Height-Prompt Mixture of Experts for multimodal 3D building change detection

Rapid global urbanization and high-rise expansion demand effective 3D building change detection techniques based on remote sensing data for critical real-world applications, including urban monitoring, disaster assessment, and 3D city modeling. Recently, integrating optical imagery with digital surface models (DSMs) has emerged as an effective solution because DSMs provide explicit geometric information that complements RGB texture, enabling reliable detection of vertical structural variations, floor additions, and roof changes that are difficult to infer from optical imagery alone. However, existing methods extract DSMs based on the image domain processing paradigms, ignoring the potential of fine-grained elevation information to prompt the degree of change. To mitigate the issue, this paper proposes the Height-Prompt Mixture of Experts (HPMoE) network, which uses multiple expert groups for fine-grained feature fusion and decision-making. First, HPMoE extracts coherent structural and morphological cues from optical and DSM data to emphasize building boundaries. Second, we introduce a hierarchical routing MoE to dispatch features to specialized expert pools that include experts on smooth surfaces, instance building areas, or natural backgrounds. Finally, a fine-grained change-decision MoE is developed for multitask decision-making and adaptive optimization based on the degree of change. Extensive experiments on the Hi-BCD, 3DCD, and SMARS benchmark datasets demonstrate that HPMoE achieves superior performance in semantic and height metrics. The code is available at https://github.com/HaoLiu-XDU/HPMoE .