Abstract
Road deterioration is driven by factors such as heavy traffic, adverse weather
conditions, overloaded vehicles, inadequate maintenance, and natural disasters,
with cracks, bumps, and potholes being the most common defects.
These issues pose safety risks and raise maintenance costs. Aerial imagery
from drones, UAVs, and aircraft provides a wide-area perspective, enabling
efficient monitoring. Advances in deep learning, computer vision, and large
datasets have improved object detection in such imagery, offering faster and
more scalable updates than ground-based or crowd-sourced methods. This
study applies MobileNetV2, a lightweight convolutional neural network, to
detect road surface defects in aerial imagery. The problem was redefined as
binary classification, and five refined MobileNetV2 models were developed
using techniques such as layer unfreezing, class weighting, threshold tuning,
and data augmentation. The best model achieved 81.52% accuracy, showing
that lightweight transfer models can deliver effective, deployable solutions
for large-scale road defect detection.