Contour-Enhanced Multi-Scale Height Estimation Network for Single Optical Remote Sensing Imagery

Under the circumstance of limited resolution in single-view optical imagery and the inherent complexity of structural variations, accurately predicting object heights to generate corresponding digital surface models (DSM) remains a highly challenging task. In this study, we propose a novel architecture termed the contour-enhanced multi-scale height estimation network (CEMS-HENet), which utilizes single-view optical imagery as input and outputs the corresponding DSM. The network initially employs an encoder-decoder U-Net architecture, augmented with a convolutional block attention module, to extract hierarchical features from the input imagery. We propose a contour enhancement module (CEM) to extract refined contour boundary features, which are subsequently fused with the U-Net output to enhance feature representation. Subsequently, a vision transformer (ViT) is leveraged for height estimation to capture the interactive relationships between global and local features at multiple scales. To address the issue of long-tailed distribution encountered in height estimation tasks, we integrate an adabins module, which adopts a classification-then-regression paradigm to model and adjust the height distribution of terrain and objects. Additionally, to mitigate the inaccuracies in height estimation along object boundaries, we design a novel joint loss function that combines contour enhancement with external contour constraints, significantly improving the estimation precision in boundary regions. Extensive experiments conducted on the DFC2019, Vaihingen, and Potsdam datasets validate the efficacy and robustness of the proposed model, demonstrating its superior performance in single-view optical imagery height estimation.