An Infrastructure Segmentation Neural Network for UAV Remote Sensing Images

Infrastructure segmentation (bridges, dams, etc.) based on remote sensing images has been widely used in urban planning, land management, disaster monitoring and other monitoring fields. As a mature remote sensing monitoring platform, unmanned aerial vehicles (UA V) have become the main source of low-altitude remote sensing image acquisition of infrastructures. However, with the increasing data acquired by UAVs, the large data volume and high resolution of remote sensing images lead to challenges for the task of infrastructure segmentation. To solve the difficulties, this paper proposes a deep learning method for infrastructure segmentation based on adaptive hidden layer feature fusion, which specifically contains an encoder network, an adaptive layer selection mechanism, and a decoder network. In addition, this paper proposes a composite loss function containing two evaluation metrics, namely, boundary evaluation loss and Tversky loss, for training the proposed neural network. The validation of this paper is carried out on a real-world dataset, and the experimental results demonstrate the high accuracy and reliability of the proposed method on the infrastructure segmentation task.