Deteksi Objek dan Perhitungan Luas Area Polusi Visual Berbasis Deep Learning (YOLO)

Devi Hasugian, Ir. Yun Prihantina Mulyani, S.T., M.Sc., Ph.D, IPM., ASEAN Eng.

2026 | Tesis | S2 Teknik Industri

Polusi visual merupakan gangguan terhadap kualitas pemandangan, tata kota, dan estetika lingkungan yang disebabkan oleh faktor alam maupun aktivitas manusia. Dampaknya dapat meningkatkan stres dan kecemasan, menurunkan kualitas hidup, mengganggu konsentrasi, meningkatkan risiko kecelakaan, serta memengaruhi produktivitas di sektor manufaktur dan jasa. Salah satu penyebab utama munculnya polusi visual adalah kurangnya pengawasan terhadap elemen elemen visual di ruang publik.

YOLO (You Only Look Once) dipandang sebagai pendekatan yang potensial untuk membantu menyelesaikan permasalahan tersebut. Penelitian ini bertujuan membangun sebuah machine learning framework yang mampu mendeteksi elemen polusi visual sekaligus menghitung luas pixel polusi visual. Data yang digunakan dianotasi menggunakan metode instance segmentation, kemudian dilakukan proses augmentation untuk memperkaya variasi data. Selanjutnya, pemilihan model terbaik dilakukan melalui hyperparameter tuning menggunakan metode Grid Search. Kombinasi hyperparameter yang diuji meliputi Epoch, Learning Rate , dan Optimizer. Tiga model yang dibandingkan adalah YOLOv8s, YOLOv9c, dan YOLOv11s. Model terbaik dipilih berdasarkan nilai mAP50 tertinggi pada hasil testing. Model terpilih tersebut selanjutnya digunakan untuk melakukan prediksi luas pixel polusi visual dalam satuan pixel.

Berdasarkan hasil eksperimen, model terbaik adalah YOLOv9c dengan Epoch 60, Initial Learning Rate 0,0001 dan Optimizer SGD. Pada tahap pengujian, performansi model untuk mask menunjukkan nilai mAP50 83,5%, mAP50-95 60,1%, Precision 87,8%, Recall 76,7?n F1-Score 81,9%. Selain itu, pengujian perhitungan luas pixel menggunakan MAPE untuk kelas Network and Communication Tower yaitu 34,01%, Outdoor Advertisement and Signage 10,12%, Street Litter 8,38?n Wire 45,5?n MAPE secara keseluruhan adalah 24,50%.

Visual pollution refers to disturbances in the quality of scenery, urban layout, and environmental aesthetics caused by natural factors and human activities. Its impacts include increased stress and anxiety, reduced quality of life, 

decreased concentration, higher accident risk, and negative effects on productivity in both manufacturing and service sectors. One of the main causes of visual pollution is the lack of effective monitoring of visual elements in public spaces. 

YOLO (You Only Look Once) is considered a potential approach to address this problem. This study aims to develop a machine learning framework capable of detecting visual pollution elements while also calculating the pixel of visual 

pollution. The dataset was annotated using an instance segmentation method, followed by data augmentation to increase data diversity. Model selection was performed through hyperparameter tuning using the Grid Search method. The tested hyperparameter combinations included Epochs, Learning Rate , and Optimizer. Three models were compared: YOLOv8s, YOLOv9c, and YOLOv11s. The best model was selected based on the highest mAP50 value on the testing results. The selected model was then used to predict the visual pollution pixel in pixel units. 

Based on the experimental results, the best-performing model was YOLOv9c with 60 Epochs, an initial Learning Rate  of 0.0001, and the SGD Optimizer. During testing, the model achieved a mask performance of 83.5% mAP50, 60.1% mAP50–95, 87.8% precision, 76.7% recall, and an F1-score of 81.9%. In addition, the pixel calculation evaluation using MAPE resulted in 

34.01% for the Network and Communication Tower class, 10.12% for Outdoor Advertisement and Signage, 8.38% for Street Litter, and 45.5% for Wire and the overall MAPE was 24.50%. 

Kata Kunci : Visual Pollution, Image Processing, Computer Vision, YOLO, Instance Segmentation, Grid Search