LAUSR

ABSTRACT With the wide application of deep learning in visual perception, height estimation has become a research hotspot in remote sensing. A monocular height estimation method based on semantic information enhancement is proposed for satellite aremote sensing images. Specifically, the height estimation problem is modelled as a classification regression task, and semantic information is introduced to guide the adaptive division of height interval. The probability score of pixel points and the centre of the height interval are linearly combined to calculate the height, enabling the learning of the height of different regions. Firstly, a feature enhancement module based on semantic information is designed to make the feature extraction process more sensitive to semantic and scale information. Secondly, the local window attention mechanism is combined with the global spatial and channel attention mechanisms. This combination effectively improves the network’s attention allocation to different image regions, thereby maintaining high estimation accuracy in complex remote sensing scenarios. The experimental results show that the proposed method can fully use semantic information, significantly improving the accuracy and reliability of monocular height estimation.