Decision-oriented multi-altitude UAV-based deep learning framework for railway track and ballast anomaly screening

This paper presents a UAV-based deep learning framework for automated railway track and ballast screening using multi-altitude RGB imagery acquired under heterogeneous operational conditions. The proposed classification framework supports scalable inspection and maintenance workflows while reducing annotation effort and manual assessment. Four operational states—normal condition, gauge anomaly, ballast degradation, and vegetation excess—are considered, whereas pixel-wise defect localization is avoided to improve computational efficiency, scalability, and robustness across varying acquisition geometries and flight altitudes. RGB images acquired at 3 m, 40 m, and 60 m are integrated through an augmentation strategy combining geometric, photometric, and scale-consistent transformations that preserve engineering plausibility and realistic UAV variability. The dataset is organized into composite multi-altitude sub-datasets and evaluated using leakage-safe segment- and session-independent partitioning protocols. Three convolutional backbones are analyzed within a unified transfer-learning framework optimized for limited and heterogeneous annotated data. Experimental results show stable performance across multi-altitude scenarios, with Macro-F1 scores up to 0.87 and ROC-AUC values close to 0.97. Ablation analyses confirm that the proposed augmentation and multi-altitude learning improve cross-scale robustness, training stability, and anomaly sensitivity compared with conventional augmentation and single-altitude configurations, supporting the framework as a scalable front-end screening stage for hierarchical railway inspection pipelines.