Unmanned Aerial Vehicle (UAV)-Based Identification of Pipeline High Consequence Areas (HCAs) in Few-Shot Scenarios: A Phased Transfer Learning Method

High Consequence Area (HCA) identification is fundamental to pipeline integrity management because building distribution and population conditions along pipeline corridors directly shape consequence assessment and riskinformed decisions. In practical unmanned aerial vehicle (UAV) inspections, however, labeled samples are often scarce and corridor environments are highly heterogeneous, making building extraction prone to false positives and missed buildings and thereby weakening the reliability of automated HCA identification. To address this challenge, this study develops a phased transfer learning framework for building-footprint extraction from UAV imagery under few-shot conditions. Built on TransUNet, the proposed method integrates CNN-based local feature extraction with Transformer-based global context modeling and adapts the model to corridor scenes through staged transfer learning with limited annotations. Under spatially isolated five-fold cross-validation, the proposed approach achieves a Dice score of 82.7 %, an intersection over union (IoU) of 79.4 %, and a false positive rate of 1.2 %. Compared with the widely used engineering baseline DeepLabV3+, Dice and IoU increase by 9.4 % and 10.1 %, respectively, while the false positive rate decreases by 29.4 %. Two engineering cases further show that the extracted building inventories support reliable building counts, population estimation, and HCA grading, thereby providing a reliable basis for periodic HCA updating and quantitative risk assessment in pipeline safety management.

Keywords: High Consequence Areas, Few-Shot Learning, Transfer Learning, TransUNet, UAV Remote Sensing, Pipeline Integrity Management.