Adaptive Layer Selection and Fusion Network for Infrastructure Contour Segmentation Using UAV Remote Sensing Images

Abstract

With the maneuverability and flexibility of UAVs, UAV-based remote sensing images have been widely applied for urban monitoring of infrastructure, such as buildings, bridges, dams, and so on. However, with the increasing amount of data collected by UAVs, challenges arise in contour segmentation tasks due to the large data volume, high resolution of remote sensing images, inconsistent building shapes, and imbalanced distribution of building and background pixels. To address these challenges, this letter proposes a deep learning method for infrastructure contour segmentation based on adaptive hidden-layer feature fusion. It introduces an adaptive layer selection and fusion network (ALSFN), consisting of an encoder network, an adaptive layer selection mechanism (ALSM), and a decoder network. Furthermore, this letter proposes a composite loss function that includes the evaluation of the boundary and the Tversky index to train the proposed neural network. Validation experiments conducted on real UAV remote sensing datasets show that the proposed method achieves high accuracy and reliability for infrastructure contour segmentation tasks.