MAC-DETR: Re-parameterized Mambaout for robust object detection in high-resolution remote sensing
Huanxu Li, Keke Xu, Xianglei Liu, Jianlong Wang, Guoqing Zhou · International Journal of Applied Earth Observation and Geoinformation · 2026
Accurate and efficient object detection in remote sensing imagery is fundamental to applications such as urban monitoring, infrastructure assessment, and disaster management. However, two persistent challenges remain: (1) small objects occupy only a few pixels in ultra-high-resolution images, making them difficult to detect reliably, and (2) the rapid growth of large-scale satellite data demands computationally efficient methods suitable for practical deployment. Existing lightweight convolutional networks provide real-time efficiency but often fail to capture sufficient discriminative information for small targets, while Transformer-based detectors improve representation power at the cost of high computational complexity. To address these challenges, we propose MAC-DETR, a CNN-Mambaout-Transformer hybrid detection framework designed for remote sensing imagery. MAC-DETR integrates three complementary modules: (1) Mambaout-RepLK, a Mambaout-style gated convolution block that reduces computational overhead while preserving expressive feature extraction, achieving + 2.5% mAP improvement in fine-grained categories on NWPU VHR-10 while reducing parameters by 23%; (2) CROSS-UP, an adaptive upsampling block that enhances multi-scale fusion and improves the detection of small objects—yielding up to + 2.6% mAP gain on multi-scale categories in NWPU VHR-10—without introducing extra complexity; and (3) ASSAIFI, a sparse attention module that strengthens feature interaction to further refine small-object representations. Experiments on four benchmark datasets demonstrate that MAC-DETR achieves 74.1% mAP on DOTA-v1.5, 95.4% on NWPU VHR-10, 94.4% on HRSC2016, and 95.4% on the infrared SIRSTv2 dataset, consistently outperforming both CNN– and Transformer-based baselines. Ablation studies show that the design reduces parameters by 23% and computation by 10–15%, offering a practical balance between accuracy and efficiency. These results highlight MAC-DETR’s effectiveness for large-scale, multi-modal Earth observation applications. The source code will be available at: https://github.com/Keykeykeykeykeykey/MAC-DETR.