<p>Accurate and interpretable multi-target recognition in complex remote-sensing landscapes remains a significant challenge due to spectral heterogeneity, multi-scale object variation, spatial overlap and limited model transparency. Although convolutional and transformer-based approaches have advanced landscape analysis, they often struggle to simultaneously achieve high detection accuracy, long-range contextual reasoning, computational efficiency and explainability. To overcome these limitations, this paper proposes a Hybrid Swin Transformer–MAMBA–DETR framework for hierarchical multi-target landscape understanding. The Swin Transformer backbone extracts multi-scale spatial representations through shifted window self-attention, effectively modeling local and hierarchical dependencies. To enhance global contextual consistency, a MAMBA state-space module is integrated to capture long-range sequential relationships with linear computational complexity, reducing memory overhead while preserving feature continuity. A DETR-based detection head enables end-to-end, anchor-free object recognition, eliminating heuristic region proposals and improving convergence stability. Furthermore, a Grad-CAM-based explainability module generates class-specific activation maps, enhancing interpretability and supporting transparent decision analysis. Comprehensive experiments on the LoveDA dataset, complemented by cross-domain evaluations on MAVSD and FAIR1M, demonstrate that the proposed framework achieves 86.53% mIoU and 91.72% mAP on LoveDA, outperforming strong CNN and transformer-based baselines. The model also shows strong robustness under illumination variation, occlusion and domain shift, with less than 2% performance degradation. These results indicate that the proposed approach offers an effective, efficient and interpretable solution for large-scale intelligent landscape monitoring.</p>