Toward intelligent pavement maintenance: a transferable deep learning framework for cross-domain crack segmentation and UAV-based field inspection

Intelligent pavement condition assessment is fundamental to infrastructure asset management and data-driven maintenance decision-making. However, deploying deep learning models in real-world inspection scenarios remains challenging due to the significant cross-domain distribution shift between laboratory training data and field conditions characterized by varying illumination, complex backgrounds, and heterogeneous pavement textures. This study presents a systematic framework for pavement crack semantic segmentation that bridges the gap between model development and engineering deployment. Through comprehensive benchmarking of CNN and Transformer architectures on a hybrid source domain dataset, the optimal base model configuration balancing segmentation accuracy and computational efficiency is identified. To address the inherent class imbalance in crack imagery, loss function hyperparameters are optimized via grid search, enhancing the model’s sensitivity to fine and sparse crack features. A multi-stage transfer learning framework comprising three complementary strategies is then proposed, combining general visual feature initialization with source domain knowledge to enable effective cross-domain adaptation. Systematic evaluation across four target domain datasets of varying scales (250 to 2000 images) demonstrates that the proposed framework consistently outperforms the non-transfer control group, general segmentation models, and advanced crack-specific methods across all dataset scales, exhibiting robust performance under challenging real-world conditions. Field validation through UAV-based pavement inspection further confirms the framework’s operational effectiveness, demonstrating stable crack detection and accurate width quantification across diverse pavement types under complex aerial imaging conditions. The proposed framework offers an end-to-end solution that formalizes deep learning-based crack segmentation into a transferable, domain-adaptive methodology, providing new insights into how computational methods can be generalized and operationalized to construct engineering knowledge for intelligent infrastructure maintenance.