跨任务交互的语义增强遥感图像高度估计

3D geo-information plays a crucial role in understanding the living environment. Due to the high costs associated with LiDAR-based data acquisition, monocular height estimation from optical remote sensing images has emerged as an attractive alternative. However, the distribution of height values in remote sensing images typically follows a long-tailed pattern, with low-height pixels (e.g., background) as the dominant category. As a result, trained networks tend to exhibit bias, often underestimating building heights, leading to significant instance-level height deviations. To address this issue, we propose a novel height estimation approach for remote sensing images, integrating semantic constraints and cross-task interactions. Specifically, a semantic segmentation module leveraging both local and global category representations is introduced to enhance the model\'s perception of height variations across different land cover classes. Additionally, a feature interaction module is employed to optimize the fusion of semantic and height information, mitigating height underestimation for high-rise buildings and other minority categories. Extensive experiments on the DFC2019 dataset demonstrate that the proposed method exhibits strong generalization capabilities and outperforms existing approaches in terms of accuracy and robustness.