Hie-DinoMamba: Hierarchical DINOv3 and Mamba Architecture for Multi-Class Building Change Detection

Abstract. Accurate multi-class building change detection in high-resolution aerial imagery is a critical task for urban analysis. However, it is hindered by two key challenges: severe class imbalance and the difficulty of obtaining robust, generalizable feature representations. While recent models show promise, encoders trained from scratch on aerial data remain limited in their representational capacity. Leveraging large-scale Visual Foundation Models (VFMs) offers a path to better features, but full fine-tuning is computationally prohibitive. To address this, we propose Hie-DinoMamba, a novel hierarchical architecture. We integrate a frozen 1.1B parameter DINOv3-L (SAT-493M) encoder, preserving its rich pre-trained knowledge. We efficiently adapt this encoder to the aerial domain using parameter-efficient Low-Rank Adaptation (LoRA). Furthermore, we design a new Hierarchical Mamba FPN decoder that uses Visual State Space Model (VSSM, Mamba) blocks to fuse and refine multi-scale feature pairs in a top-down manner. The model is optimized using a dual-loss strategy (Semantic and Boundary) to ensure both classification accuracy and precise boundary delineation. On the 4-class aerial building change detection benchmark, Hie-DinoMamba achieves state-of-the-art performance with an mIoU of 65.12%, a significant improvement of 2.1 percentage points over the strong VSSM-based baseline (ChangeMamba-MC). Qualitative analysis further demonstrates our model’s superior generalization, successfully detecting complex changes in unseen geographic regions where other models fail.