An informatics workflow for label-scarce cross-domain segmentation of geological outcrop fractures from UAV photogrammetric imagery

Abstract Accurate extraction of geological outcrop fracture networks from UAV photogrammetric imagery supports rock-mass characterization, discontinuity mapping, and geotechnical interpretation. However, automated segmentation remains challenging because fractures are often thin, low-contrast, discontinuous, and visually confused with heterogeneous textures, weathering patterns, and shadows. Pixel-level annotation is also labor-intensive, resulting in limited labeled data for newly surveyed areas. This study develops a label-scarce cross-domain segmentation workflow using GeoCrack as the source domain and RockCrack, a small, site-specific UAV dataset, as the target domain. GRNet, the segmentation model used in the workflow, combines Deformable Large-Kernel Attention at the bottleneck with Spatial-Channel Squeeze-and-Excitation in the encoder. Feature-space analysis identifies a measurable discrepancy between the two domains. On GeoCrack, GRNet achieves an IoU of 66.84% and an F1-score of 80.12%. When trained from scratch on RockCrack, it achieves a five-fold mean crack-class IoU of 56.69 ± 4.34% and an F1-score of 72.28 ± 3.61%. Following GeoCrack pretraining and target-domain fine-tuning, the best-performing configuration achieves a mean IoU of 58.96 ± 3.64% and an F1-score of 74.13 ± 2.93%. Fine-tuning substantially outperforms direct source-only deployment, although differences between the leading strategies are small and fold-dependent. The workflow provides pixel-level fracture maps that can support subsequent skeletonization, trace vectorization, and fracture-network characterization; however, further spatial calibration, uncertainty assessment, and geological validation are required before the outputs can be used for quantitative rock-mass analysis or engineering decisions.