UAV-based video multi-object tracking (MOT) is a significant task in the field of remote sensing. However, current research still faces critical issues: (1) the limitations of the single architecture of DNNs inherently hinder performance improvement of object detection; and (2) current linear modeling approaches for spatiotemporal relationships fail to capture complex motion patterns in the real world. To overcome the aforementioned issues, a hybrid architecture tracker (HA-Tracker) with a spatiotemporal Mamba motion model for UAV-based video MOT is the first to be proposed, which has the following innovations and contributions: (1) a CNN–Transformer–Mamba detector (CTM detector) is proposed to enhance the capability of object detection, which is a novel synergistic fusion framework for simultaneously fusing the local details of a CNN, the global context of a Transformer, and the long-range dependency of Mamba; and (2) a spatiotemporal Mamba motion model (STM³) is proposed to improve tracking accuracy by modeling the nonlinear spatiotemporal motion relationships of object trajectories. Extensive experimental results indicate that our HA-Tracker achieved outstanding performance, with multiple object tracking accuracy (MOTA) metrics of 44.76% and 52.22% and identity F1 scores (IDF1) of 60.33% and 72.34% on the Visdrone and UAVDT datasets, respectively. These results validate the effectiveness of HA-Tracker, which outperforms the existing MOT networks.

The Mamba network has recently emerged as a powerful architecture for remote sensing change detection (RSCD). Notably, the fusion of multi-scale features and heterogeneous attention mechanisms has proven effective in enhancing the performance of Mamba network. However, existing methods fail to quantitatively evaluate the actual contribution weights of different components during feature fusion, thereby limiting the interpretability of their improvements. To address this issue, this paper proposes a novel mixture-of-scale-and-attention-experts Mamba Network (MoEMamba) for the RSCD task, which has the following innovations and contributions: (1) a Siamese Mamba architecture is employed to extract features from bi-temporal images; (2) a mixture-of-scale-experts module is designed to enhance the generalization ability across multi-scale objects, which quantitatively evaluates the contribution of multiple scale-specific experts; and (3) a mixture-of-attention-experts module is proposed to improve the accuracy, which adaptively fuses complementary responses of heterogeneous attention mechanisms based on the learned weights. Extensive experiments on the WHU-CD and LEVIR-CD datasets demonstrate that MoEMamba outperforms existing state-of-the-art methods. Specifically, it achieves F1 scores of 94.20% and 92.62%, and IoU values of 89.04% and 86.25%, respectively. Importantly, this study provides valuable insights into the interpretability of multi-scale feature fusion and heterogeneous attention fusion for RSCD research.

Mamba has recently become a popular visual representation architecture for building change detection (BCD) in the remote sensing field. Although it effectively models global spatial dependencies, it still struggles to capture fine-grained local structures, thereby limiting its representation capability. To overcome this limitation, we propose ChangeMoE, a mixture-of-experts-based Mamba network for BCD from high-resolution remote sensing imagery. The main contributions and innovations are as follows: (1) A scan-specific mixture-of-experts (ScanMoE) module is proposed, which leverages a scan gating network to enhance global feature representations using eight selective scan mechanisms defined as scan-specific experts. We further quantify their contributions to improve the interpretability of global representation learning. (2) A global-and-local mixture-of-experts (GLMoE) module is proposed. We design the local and global expert networks to capture local details and global dependencies. A gating network is used to reveal their relative importance and adaptively fuse their complementary features. Experimental results on the LEVIR-CD and WHU-CD benchmarks validate the superiority of ChangeMoE over existing state-of-the-art networks. ChangeMoE achieves F1-scores of 92.55% and 94.11%, with IoU values of 86.13% and 88.88%, respectively. Compared with ChangeMamba as the baseline, the proposed ChangeMoE achieves improvements of 1.96% in F1-score and 3.33% in IoU on the LEVIR-CD dataset. These results demonstrate the effectiveness of the proposed ChangeMoE and provide an interpretable expert-fusion paradigm for understanding feature representation learning in BCD.