | S2 Teknik Industri

Sampah padat perkotaan terus meningkat seiring pertumbuhan penduduk dan aktivitas ekonomi, sementara proses pemilahannya di Indonesia masih banyak bergantung pada tenaga manual yang berisiko terhadap kesehatan pekerja dan kurang efisien. Penelitian ini berfokus pada bagaimana perbandingan performa YOLOv7, YOLOv8, dan YOLOv11 dalam mendeteksi dan mengklasifikasikan jenis sampah secara otomatis, guna memudahkan proses pemilahan sampah untuk daur ulang.

Penelitian menggunakan dataset sebanyak 21.791 gambar yang terdiri dari kombinasi data sampah pribadi, Tempat Pengolahan Sampah (TPS), dan dataset publik, mencakup 20 kelas sampah dengan total 19.394 objek hasil anotasi bounding box menggunakan Roboflow. Dataset dibagi menjadi data training (70%) sebanyak 15.253 gambar, testing (15%) sebanyak 3.270 gambar, dan validation (15%) sebanyak 3.268 gambar, serta diperbanyak dengan teknik augmentasi flip, crop, blur, dan noise. Ketiga model yaitu YOLOv7, YOLOv8, dan YOLOv11, dilatih dengan lima variasi epoch dengan jumlah epoch 50, 75, 100, 125, dan 150 serta dievaluasi menggunakan metrik precision, recall, F1-score, mAP@0.5, dan waktu inferensi.

Hasil penelitian menunjukkan performa terbaik pada seluruh model dicapai di 150 epoch. YOLOv7 menghasilkan mAP@0.5 sebesar 0,878, F1-score 0,843, dan accuracy 69,47%. YOLOv8 meningkatkan performa menjadi mAP@0.5 0,929, F1-score 0,892, dan accuracy 76,93%. YOLOv11 menjadi model terbaik dengan mAP@0.5 0,932, F1-score 0,897, accuracy 79,62%, serta waktu inferensi tercepat yaitu 5,5 ms dengan 181 FPS, melebihi YOLOv8 yaitu 5,56 ms dengan 179 FPS dan YOLOv7 yaitu 10 ms dengan 100 FPS. Dengan demikian, YOLOv11 pada 150 epoch merupakan model paling optimal untuk diterapkan dalam sistem deteksi dan klasifikasi sampah otomatis pada proses sortasi.

Municipal solid waste continues to increase as a result of population growth and economic activities, while waste sorting in Indonesia still largely relies on manual labor, which poses health risks to workers and operates with low efficiency. This study compares the performance of YOLOv7, YOLOv8, and YOLOv11 in automatically detecting and classifying different types of waste to support the waste sorting process for recycling.

The study used a dataset of 21,791 images consisting of a combination of self collected waste images, images from waste processing facilities, and publicly available datasets. The dataset included 20 waste classes with a total of 19,394 annotated objects using bounding boxes created in Roboflow. The dataset was divided into 70 percent for training with 15,253 images, 15 percent for testing with 3,270 images, and 15 percent for validation with 3,268 images. Data augmentation techniques including flipping, cropping, blurring, and noise addition were applied to increase data diversity. The three models, namely YOLOv7, YOLOv8, and YOLOv11, were trained using five different numbers of epochs consisting of 50, 75, 100, 125, and 150 epochs. Model performance was evaluated using precision, recall, F1 score, mAP at 0.5, and inference time.

The experimental results show that all models achieved their best performance after 150 epochs. YOLOv7 achieved an mAP at 0.5 of 0.878, an F1 score of 0.843, and an accuracy of 69.47 percent. YOLOv8 improved the results with an mAP at 0.5 of 0.929, an F1 score of 0.892, and an accuracy of 76.93 percent. YOLOv11 achieved the best overall performance with an mAP at 0.5 of 0.932, an F1 score of 0.897, an accuracy of 79.62 percent, and the fastest inference time of 5.5 milliseconds with 181 frames per second. This performance exceeded YOLOv8, which recorded 5.56 milliseconds with 179 frames per second, and YOLOv7, which recorded 10 milliseconds with 100 frames per second. Therefore, YOLOv11 trained for 150 epochs is the most suitable model for implementation in automatic waste detection and classification systems to support waste sorting processes.

Kata Kunci : multi class waste, waste detection, deep learning, YOLO

  1. S2-2026-530109-abstract.pdf  
  2. S2-2026-530109-bibliography.pdf  
  3. S2-2026-530109-tableofcontent.pdf  
  4. S2-2026-530109-title.pdf