# UAV-Based Pothole Segmentation for Area Quantification: A Comparative Study of Mask RCNN and YOLOv11
> Ramadhani S.F.

URL kanonis: https://discover.unhas.ac.id/publications/pub_scopus_105041685354
Jurnal / Konferensi: Proceedings 2026 International Conference on Current Research in Artificial Intelligence and Data Science Iccraids 2026
Tahun terbit: 2026
DOI: https://doi.org/10.1109/ICCRAIDS67816.2026.11519665
Citations: 0

## Authors
- Ramadhani S.F.

## Abstract
Potholes are a persistent form of road surface damage that endangers driving safety and increases maintenance costs. This paper presents an intelligent UAVbased system for pothole detection, instance segmentation, and area measurement using a Ground Sampling Distance (GSD) approach. The system applies deep learning models (Mask RCNN and YOLOv11 segmentation) to predict pothole masks from UAV imagery and converts the mask pixel area into a realworld area through GSD-based scaling. A dataset of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{1, 1 0 0}$</tex> annotated images was collected at four altitudes (5, 10, 15, and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$20 ~\mathrm{m})$</tex> from two distinct road locations. To prevent evaluation leakage from correlated frames captured within the same UAV flight, we adopted a grouped split by flight session and report final performance on a held-out test set <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\boldsymbol{(} \mathbf{7 0 \%} \boldsymbol{/} \mathbf{2 0 \%} \boldsymbol{/} \mathbf{1 0 \%}$</tex> train/validation/test). We benchmarked four models: Mask RCNN with ResNet-50 and MobileNetV2 backbones, YOLOv11nseg and YOLOv11s-seg. On the held-out test set, Mask R-CNN with ResNet-50 achieves the best overall segmentation quality <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\text{mAP} {@} 0.50=0.956, \text{mIoU}=0.832$</tex>, Dice <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$=0.908)$</tex>, while YOLOv11n-seg provides a strong efficiency trade-off by running at <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4 5. 2 5}$</tex> FPS (approximately <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{4. 6 - 5. 9} \boldsymbol{\times}$</tex> faster than Mask R-CNN) with competitive accuracy (<tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{m A P} {@} 0.50=0.952$</tex>). YOLOv11s-seg obtains the best mAP@0.50:0.95 (0.585). For the GSD-based area estimation of two potholes, the RMSE values are 1,442.3 <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{cm}^{2}$</tex> and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$379.8 ~\text{cm}^{2}$</tex>, with average absolute percentage errors of 13.1 % and 19.2 %, respectively, which highlights that measurement reliability is more sensitive to segmentation consistency for smaller potholes.

## Keywords
- Artificial intelligence
- Pothole (geology)
- Segmentation
- Computer science
- Pattern recognition (psychology)
- Computer vision
- Image segmentation
- Feature (linguistics)
- Object detection
- Object (grammar)
- Feature extraction
- Noise (video)
- Artificial neural network

---
Sumber: Discover Unhas — RIMS Universitas Hasanuddin.
Saat mengutip, gunakan DOI bila tersedia atau URL kanonis di atas.