MAR-YOLO: multi-scale feature adaptive selection and asymptotic pyramid for oriented Building detection in remote sensing images Authors: Yuzhe Zhao, Haizhong Qian Published: 2025 ABSTRACT AbstractBuilding detection in remote sensing imagery confronts three interdependent challenges including extreme scale variance under dense spatial distributions, orientation instability in off-nadir imagery, and semantic gaps during multi-scale feature fusion. Existing methods address these challenges in isolation, resulting in performance degradation when challenges co-occur. MAR-YOLO establishes an integrated rotated object detection framework extending YOLOv11-OBB through three synergistic innovations. The Multi-scale Feature Adaptive Selection (MFAS) module adaptively filters P2-P5 features through dual- domain weighting, enhancing small building perception while suppressing redundancy. The adapted Adaptive Feature Pyramid Network (AFPN) employs progressive fusion with scale-matched kernels and learned spatial weights, eliminating semantic inconsistencies inherent in direct multi-scale concatenation. The RepVGG-based Enhanced Rotated Detection Head (RRD-Head) applies branch-specialized structural reparameterization addressing angle regression instability. Validation on BONAI demonstrates 87.2% mAP50 and 65.3% mAP50-95, representing 2.9% and 2.6% improvements over YOLOv11s-OBB at 95 FPS. Cross-dataset experiments on DOTA, DIOR-R, and HRSC2016 confirm architectural robustness across diverse detection scenarios.Keywords: Remote sensing images, Building detection, Rotated object detection, YOLOv11, Multi-scale feature fusion, ReparameterizationSubject terms: Engineering, Mathematics and computing Introduction With the rapid development of remote sensing technology, high-resolution remote sensing images have been widely applied in fields such as geographic information systems, environmental monitoring, urban planning, and disaster assessment1,2. As a core task in remote sensing image interpretation, object detection aims to accurately identify and locate specific targets in images, providing fundamental data support for subsequent information extraction and analysis3,4. Among these applications, building detection, as an important component of remote sensing image processing, holds significant importance for urban change monitoring, land use analysis, and in-frastructure management. Deep learning object detection algorithms are mainly divided into two categories: two-stage and single-stage methods. Two-stage detection algorithms such as Faster R-CNN5 and Mask R-CNN6 achieve high-precision detection through candidate region generation and refined classification and localization. However, these methods have high computational complexity and relatively slow inference speed. In contrast, single-stage detection algorithms such as the YOLO series7 and SSD8 transform object detection into a unified regression problem, directly predicting object categories and bounding box coordinates, significantly improving detection efficiency while maintaining high accuracy. Particularly, the YOLO series algorithms have gained widespread attention and application in remote sensing image object detection due to their excellent real-time performance and detection accuracy. Remote sensing building detection confronts three interconnected challenges that existing methods address only in isolation. Small object detectors experience performance degradation when extreme scale variance (ranging from 8 to 128 pixels) co-occurs with dense spatial distributions where inter-building spacing approaches or falls below network receptive field dimensions9,10. Oriented detection frameworks achieve rotational regression capabilities yet demonstrate instability when processing irregular building geometries and perspective distortions characteristic of off-nadir aerial imagery11. Feature preservation architectures alleviate downsampling-induced information loss but introduce semantic inconsistencies during non-adjacent pyramid level fusion, with such discontinuities becoming more pronounced upon incorporating fine-grained detection layers. While prior work addresses scale variance, orientation, and feature preservation independently, MAR- YOLO targets their interdependent co-occurrence through integrated architectural design rather than component-wise optimization. Addressing these challenges requires an integrated architectural approach rather than isolated component improvements. MAR-YOLO establishes a building-oriented detection framework based on YOLOv11-OBB12, where multi-branch aggregation networks preserve target information throughout the feature extraction hierarchy. The architecture combines dynamic interpolation with multi-feature fusion reconstruction within the feature pyramid, thereby strengthening multi-scale representation capabili-ties while enriching spatial semantic information across pyramid levels. Large-depth separable kernel modules expand the effective receptive field through contextual in-formation utilization, consequently enhancing small target detection performance un-der dense distribution scenarios. Additionally, the integration of NWD loss function13 improves the model’s target recognition capability in complex backgrounds. Experi-mental results demonstrate that these optimization strategies significantly improve the performance of remote sensing image object detection. The main contributions of this paper are as follows: Designed a Multi-scale Feature Adaptive Selection (MFAS) module addressing feature redundancy arising from P2 layer integration—a limitation unaddressed by standard attention mechanisms operating solely in spatial or channel domains. Unlike FocusedLinearAttention’s14 uniform spatial filtering across pyramid levels, MFAS introduces scale-specific learnable parameters enabling differentiated feature importance learning at P2, P3, and P4 layers, with dual-domain weighting achieving 0.6% mAP50 improvement over single-domain attention baseline in ablation studies. Designed a Multi-scale Feature Adaptive Selection (MFAS) module addressing feature redundancy arising from P2 layer integration—a limitation unaddressed by standard attention mechanisms operating solely in spatial or channel domains. Unlike FocusedLinearAttention’s14 uniform spatial filtering across pyramid levels, MFAS introduces scale-specific learnable parameters enabling differentiated feature importance learning at P2, P3, and P4 layers, with dual-domain weighting achieving 0.6% mAP50 improvement over single-domain attention baseline in ablation studies. Proposed building-oriented adaptations to Asymptotic AFPN’s12 progressive fusion framework addressing semantic gap challenges in dense small-target scenarios. Beyond the original design’s uniform 3 × 3 convolutions, two architectural modifications target building-specific characteristics: scale-matched convolution kernels (3 × 3 preserving small building details in P2-P3 fusion, 5 × 5 capturing elongated structures in P4-P5 integration) and learned spatial fusion weights replacing fixed averaging to handle density variations. Comparative analysis with original AFPN implementation shows 0.4% mAP50 gain attributed to these adaptations. Proposed building-oriented adaptations to Asymptotic AFPN’s12 progressive fusion framework addressing semantic gap challenges in dense small-target scenarios. Beyond the original design’s uniform 3 × 3 convolutions, two architectural modifications target building-specific characteristics: scale-matched convolution kernels (3 × 3 preserving small building details in P2-P3 fusion, 5 × 5 capturing elongated structures in P4-P5 integration) and learned spatial fusion weights replacing fixed averaging to handle density variations. Comparative analysis with original AFPN implementation shows 0.4% mAP50 gain attributed to these adaptations. Constructed a RepVGG15 reparameterization-based Enhanced Rotated Detection Head (RRD-Head) establishing explicit branch role specialization for angle regression—differentiating from standard RepVGG’s uniform branch configurations. While conventional RepVGG applications employ identical 3 × 3, 1 × 1, and identity branches across all tasks, RRD-Head assigns distinct responsibilities: 3 × 3 branches extract edge orientation features addressing perspective distortion, 1 × 1 branches model inter-channel angle correlations, and identity branches stabilize gradient flow preventing orientation ambiguity during backpropagation. This role-differentiated design yields 0.7% mAP50 improvement over direct RepVGG application in ablation studies. Constructed a RepVGG15 reparameterization-based Enhanced Rotated Detection Head (RRD-Head) establishing explicit branch role specialization for angle regression—differentiating from standard RepVGG’s uniform branch configurations. While conventional RepVGG applications employ identical 3 × 3, 1 × 1, and identity branches across all tasks, RRD-Head assigns distinct responsibilities: 3 × 3 branches extract edge orientation features addressing perspective distortion, 1 × 1 branches model inter-channel angle correlations, and identity branches stabilize gradient flow preventing orientation ambiguity during backpropagation. This role-differentiated design yields 0.7% mAP50 improvement over direct RepVGG application in ablation studies. Conducted comprehensive experimental validation on the BONAI dataset16, where MAR-YOLO achieved improvements of 2.9% and 2.6% in mAP50 and mAP50-95 respectively compared to the original YOLOv11s-OBB. Additionally, generalization experiments on DOTA, DIOR-R, and HRSC2016 datasets demon-strated the robustness and effectiveness of the method. Conducted comprehensive experimental validation on the BONAI dataset16, where MAR-YOLO achieved improvements of 2.9% and 2.6% in mAP50 and mAP50-95 respectively compared to the original YOLOv11s-OBB. Additionally, generalization experiments on DOTA, DIOR-R, and HRSC2016 datasets demon-strated the robustness and effectiveness of the method. The remainder of this paper is organized as follows: “Related work” reviews related work; “Materials and methods” details the proposed method; “Results” validates the effectiveness of the method through ablation experiments, comparative experiments, and visualization analysis; “Discussion” summarizes the research work and discusses future development directions. Related work Object detection algorithms can be divided into two major categories based on detection pipeline: two-stage and single-stage methods. Two-stage algorithms such as Faster R-CNN and Mask R-CNN achieve high-precision detection through candidate region generation and refined classification regression5,6. Single-stage algorithms such as the YOLO series and SSD series transform object detection into a regression problem, achieving fast detection through end-to-end network structures7,8. Two-stage rotated object detection Rotated object detection extends conventional horizontal bounding box frameworks through angle parameter integration, enabling precise target localization under arbitrary orientations. Two-stage architectures pioneered this direction, with Rotated Faster R-CNN introducing rotated proposal generation within the Region Proposal Network while maintaining the R-CNN refinement pipeline17. Despite achieving high accuracy, computational complexity and inference latency constrain real-time deployment. Oriented R-CNN enhanced prediction precision via rotated RoI operations addressing feature-box misalignment18, whereas S2ANet resolved alignment challenges through feature refinement combined with single-shot anchor transformation19. R3Det advanced angle regression stability by implementing progressive bounding box refinement across multiple feature hierarchy levels20. These two-stage methods establish strong accuracy baselines yet require architectural simplification for efficiency-critical remote sensing applications. Single-stage YOLO-based rotated detection Single-stage detector evolution motivated YOLO framework adaptations for rotated object detection. YOLOv5s-OBB and YOLOv8s-OBB architectures incorporate angle prediction branches within detection heads, achieving rotational regression while preserving computational efficiency21. These adaptations maintain real-time inference capabilities essential for large-scale remote sensing analysis. Nevertheless, angle prediction accuracy degrades substantially in dense small building scenarios, exhibiting systematic localization biases attributable to detection head architectural limitations and loss function formulations inadequate for handling orientation ambiguity under scale variance. Feature fusion strategies for multi-scale detection Multi-scale feature fusion constitutes a fundamental challenge in remote sensing object detection. Feature Pyramid Networks (FPN) established the paradigm of top-down semantic propagation, enabling cross-scale feature integration through lateral connections22. Path Aggregation Network (PANet) augmented this architecture with bottom-up pathways, strengthening bidirectional information flow23. Subsequent variants including NAS-FPN24 and BiFPN25 introduced learnable fusion weights and architecture search strategies, improving computational efficiency. Despite these advances, direct cross-scale fusion introduces semantic inconsistencies when combining features from non-adjacent pyramid levels—high-level features encode abstract semantics with coarse spatial resolution, whereas low-level features preserve fine spatial details but lack semantic discriminability. This semantic gap manifests as increased false positives and missed detections for small targets, particularly when dense building distributions require simultaneous preservation of fine-grained localization and contextual reasoning capabilities. SPD-Conv effectively reduces information loss in traditional downsampling operations through spatial-depth conversion, with particularly significant protective effects on small target features26. Li et al. alleviated information degradation problems by adopting GhostNetV227 convolution modules. Zhang et al. proposed methods to extract discriminative semantic information through multi-branch convolution structures and dilated convolution. Attention mechanisms enhance the model’s perception of key information by dynamically adjusting feature weights, but many algorithms typically focus on solving problems from a global feature perspective, lacking sufficient perception of local details and unable to fully utilize target information, resulting in limited improvement in the model’s small target detection capability28. Tian et al. designed deformable attention modules that adaptively learn significant offsets to optimize meaningful local features, enhancing the network’s local perception capability25. Yi et al. proposed CoAttention, a dual-branch architecture combining downsampling attention with local feature enhancement, performing global context modeling through depthwise convolution and gating mechanisms. However, most feature enhancement strategies fail to fully utilize contextual information and cannot effectively capture meaningful contextual dependencies, with interactions between targets in complex scenes being largely ignored, undermining the model’s overall understanding of complex scenes. Building-specific detection methods Building detection in remote sensing imagery has motivated domain-specific architectural adaptations beyond generic rotated object detection. Beyond scale variance challenges common to small object detection, building targets present two domain-specific characteristics distinguishing them from vehicles or ships. Dense spatial arrangements characterize urban and suburban regions where inter- building proximity frequently approaches network receptive field limitations, creating feature overlap and occlusion ambiguities absent in isolated vehicle or ship scenarios where targets maintain spatial separation. Architectural geometry complexity involving L-shaped, U-shaped, and irregular polygonal footprints introduces orientation ambiguities beyond the predominantly rectangular or elliptical shapes of vehicles and ships, requiring detection frameworks to handle multi-component structures where individual building sections exhibit varying orientations. Scale-adaptive detection mechanisms addressing these characteristics have demonstrated effectiveness through dynamic receptive field adjustment29, cross-scale feature calibration30, and knowledge distillation-based fine-grained recognition31, validating the necessity of adaptive multi-scale feature selection in building detection. Attention-enhanced frameworks incorporate spatial-channel weighting mechanisms addressing these building-specific challenges. Li et al. proposed RSI-YOLO, which strengthens neural network-fused features by introducing channel attention and spatial attention mechanisms, improving the original network’s multi-scale feature fusion structure to a weighted bidirectional feature pyramid structure, achieving more efficient and richer feature fusion32. Liu et al. proposed MCRS-YOLO, designing a multi-branch aggregation network to enhance information flow and reconstructing the feature pyramid through dynamic interpolation and multi-feature fusion, further improving the model’s multi-scale feature representation capability33. Yang et al. designed UAV-YOLOv8, integrating BiFormer34 attention mechanisms to optimize the backbone network and introducing new detection scales to effectively fuse shallow and deep feature information35. Adaptive multi-scale fusion strategies attempt to balance semantic richness and spatial precision across pyramid hierarchies. Region-based refinement approaches employ dense sampling grids targeting precise small building localization. Notwithstanding these contributions, existing building detection methods exhibit three critical limitations: evaluation predominantly focuses on vertical imagery with moderate building densities, rotated bounding box prediction remains a secondary adaptation rather than integral architectural consideration, and performance degrades substantially under off-nadir viewing angles where perspective distortion compounds scale variance challenges. These gaps necessitate detection frameworks where multi-scale feature selection, progressive fusion, and orientation-aware regression jointly address building-specific detection requirements. Recent building detection advances (2022–2025) extend these directions with novel paradigms. Transformer-based architectures demonstrate improved small-target localization through deformable attention mechanisms, though computational costs remain prohibitive for real-time applications. Lightweight CNN variants reduce model complexity through depthwise separable convolutions but exhibit limited performance in dense building scenarios. Rotation-aware loss functions such as Gaussian Wasserstein Distance enhance angle regression stability yet introduce hyperparameter tuning complexity. Newly released datasets including FAIR1M and SODA-A provide multi-altitude annotations enriching evaluation benchmarks. Despite significant progress in existing algorithms, the following limitations remain: (1) Insufficient feature representation capability - existing algorithms are prone to feature loss when processing small targets, especially in deep network layers36,37; (2) Low multi-scale fusion efficiency - traditional feature pyramid networks easily generate semantic gap problems when fusing features at different scales38,39; (3) Limited rotated angle prediction accuracy - existing methods still have room for improvement in angle regression, especially in dense target scenarios40,41. MAR-YOLO addresses these technical challenges through three synergistic architectural innovations. The Multi-scale Feature Adaptive Selection (MFAS) module introduces a P2 detection layer combined with Focused Linear Attention mechanisms and learnable weight parameters to achieve adaptive filtering of multi-scale features. The AFPN adopts a three-stage progressive fusion strategy and adaptive spatial fusion mechanism to alleviate semantic gap problems caused by direct fusion of multi-scale features. The RepVGG reparameterization-based Enhanced Rotated Detection Head (RRD-Head) strengthens feature learning through multi-branch structures during training while maintaining single-path efficient structure during inference. Materials and methods Overview of MAR-YOLO MAR-YOLO extends YOLOv11-OBB for oriented building detection in remote sensing data through three architectural enhancements. YOLOv11-OBB is the latest rotated object detection algorithm released by Ultralytics, which stands out with an enhanced model structure by replacing the C2f structure with improved C3k2 modules based on the YOLOv8 architecture and introducing C2PSA attention modules after SPPF, achieving more efficient feature extraction capabilities and excellent performance in scenarios requiring precise angle information such as remote sensing images, text detection, and aerial targets. Remote sensing building detection confronts extreme scale variance coupled with dense small-target distribution. MAR-YOLO addresses these challenges through three architectural modifications to YOLOv11-OBB. The architecture incorporates a P2 detection layer integrated with the Multi-scale Feature Adaptive Selection (MFAS) module. The AFPN replaces traditional neck structures. RepVGG reparameterization mechanisms construct the Enhanced Rotated Detection Head (RRD-Head), strengthening feature learning through multi-branch structures during training while maintaining efficient single-path structures during inference. The overall network architecture follows the “backbone network - MFAS feature selection - AFPN progressive fusion - RRD detection head” pipeline, forming a complete technical chain from feature extraction, adaptive selection, progressive fusion to precise detection, significantly improving remote sensing building detection accuracy while maintaining model lightweightness. The overall MAR-YOLO architecture integrates four components into an end-to-end detection pipeline, as illustrated in Fig. 1. MAR-YOLO network. The network processes RGB input images (640 × 640 pixels) through a backbone comprising alternating C3k2 and CBS (Conv-BN-SiLU) modules, progressively generating four-scale features: P2 (4× downsampling), P3 (8×), P4 (16×), and P5 (32×). Before entering the neck network, P2, P3, and P4 features pass through red MFAS blocks for adaptive selection, while P5 features are weighted by learnable parameter α₅ after C2PSA attention. The AFPN-based neck (dashed box) fuses multi-scale features through progressive pathways indicated by green connection lines. The Enhanced Rotated Head (lower orange box) receives fused multi-scale features from the neck through a unified pathway: the upper OBB branch employs RepVGG blocks for oriented bounding box regression, while the lower Cls branch uses depthwise convolutions for classification, with both branches processing the same fused features but specializing in different prediction tasks. Feature adaptive P2 layer backbone network Significant scale differences and dense small-target distribution in remote sensing building images necessitate enhanced feature capture at finer resolutions. A P2 detection layer incorporates into the YOLOv8-OBB backbone network to address this requirement. As shown in Fig. 2(a), an additional downsampling branch is added after the third C3k2 module, generating a P2 feature layer with 4× downsampling through 3 × 3 convolution with stride 2. This feature layer, together with the original P3 (8× downsampling), P4 (16× downsampling), and P5 (32× downsampling), forms a four-scale feature pyramid, providing finer-grained spatial information for subsequent multi-scale feature fusion. Backbone of YOLOv11_P2 and YOLOv11_P2_MFAS. Direct incorporation of the P2 layer introduces backbone network feature redundancy, as different scale features contribute variably to final detection outcomes. The Multi-scale Feature Adaptive Selection Module (MFAS) addresses this challenge through dual-domain weighting mechanisms acting before neck network input, utilizing FocusedLinearAttention (FLA)13 as the core component and linking adaptive learning parameters α={α₂, α₃, α₄}. Figure 2(b) illustrates how MFAS alleviates information imbalance arising from small target detection layer integration, as expressed in Formula (1): where represents the input feature map at pyramid level i, denotes the adaptively weighted features, and is the learnable weighting parameter. FLA achieves efficient attention computation through focusing factors and depthwise separable convolution, as shown in Formula (2). Given the P5 layer already incorporates the C2PSA self-attention module, only learnable weight parameter α₅ applies to P5 features, whereas complete MFAS modules integrate into P2, P3, and P4 layers. where , , represent query, key, and value matrices; denotes the focusing function with focusing factor ; DWC represents depthwise convolution with kernel size 5 for local feature enhancement. FAS design specifically addresses remote sensing building detection characteristics through two complementary mechanisms. FLA enhances focusing capability on key features including building edges and corners via the focusing factor = 3 configuration, which proves crucial for identifying densely distributed building contours. The embedded depthwise separable convolution (kernel size = 5) captures local structural information of buildings while maintaining computational efficiency. This dual weighting architecture enables dynamic adjustment of feature contributions from different backbone layers according to building scale distributions across diverse scenarios. Compared with traditional fixed-weight fusion strategies, MFAS achieves adaptivity in feature selection through end-to-end learning. During training, weight parameters α={α₂, α₃, α₄, α₅} are automatically learned through backpropagation, enabling the model to adjust feature fusion strategies for different types of remote sensing images (such as different resolutions and different building densities). AFPN-based progressive feature pyramid network In remote sensing image building detection tasks, directly fusing multi-scale features including the P2 layer faces significant semantic gap problems. P2 layer features undergo only 4× downsampling, retaining rich spatial detail information but with weak semantic expression; while P5 layer features undergo 32× downsampling, possessing strong semantic representation but low spatial resolution. Traditional FPN and PAN structures adopt layer-by-layer transmission, where high-level semantic information must pass through multiple intermediate scales to reach the bottom layer, inevitably causing information loss and feature degradation during this process. Figure 3 illustrates architectural evolution from traditional FPN to the proposed AFPN. Traditional FPN employs uniform top-down fusion with fixed convolution kernels, causing semantic gaps when combining non-adjacent pyramid levels such as P2 and P5 features. The original Asymptotic AFPN addresses this limitation through progressive three-stage fusion, sequentially integrating adjacent levels to reduce semantic discrepancies during feature propagation. However, its uniform 3 × 3 convolutions across all fusion stages lack adaptability to scale-specific feature characteristics. The proposed AFPN extends this progressive framework through two building- oriented modifications: scale-matched convolution kernels (3 × 3 for fine- detail preservation in P2-P3 fusion, 5 × 5 for elongated structure capture in P4-P5 integration) and learned spatial fusion weights replacing fixed averaging, enabling position-dependent adaptation to building density variations in remote sensing imagery. Comparison of feature pyramid network architectures. The AFPN architecture implements progressive multi-scale integration through three sequential fusion stages, where Stage 1 combines P2 and P3 features, Stage 2 incorporates the fused result with P4, and Stage 3 integrates P5 to complete full-scale feature fusion. This progressive design philosophy addresses semantic gap problems arising from direct fusion of feature layers with excessive semantic differences. The fusion process operates through semantic proximity preservation. Adjacent low-level features P2 and P3 undergo initial fusion to reduce inter-level semantic discrepancies. Subsequent integration introduces mid-level feature P4 into the fusion process, at which point P2 has acquired semantic information through P3 fusion, thereby reducing direct fusion difficulty with P4. Final integration incorporates top-level feature P5, completing full-scale feature fusion while ensuring each stage operates between semantically similar features, consequently avoiding information conflicts caused by direct fusion of non-adjacent levels. The implemented AFPN adapts the asymptotic fusion framework14 through building- oriented architectural modifications addressing scale-specific feature characteristics. Departing from the original design’s uniform 3 × 3 convolutions across all fusion stages, the adapted version introduces scale-matched kernels differentiated by pyramid levels: 3 × 3 convolutions preserve small building details during P2-P3 fusion, whereas 5 × 5 convolutions capture elongated building structures throughout P4-P5 integration. Additionally, Adaptive Spatial Fusion coefficients , , in Formula (3) undergo computation through 1 × 1 convolutions rather than fixed weights, thereby enabling position-dependent fusion adaptive to building density variations, with ablation experiments demonstrating scale-matched kernels contributing 0.3% mAP50 gain while learned weights adding 0.1% improvement over the original design. Targeting the characteristics of dense building distribution in remote sensing images, AFPN introduces an Adaptive Spatial Fusion (ASF) mechanism to handle spatial position conflicts during multi-layer feature fusion. At the same spatial position, different scale features may contain information from different buildings, and direct element-wise addition would cause feature confusion. ASF achieves adaptive feature fusion by assigning learnable weight coefficients to different level features at each spatial position. For feature fusion at position (i, j), it is defined as shown in Formula (3). where represents the fused feature at spatial position for fusion stage l; denotes the feature from fusion path k at stage l; , , are learnable spatial fusion weights computed through 1 × 1 convolutions and normalized to sum to 1 at each position, enabling position-dependent adaptive fusion. AFPN forms effective synergy with the feature adaptive P2 layer backbone network. In forward propagation, P2-P5 features weighted by the MFAS module are input to AFPN for progressive fusion, with MFAS adaptive weights ensuring the quality and relevance of input features. AFPN’s progressive fusion strategy further enhances feature representation capability, particularly with P2 layer detail information being effectively preserved through multi-stage fusion. Enhanced rotated object detection head Building targets in remote sensing images have characteristics of directional diversity, clear boundaries but complex geometric shapes, placing extremely high demands on rotated box localization accuracy. The original YOLOv11-OBB detection head employs cascaded standard convolutional layers for bounding box regression. While this single-path design is efficient for inference, it struggles to fully learn the complex geometric features of buildings during training. RepVGG15 achieves equivalent transformation from multi-branch architecture during training to single-branch architecture during inference through structural reparameterization technology. Standard RepVGG applications employ identical branch configurations across all convolutional layers regardless of task-specific requirements. RRD-Head establishes architectural differentiation through explicit branch role specialization addressing angle regression requirements in rotated detection. The three-branch structure assigns distinct responsibilities where 3 × 3 convolutions extract edge orientation features capturing building boundaries under perspective distortion, 1 × 1 convolutions model inter-channel angle correlation patterns encoding geometric relationships, while identity mappings stabilize gradient flow preventing orientation ambiguity during backpropagation. This role-differentiated design contrasts with uniform branch configurations in conventional RepVGG implementations, enabling enhanced feature learning for oriented bounding box regression without increasing inference computational cost through training-inference decoupled architecture, as illustrated in Fig. 4. Detection head architecture comparison. RRD-Head establishes architectural differentiation from standard RepVGG implementations through explicit branch role specialization addressing angle regression requirements: 3 × 3 branches extract edge orientation features, 1 × 1 branches model angle correlation patterns, while identity branches stabilize gradient flow throughout backpropagation. Ablation experiments systematically removing this specialization mechanism demonstrate 0.5% mAP50 degradation, thereby validating the necessity of role-differentiated branch design for oriented bounding box regression tasks. In RRD-Head, the convolutional layers of the localization branch are replaced with RepVGG modules containing three parallel branches: 3 × 3 convolution extracts local spatial features, 1 × 1 convolution achieves cross-channel information interaction, and identity mapping maintains feature transmission. During training, the three-branch outputs are fused through element-wise addition, as shown in Formula (4). where x is the input feature; , , and identity branches with batch normalization extract local features, enable cross-channel interaction, and preserve original features respectively. During inference, the three branches are reparameterized into a single 3 × 3 convolution. The key to structural reparameterization lies in equivalently transforming the multi-branch structure into a single 3 × 3 convolution. This process is based on the linear properties of convolution operations and the mathematical characteristics of batch normalization: first, the 1 × 1 convolution kernel is expanded to a 3 × 3 convolution kernel through zero padding, and the identity mapping is represented as a 3 × 3 convolution kernel with the center element as 1 and others as 0; then, utilizing the linear transformation properties of BN layers, the BN parameters of the three branches are respectively fused into the corresponding convolution weights; finally, the three equivalent 3 × 3 convolution kernels and their bias terms are added to obtain the final single 3 × 3 convolutional layer. This transformation ensures mathematical equivalence, making the computational graph during inference completely consistent with traditional VGG structures. For building rotation detection, the multi-branch structure of RRD-Head enhances the perception capability of features at different scales, and multi-path gradient propagation helps accurate prediction of rotation angle θ. RepVGG modules contain no nonlinear activation functions, ensuring the feasibility of reparameterization and the linear additivity of features, suitable for continuous angle regression of rotated boxes. During inference, it transforms to a single-path structure, eliminating the memory access overhead of multi-branches, improving localization accuracy in dense building scenarios while maintaining real-time performance. Datasets BONAI (Buildings in Off-Nadir Aerial Images) is a large-scale building detection dataset specifically designed for building footprint extraction tasks in non-vertical aerial images. The dataset contains 3,300 high-resolution aerial images with a total of 268,958 annotated building instances. The images were collected from six representative cities in China: Shanghai, Beijing, Harbin, Jinan, Chengdu, and Xi’an, covering different geographical regions, climatic conditions, and architectural styles. Unlike traditional vertical aerial datasets, buildings in BONAI exhibit significant offsets between rooftop and ground footprints, with each building instance providing three types of annotations: rooftop contours, ground footprints, and offset vectors. In the experiments of this paper, to adapt to the input requirements of YOLOv11-OBB and ensure training efficiency, the following processing was applied to the BONAI dataset16: The training set, validation set, and test set were divided in a 70:15:15 ratio, containing 2,310, 495, and 495 images respectively. A stratified sampling strategy was employed to ensure that the distribution of different cities and building density regions in each subset remained consistent with the original dataset. Since the original annotations were in instance segmentation format, OpenCV’s minimum area bounding rectangle algorithm was used to convert them to rotated bounding box representation, extracting center coordinates, width-height, and rotation angle parameters. Original BONAI images exhibit variable spatial resolutions across different urban regions as documented in the dataset source16. Since the original annotations were in instance segmentation format, OpenCV’s minimum area bounding rectangle algorithm was used to convert them to rotated bounding box representation, extracting center coordinates, width-height, and rotation angle parameters. All images were uniformly resized to 640 × 640 pixels to improve model training efficiency, and radiometric correction and histogram equalization were performed to enhance building edge contrast. Rotation angles were normalized to the [-90°, 90°] range, and finally saved in YOLO format, with each line containing the class index (building class as 0) and normalized five-parameter rotated box coordinates. The processed dataset examples are shown in Fig. 5. Dataset examples. To validate cross-domain generalization capability, three additional public benchmarks complement the building-centric BONAI evaluation. DOTA v1.042 comprises 2,806 aerial images with original resolutions ranging from 800 × 800 to 4000 × 4000 pixels, containing 188,282 annotated instances across 15 object categories including planes, ships, and vehicles, providing diverse rotated object detection scenarios beyond building-specific contexts. DIOR-R2 contains 23,463 uniformly sized 800 × 800 pixel images with 192,472 oriented bounding box annotations spanning 20 object classes in optical remote sensing imagery, emphasizing small object detection challenges across varied geographical environments. HRSC201643 focuses on ship detection with 1,061 images exhibiting variable resolutions ranging from 300 × 300 to 1500 × 900 pixels, containing 2,976 ship instances captured from Google Earth, representing concentrated object distributions under varying scales and orientations. All cross-dataset experiments employ official training-validation-test splits with images uniformly resized to 640 × 640 pixels through letterbox padding. Dataset-specific training configurations maintain consistency with BONAI protocols described in Sect. 4.1.1. DOTA v1.0 utilizes the standard benchmark partition with 1,411 training images, 458 validation images, and 937 test images as specified in the original publication. DIOR-R follows the official split allocating samples across training, validation, and test subsets according to the dataset documentation. HRSC2016 adopts 436 training images, 181 validation images, and 444 test images following the established benchmark protocol. Training employs 300 epochs with SGD optimizer, initial learning rate 0.01, momentum 0.937, weight decay 0.0005, and cosine annealing learning rate schedule. Data augmentation applies random horizontal-vertical flipping with 0.5 probability and mosaic augmentation disabled in the final 10 epochs to stabilize convergence. Batch size adjusts according to dataset scale: 32 for DOTA v1.0 given lower per-image object density, 16 for DIOR-R matching BONAI configuration, and 8 for HRSC2016 due to limited training samples. These configurations ensure reproducibility while accommodating dataset-specific characteristics. Evaluation indicators This paper uses Precision, Recall, mAP50, and mAP50-95 as model accuracy evaluation metrics, as shown in Formula (5) to Formula (7). Precision and Recall are fundamental evaluation metrics that reflect the accuracy and completeness of model detection, respectively. Mean Average Precision (mAP) is calculated from the precision-recall curve and serves as the core metric for comprehensive evaluation of model detection performance. mAP50 represents the average precision at IoU threshold of 0.5, while mAP50-95 represents the average mAP value from IoU 0.5 to 0.95 (with step size 0.05), with the latter requiring stricter localization accuracy. Additionally, this paper uses parameter count (MB) and computational complexity (GFLOPs) to evaluate model complexity, reflecting the model’s storage requirements and computational overhead respectively. Inference speed is measured in FPS (Frames Per Second). where TP, FP, and FN represent the number of correctly detected buildings, falsely detected buildings, and undetected buildings, respectively. AP is the average precision for a single class; mAP is the average value across all detection classes. When IoU is set to 0.5, mAP50 refers to the average precision at IoU threshold of 0.5, while mAP50-95 refers to the average mAP value from IoU 0.5 to 0.95 with step size of 0.05, consistent with MS COCO evaluation metrics. Results Ablation study Ablation experiments validate individual module contributions and their synergistic effects through systematic component removal and progressive integration. The analysis comprises two stages: MFAS module evaluation examines P2 layer integration strategies, whereas progressive module integration assesses cumulative performance gains from MFAS, AFPN, and RRD-Head combination. MFAS module ablation analysis The experimental environment comprises Windows 11 OS equipped with Intel i9-12900 K processor and NVIDIA RTX 3090 graphics card (24GB VRAM), using PyTorch 2.1.0 and CUDA 12.1. Training employs systematic hyperparameter configuration ensuring reproducibility: input images resize to 640 × 640 pixels through letterbox padding preserving aspect ratios, batch size of 16 balancing GPU memory utilization and gradient stability, and 300 training epochs achieving full convergence. The SGD optimizer applies momentum 0.937 and weight decay 0.0005, with initial learning rate 0.01 decaying through cosine annealing schedule reaching 0.0001 at final epoch. Data augmentation strategies include random horizontal-vertical flipping (probability 0.5) and mosaic augmentation (4-image tiling) disabled in the final 10 epochs to stabilize convergence, with additional augmentations following YOLOv11 default configurations including HSV color space transformations and random scaling. No pre-trained weights initialize training, enabling evaluation of architectural innovations from random initialization. All experiments maintain identical configurations across datasets except batch size variations adapting to dataset scale: batch size 16 for BONAI, 32 for DOTA v1.0, 16 for DIOR-R, and 8 for HRSC2016. To verify the effectiveness of the Multi-scale Feature Adaptive Selection Module (MFAS), this experimental section designs three schemes for comparative analysis: The baseline model adopts the original YOLOv11s-OBB, using only three feature layers P3, P4, and P5 for detection. baseline + P2_only introduces the P2 detection layer based on the baseline model, but directly inputs the four feature layers to the neck network without feature selection; baseline + P2_MFAS introduces the P2 layer while adding the MFAS module to perform adaptive weighted selection on P2-P5 features. The experimental results are shown in Table 1. Table 1MFAS component ablation experimental results.ModelPrecisionRecallmAP50mAP50-95Parameters(MB)GFLOPsFPSBaseline87.281.584.362.7 9.42 28.4 142 Baseline + P287.882.485.163.311.3635.7118Baseline + P2_MFAS 88.5 83.1 85.7 63.8 12.1537.2112 MFAS component ablation experimental results. Experimental outcomes reveal that P2 detection layer incorporation (baseline + P2_only) yields 0.8% mAP50 improvement and mAP50-95 by 0.6%, verifying the importance of fine-grained features for small building detection. However, directly introducing the P2 layer results in a 20.6% increase in parameters and a 16.9% decrease in FPS. After adding the MFAS module (baseline + P2_MFAS), mAP50 further improves by 0.6% to reach 85.7%, and mAP50-95 increases to 63.8%. Although the parameter count increases by only 6.9% compared to baseline + P2_only, the performance shows stable improvement, indicating that MFAS effectively suppresses redundant information through adaptive feature selection and improves feature utilization efficiency. Examining parameter efficiency reveals that introducing P2 layer alone achieves 0.41% mAP50 gain per million parameters (0.8% gain / 1.94 M), while MFAS achieves 0.76% per million parameters (0.6% gain / 0.79 M), demonstrating that MFAS contributes to accuracy improvement through architectural design rather than merely adding computational capacity. Considering both detection accuracy and computational efficiency, the baseline + P2_MFAS scheme is selected for subsequent research. Progressive module integration analysis To comprehensively verify the effectiveness of the three innovative modules and their synergistic effects, ablation experiments evaluate both progressive integration and isolated module contributions. Table 2 presents results across seven configurations including baseline, three single-module additions (MFAS-only, AFPN-only, RRD-Head-only), P2 + MFAS combination, P2 + MFAS + AFPN integration, and the complete model with all components. This experimental design enables analysis of individual module contributions, synergistic effects between P2 and MFAS, and cumulative performance gains from progressive integration. Table 2Comprehensive ablation experiment results.ModelP2MFASAFPNRRD-HeadmAP50mAP50-95Parameters(MB)Baseline✗✗✗✗84.362.7 9.42 +MFAS only✗✓✗✗84.863.010.21+AFPN only✗✗✓✗85.263.511.14+RRD-Head only✗✗✗✓84.963.19.81P2 + MFAS✓✓✗✗85.763.812.15P2 + MFAS + AFPN✓✓✓✗86.564.613.87Full model✓✓✓✓ 87.2 65.3 14.26 Comprehensive ablation experiment results. Table 2 reveals AFPN achieves the largest isolated gain (0.9% mAP50), while MFAS-only contributes 0.5%, confirming MFAS requires P2 features for effectiveness (0.5% alone vs. 1.4% with P2). RRD-Head provides consistent 0.6–0.7% improvements regardless of configuration. The non-additive synergy between P2 and MFAS validates targeted architectural design over capacity expansion. Figure 6 visualizes feature activation intensity through heatmaps overlaid on five representative building images, comparing YOLOv8s-OBB (row a) and MAR-YOLO (row b). The color scale ranges from blue (low activation, 0) to yellow (high activation, 1). MAR-YOLO exhibits more concentrated activation patterns on building structures with reduced background noise across all scenes. The enhanced feature localization validates the effectiveness of MFAS in filtering multi-scale features and AFPN in preserving discriminative information through progressive fusion. Comparison of feature activation heatmaps. Comparative experiments To verify the proposed method’s effectiveness and generalization capability, comparative experiments were conducted with representative rotated object detection algorithms including single-stage YOLO variants (YOLOv9s-OBB, YOLOv8s-OBB, YOLOv5s-OBB), Transformer-based detector (Rotated DETR), two-stage methods (Rotated Faster R-CNN, Oriented R-CNN), and S2ANet. All models were trained using unified strategies with dataset-specific official splits: BONAI employs 70:15:15 train-validation-test partitioning as described in Sect. 3.5, whereas DOTA v1.042, DIOR-R2, and HRSC201643 utilize respective official test sets for evaluation following community-standard protocols. Results are presented in Tables 3, 4, 5 and 6, with all reported metrics computed on designated test partitions to ensure fair comparison and reproducibility. Table 3Comparison with mainstream rotated object detection algorithms on BONAI dataset.Detector typeModelPrecisionRecallmAP50mAP50-95Parameters (MB)GFLOPsFPSSingle-stageYOLOv5s-OBB85.779.882.660.4 7.28 17.2 168 YOLOv8s-OBB87.682.384.863.111.1328.6135YOLOv9s-OBB88.182.785.463.615.839.191S2ANet87.980.684.162.538.9198.342Two-stageRotated Faster R-CNN88.478.583.261.841.35134.525Oriented R-CNN89.181.985.363.741.12129.728TransformerRotated DETR87.981.584.662.841.5152.323ProposedMAR-YOLO 89.8 84.7 87.2 65.3 14.2643.895 Comparison with mainstream rotated object detection algorithms on BONAI dataset. Table 4Comparison with mainstream rotated object detection algorithms on DOTA v1.0 dataset.Detector typeModelPrecisionRecallmAP50mAP50-95Single-stageYOLOv5s-OBB75.368.171.250.8YOLOv8s-OBB76.869.572.952.1YOLOv9s-OBB77.570.373.652.9S2ANet76.968.872.451.8Two-stageRotated Faster R-CNN77.167.271.851.2Oriented R-CNN78.369.773.252.6TransformerRotated DETR76.268.972.151.5ProposedMAR-YOLO 79.2 72.9 75.8 54.6 Comparison with mainstream rotated object detection algorithms on DOTA v1.0 dataset. Table 5Comparison with mainstream rotated object detection algorithms on DIOR-R dataset.Detector typeModelPrecisionRecallmAP50mAP50-95Single-stageYOLOv5s-OBB70.262.165.845.9YOLOv8s-OBB71.963.867.547.3YOLOv9s-OBB72.864.968.448.1S2ANet72.063.567.347.1Two-stageRotated Faster R-CNN71.861.566.246.3Oriented R-CNN73.164.368.147.9TransformerRotated DETR71.363.166.946.8ProposedMAR-YOLO 74.6 66.8 70.1 49.7 Comparison with mainstream rotated object detection algorithms on DIOR-R dataset. Table 6Comparison with mainstream rotated object detection algorithms on HRSC2016 dataset.Detector typeModelPrecisionRecallmAP50mAP50-95Single-stageYOLOv5s-OBB90.585.287.569.2YOLOv8s-OBB91.286.488.770.5YOLOv9s-OBB91.786.989.271.1S2ANet91.386.188.670.3Two-stageRotated Faster R-CNN91.585.188.169.8Oriented R-CNN93.187.289.571.5TransformerRotated DETR90.985.888.370.1ProposedMAR-YOLO92.788.590.672.8 Comparison with mainstream rotated object detection algorithms on HRSC2016 dataset. Tables 3, 4, 5 and 6 demonstrate MAR-YOLO’s consistent superiority across diverse datasets and detector architectures. Comparative methods partition into single-stage detectors (YOLOv5s-OBB, YOLOv8s-OBB, YOLOv9s-OBB, S2ANet) emphasizing real-time inference, two-stage detectors (Rotated Faster R-CNN, Oriented R-CNN) prioritizing accuracy through region proposals, and transformer-based architectures (Rotated DETR) employing global attention. On BONAI (Table 3), MAR-YOLO achieves 87.2% mAP50, outperforming the strongest single-stage baseline YOLOv9s-OBB by 1.8% and two-stage Oriented R-CNN by 1.9% at 3.4× faster inference (95 FPS vs. 28 FPS). Cross-dataset validation reveals robust gains: 2.2% improvement on DOTA v1.0 (Table 4), 1.7% on DIOR-R (Table 5), and 1.4% on HRSC2016 (Table 6) over YOLOv9s-OBB. Against transformer-based Rotated DETR, MAR-YOLO achieves 2.6–3.7% higher mAP50 while requiring only 28.7% computational cost (43.8 vs. 152.3 GFLOPs) and 4.1× faster inference. The consistent advantages validate architectural innovations: MFAS dual-domain selection addresses extreme scale variance through adaptive P2-layer filtering (0.76%/M efficiency); AFPN progressive fusion with scale-matched kernels eliminates semantic gaps at 3.5× lower cost than global attention; RRD-Head branch-specialized reparameterization achieves 1.79%/M efficiency, resolving orientation instability across detection scenarios. To validate the effectiveness of the proposed modules in complex scenarios, we selected two representative challenging scenes for qualitative comparison. These scenes demonstrate densely arranged urban buildings and regularly distributed building areas, covering different lighting conditions and scale variations. YOLOv11s-OBB was chosen as the baseline for comparison, which shares the same backbone network as MAR-YOLO but does not integrate the three core modules proposed in this work. This setting is consistent with the ablation study and can intuitively demonstrate the synergistic effects of MFAS, RRD-Head, and AFPN. As shown in Fig. 7, from left to right are the detection results of YOLOv11s-OBB, MAR-YOLO, and Ground Truth, respectively. The comparison clearly shows the detection completeness advantage of our method. Qualitative comparison of detection results on challenging scenarios. As shown in the comparison results in Fig. 7, MAR-YOLO demonstrates superior detection completeness compared to YOLOv11s-OBB in both scenarios. In the top row showing dense urban buildings, YOLOv11s-OBB exhibits noticeable missed detections, while MAR-YOLO detects more small-scale building targets, approaching the annotation completeness of Ground Truth. In the bottom row showing regularly arranged buildings, MAR-YOLO also demonstrates more comprehensive detection coverage. These results validate the effectiveness of the MFAS, RRD-Head, and AFPN modules in improving detection performance. Model generalization verification To verify the generalization capability of the proposed method, comparative experiments were conducted on three authoritative remote sensing rotated object detection datasets. BONAI serves as the primary benchmark for ablation studies due to its building-specific annotations and off-nadir characteristics, while DOTA v1.042, DIOR-R2, and HRSC201643 provide cross-domain validation across different object types and imaging conditions. Mainstream YOLO variants (YOLOv8s-OBB, YOLOv11s-OBB) and MAR-YOLO were trained using identical strategies on each dataset’s official splits. The experimental results are shown in Table 7. Table 7Verification of model generalization performance on different datasets.DatasetModelPrecisionRecallmAP50mAP50-95DOTA v1.0[42]YOLOv8s-OBB76.869.572.952.1YOLOv11s-OBB77.370.273.552.8MAR-YOLO 79.2 72.9 75.8 54.6 DIOR-R[2]YOLOv8s-OBB71.963.867.547.3YOLOv11s-OBB72.564.768.248.1MAR-YOLO 74.6 66.8 70.1 49.7 HRSC2016[43]YOLOv8s-OBB91.286.488.770.5YOLOv11s-OBB91.887.189.371.2MAR-YOLO 92.7 88.5 90.6 72.8 Verification of model generalization performance on different datasets. Table 7 validates MAR-YOLO’s superiority across datasets compared to mainstream YOLO variants. Against YOLOv8s-OBB, MAR-YOLO achieves 2.9% mAP50 gain on DOTA, 2.6% on DIOR-R, and 1.9% on HRSC2016, demonstrating MFAS dual-domain selection addresses extreme scale variance where traditional pyramids miss features. Against YOLOv11s-OBB baseline, the consistent 1.3–2.3% improvements confirm AFPN progressive fusion eliminates semantic gaps universally, while RRD-Head branch specialization stabilizes angle regression across diverse orientations. The cross-dataset robustness validates architectural innovations solve fundamental limitations rather than dataset-specific tuning. To validate how architectural innovations adapt across diverse scenarios and address concerns of dataset-specific overfitting, per-module ablation experiments were conducted across all four evaluation datasets. Table 8 presents the results. Table 8Per-module cross-dataset ablation results (mAP50%).DatasetBaseline+MFAS+AFPN+RRD-HeadBONAI84.385.786.5 87.2 DOTA v1.073.574.975.3 75.8 DIOR-R68.269.469.8 70.1 HRSC201689.389.990.4 90.6 Per-module cross-dataset ablation results (mAP50%). Table 8 reveals how innovations address dataset-specific limitations. MFAS contribution variance (BONAI 1.4% vs. HRSC2016 0.6%) demonstrates dual-domain selection effectively filters P2-layer redundancy for extreme scale scenarios while requiring less filtering for concentrated ship scales, validating scale-specific design. AFPN stable contributions (0.4–0.8%) confirm progressive fusion with scale-matched kernels universally eliminates semantic gaps regardless of object type, demonstrating advantages over direct multi-scale concatenation. RRD-Head correlation with orientation complexity (BONAI 0.7% vs. HRSC2016 0.2%) validates branch specialization—3 × 3 edge-oriented and 1 × 1 correlation-oriented branches—stabilizes angle regression for diverse rotations. The consistent cross-dataset improvements confirm innovations solve fundamental architectural limitations. Despite consistent improvements, dataset-specific challenges reveal architectural limitations. On HRSC2016, MAR-YOLO’s modest gains (1.3% over baseline, Table 9) reflect difficulties in densely packed ship scenarios where inter-vessel spacing falls below 10 pixels—MFAS P2-layer filtering provides limited benefits when targets concentrate in 32–96 pixel range, reducing scale variance advantages. Dense arrangements cause merged detections when overlapping features confuse the neck network, while AFPN’s progressive fusion struggles to preserve individual ship boundaries under extreme occlusion (> 40% overlap). RRD-Head angle prediction degrades for partially occluded ships where edge gradients become unreliable, suggesting future integration of contour refinement and occlusion-aware mechanisms could address these limitations. Failure case analysis Despite consistent cross-dataset performance improvements demonstrated in comparative and generalization experiments, understanding architectural boundaries under extreme conditions provides essential insights for future development. Failure case analysis serves three critical purposes: identifying specific scenarios where current architectural innovations reach performance limits, validating hypothesized relationships between module designs and failure patterns, and establishing concrete directions for subsequent research. Representative challenging scenarios were selected from BONAI, DOTA v1.0, and HRSC2016 test sets through systematic examination focusing on three recurring limitation patterns observed during qualitative analysis. The first pattern involves extreme density scenarios where buildings or objects exhibit minimal inter- target spacing, causing feature overlap during neck network fusion despite MFAS adaptive filtering mechanisms. The second pattern encompasses perspective distortion challenges where off-nadir viewing angles introduce geometric deformations affecting RRD-Head orientation prediction stability through weakened edge gradients and ambiguous corner features. The third pattern addresses severe occlusion conditions in densely packed arrangements where AFPN progressive fusion struggles to maintain individual target boundaries when semantic information from deeper pyramid levels dominates spatial details from shallow layers. Figure 8 presents these failure patterns through color-coded annotations highlighting specific detection errors, merged predictions, and orientation instabilities across the three datasets. Representative failure cases demonstrating architectural limitations across three datasets. Analysis of color-coded annotations reveals module-specific limitations under extreme conditions. (a) demonstrates density-related failures where red overlay indicates adjacent buildings merged into single detections due to feature overlap during neck network fusion, while yellow circles mark small residential structures missed despite MFAS adaptive filtering, confirming P2-layer features remain insufficient at extreme scale lower bounds. (b) presents orientation instability where orange dashed boxes identify storage tanks exhibiting prediction-truth angle misalignment under perspective distortion, revealing RRD-Head challenges when oblique viewing weakens edge gradients essential for rotation regression. (c) illustrates occlusion-induced failures where cyan overlay encompasses ship clusters producing fewer detections than actual vessel counts, demonstrating AFPN progressive fusion limitations when deeper semantic information dominates P2 spatial details under severe overlap. These patterns establish performance boundaries—extreme density, perspective distortion, and high occlusion—motivating future research on contour-aware fusion mechanisms and occlusion-robust regression strategies. Discussion MAR-YOLO establishes a rotated object detection framework addressing building detection challenges in remote sensing imagery through three synergistic architectural innovations integrated into the YOLOv11-OBB baseline. The Multi-scale Feature Adaptive Selection Module (MFAS) resolves feature redundancy problems arising from P2 detection layer introduction within the backbone network, where FocusedLinearAttention mechanisms combined with learnable weight parameters achieve adaptive filtering of P2-P5 multi-scale features, thereby significantly improving small building detection capabilities while maintaining computational efficiency. Direct fusion of multi-scale features exhibiting large semantic differences causes information conflicts, necessitating the AFPN integration for neck network feature fusion optimization, which through three-stage progressive fusion strategies coupled with adaptive spatial fusion mechanisms, effectively alleviates semantic gap problems between P2-P5 features across pyramid levels. Rotated object detection angle regression challenges motivate the RepVGG reparameterization-based Enhanced Rotated Detection Head (RRD-Head) design, where multi-branch structures enhance feature learning capabilities during training while ensuring inference efficiency through structural reparameterization technology. Experimental results show that the MAR-YOLO model achieves 87.2% mAP50 and 65.3% mAP50-95 on the BONAI dataset, representing improvements of 2.9% and 2.6% respectively compared to the original YOLOv11s-OBB. Compared with other mainstream rotated object detection algorithms, the proposed method surpasses two-stage methods such as Oriented R-CNN in detection accuracy while maintaining 95 FPS inference speed. Furthermore, generalization validation experiments on three public datasets—DOTA v1.0, DIOR-R, and HRSC2016—demonstrate the robustness and universality of the proposed method, fully meeting the requirements for real-time building detection in remote sensing images. Beyond empirical performance improvements, MAR-YOLO contributes three methodological principles transferable to multi-scale object detection. The dual-domain feature selection paradigm—spatial attention combined with channel-wise weighting—establishes an efficiency benchmark where adaptive mechanisms achieve 0.76% mAP50 per million parameters versus 0.41% for capacity-only expansion, demonstrating architectural design superiority over parameter scaling. The scale-matched progressive fusion framework validates that kernel size serves as a critical fusion design parameter: 3 × 3 convolutions preserve fine details in adjacent-level fusion while 5 × 5 convolutions capture elongated structures in distant-level integration, offering empirical guidance for feature pyramid construction. The branch role differentiation in structural reparameterization—3 × 3 for edge orientation, 1 × 1 for angle correlation, identity for gradient stability—extends reparameterization applicability from classification to regression tasks, achieving 1.79% mAP50 per million parameters. These architectural innovations address fundamental challenges in remote sensing object detection, providing methodological foundations applicable to drone surveillance, disaster assessment, and infrastructure monitoring. Despite significant improvements achieved by MAR-YOLO across all aspects, several limitations warrant further investigation. While the additional P2 layer introduced by the MFAS module enhances small target detection capability, further lightweighting is needed for scenarios with extreme real-time requirements, and the current attention weight allocation strategy still has room for improvement in recognizing building boundary features under complex backgrounds. Although AFPN’s three-stage progressive fusion strategy improves feature quality, it correspondingly increases network complexity and parameter count, potentially limiting deployment on resource-constrained edge devices, and the current fusion strategy, primarily optimized for building targets, requires further validation of generalization capability to other rotated objects. While RRD-Head’s multi-branch structure during training enhances learning capability, it presents memory overhead challenges when training on large-scale datasets, and angle prediction stability for highly irregular or occluded buildings requires further enhancement through incorporating shape prior knowledge or multi-task learning strategies. Conclusions This paper proposes MAR-YOLO addressing unresolved interdependencies of scale variance, orientation instability, and semantic gaps in remote sensing building detection through three synergistic architectural innovations. The MFAS module introduces dual-domain feature selection combining spatial FocusedLinearAttention with channel-wise learnable parameters, achieving adaptive P2-P5 filtering while suppressing redundancy. The adapted AFPN employs progressive fusion with scale-matched kernels (3 × 3 for P2-P3, 5 × 5 for P4-P5) and learned spatial weights, eliminating semantic gaps inherent in direct multi-scale concatenation. The RRD-Head applies branch-specialized structural reparameterization—differentiating 3 × 3 edge-oriented, 1 × 1 correlation-oriented, and identity gradient-stabilizing branches—specifically targeting angle regression challenges. Experimental validation demonstrates 2.9% mAP50 and 2.6% mAP50-95 improvements over YOLOv11s-OBB on BONAI (achieving 87.2% and 65.3% respectively) at 95 FPS, with parameter efficiency analysis confirming architectural innovation drives performance: MFAS achieves 0.76%/M, AFPN 0.47%/M, and RRD-Head 1.79%/M efficiency, validating design superiority over capacity expansion. Cross-dataset validation on DOTA v1.0, DIOR-R, and HRSC2016 reveals consistent module contributions (MFAS: 0.6–1.4%, AFPN: 0.4–0.8%, RRD-Head: 0.2–0.7%), confirming robustness across diverse scenarios. Future research encompasses three promising extensions addressing deployment constraints and application scenarios. Knowledge distillation strategies could transfer MAR-YOLO’s architectural advantages to lightweight variants suitable for edge deployment while preserving detection accuracy, thereby addressing real-time constraints in resource-limited environments. Integration with emerging foundation models—particularly vision-language models for zero-shot detection and diffusion models for data augmentation—may enhance generalization to rare building types and extreme imaging conditions beyond current training distributions. Extension of the progressive fusion framework to temporal sequences could enable building change detection in multi-temporal remote sensing imagery, supporting applications in urban expansion monitoring and disaster damage assessment where temporal consistency proves critical. List of symbols Feature pyramid levels with 4×, 8×, 16×, 32× downsampling Learnable weighting parameters for adaptive feature selection Adaptive spatial fusion coefficients Rotation angle of oriented bounding box True positive, false positive, false negative Mean average precision at IoU threshold 0.5 Mean average precision from IoU 0.5 to 0.95 Frames per second Appendix A: system parameters and design values ParameterValueDescriptionTraining configuration Input size640 × 640Image resolution Batch size16Training batch size Learning rate0.01Initial learning rate OptimizerSGDStochastic Gradient Descent Training epochs300Total training iterationsMFAS module parameters FLA focusing factor (f)3Attention focusing parameter FLA kernel size5Depthwise convolution kernel P2 downsampling4×Feature stride at P2 layer P3 downsampling8×Feature stride at P3 layer P4 downsampling16×Feature stride at P4 layer P5 downsampling32×Feature stride at P5 layerAFPN fusion parameters P2-P3 fusion kernel3 × 3Small building detail preservation P4-P5 fusion kernel5 × 5Elongated structure capture Fusion stages3Progressive fusion levels ASF coefficients Learnable spatial weightsRRD-head configuration RepVGG branches33 × 3 conv + 1 × 1 conv + identity Training modeMulti-branchThree parallel paths Inference modeSingle-branchReparameterized 3 × 3 conv Angle normalization[-90°, 90°]Rotation angle range Author contributions Conceptualization, Yu-Zhe ZHAO and Hai-Zhong Qian; methodology, Yu-Zhe ZHAO and Hai-Zhong Qian; software, Yu-Zhe ZHAO; validation, Yu-Zhe ZHAO; resources, Hai-Zhong Qian; data curation, Yu-Zhe ZHAO; writing—original draft preparation, Yu-Zhe ZHAO; writing—review and editing, Hai-Zhong Qian; visualization, Yu-Zhe ZHAO; project administration, Hai-Zhong Qian; funding acquisition, Hai-Zhong Qian. All authors have read and agreed to the published version of the manuscript. Funding This research was funded by the National Natural Science Foundation of China, grant numbers 42271463, 42101453. The APC was funded by the National Natural Science Foundation of China, grant numbers 42271463, 42101453. Data availability The BONAI dataset is publicly available at https://github.com/jwwangchn/BONAI. DOTA v1.0, DIOR-R, and HRSC2016 datasets can be accessed through their respective official websites. The source code and trained model weights are available from the corresponding author upon reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The BONAI dataset is publicly available at https://github.com/jwwangchn/BONAI. DOTA v1.0, DIOR-R, and HRSC2016 datasets can be accessed through their respective official websites. The source code and trained model weights are available from the corresponding author upon reasonable request.