MAFD-NET: multi-head attention dynamic sampling fusion network for power tower detection in high-resolution remote sensing images

1 Introduction

Power towers are core components of power transmission networks, and their precise spatial positioning serves as a fundamental prerequisite for power inspections, operation and maintenance planning, and safety management (Savva et al., 2025). Traditional manual on-site positioning methods not only incur high labor and material costs, but also suffer from low efficiency, large positioning deviation and limited coverage, making them poorly suited for the wide-area and high-precision demands of modern power operation and maintenance. With the popularization of artificial intelligence technology, power tower detection has achieved an intelligent upgrade. Although AI and remote sensing technologies have been extensively adopted in intelligent tower inspection and disaster prevention, accurately identifying the spatial locations of power towers from high-resolution remote sensing images still poses a critical challenge. Addressing this issue is essential to guarantee the operational safety and maintenance reliability of modern power systems (Kaku, 2019).

With the rapid development of deep learning techniques, typical detection models including Faster R-CNN (Girshick et al., 2014), the YOLO series (Redmon et al., 2016) and SSD (single shot multiBox detector) (Liu et al., 2016) have been extensively applied in power tower detection tasks. Nevertheless, most existing deep learning models are not specially designed according to the characteristics of remote sensing images and the slender morphological structure of power towers. Those models still show low detection precision for small and medium-sized distribution power towers in practical use, and their detection robustness under complex backgrounds still needs further enhancement.

To address these issues, targeted improvements in existing literature have been made from three perspectives: feature fusion architecture optimization, scene adaptation enhancement, and data and theoretical support. In terms of feature fusion architectures, cascaded feature fusion strategies effectively strengthen the multi-scale feature representation of power tower detection models (Hou et al., 2020). On this basis, bidirectional feature fusion architectures further improve the fusion efficiency of multi-scale tower features (Fu et al., 2020). The shape-aware and attention-based fusion mechanisms have been introduced for power line detection, where a shape-aware block enhances feature propagation from elongated regions, and a spatial and channel attention fusion block adaptively integrates multi-level features, effectively alleviating class imbalance and boundary ambiguity issues (Zhang et al., 2026a).

For scene adaptation optimization, HighDAN, a high-resolution attention network, enhances feature discrimination performance in complex scenarios (Hong et al., 2023). As a general-purpose remote sensing foundation model, SpectralGPT further expands the generalization boundary across diverse remote sensing tasks (Hong et al., 2024). The task-specific LSKF-YOLO network is developed for tower detection, optimizing scene adaptation performance for this specific scenario (Shi et al., 2024). For the insufficient transmission line detection accuracy caused by misalignment between visible and infrared images in UAV-based power line inspection scenarios, the Hierarchical Multi-Modal Enhancement Network (HMMEN) significantly improves detection accuracy and robustness by leveraging mutual enhancement and feature alignment modules (Zhang et al., 2026b).

In terms of data and theoretical support, data augmentation strategies targeting the shortage of annotated tower samples provide essential data support for model training (Fang et al., 2021). From the perspective of SAR imaging mechanism, an analysis of the backscattering characteristics of towers and transmission lines in X-band SAR imagery clarifies the correlation between line orientation and target visibility, offering technical references for tower detection under complex weather conditions (Ma et al., 2024). Systematic research on the scattering characteristics of UHV transmission lines in spaceborne SAR imagery also establishes a theoretical basis for spaceborne SAR-based power facility monitoring (Shi et al., 2021). For optical imagery, adverse weather such as haze and low light also degrades image quality, and dehazing methods developed in intelligent transportation, such as the wavelet-based physically guided normalization network (Zhang et al., 2025) and the VNDHR variational nighttime dehazing approach (Liu et al., 2025), these methods are also instructive for power tower detection in adverse weather environments like haze.

Although a large number of object detection models with novel architectures and excellent performance have emerged in recent years, three under-addressed issues remain when general detection models are directly applied to remote sensing power facility inspection: (1) There is a lack of customized feature extraction design at the backbone level tailored to the characteristics of power towers. Most existing improvements focus on the neck feature fusion stage, making it difficult to simultaneously capture the global topological structure and local texture details of power towers from the source of feature extraction. (2) The issues of feature interference and fusion under complex backgrounds have not been fully resolved. Traditional single-branch attention modules cannot simultaneously achieve global background suppression and local feature enhancement, which easily causes feature confusion between backgrounds and targets and directly undermines the accurate identification and precise localization of power towers. (3) The upsampling process in multi-scale feature fusion generally relies on a fixed-weight strategy and lacks content awareness, which readily leads to the loss of edge details of small-scale distribution towers. Moreover, there is still no customized dynamic sampling scheme for such slender targets to date.

In light of this, this paper proposes a multi-head attention dynamic sampling fusion network (MAFD-Net) based on the YOLO11 framework (Wang and Wang, 2025), targeting the challenges of large scale variation and complex background interference in power tower detection. The main contributions of this paper are as follows:

(1) We propose the MAFD-Net to address issues such as the wide variation in power tower scales, complex backgrounds, and the high miss rate of small-scale distribution towers in remote sensing imagery. This network achieves high-precision and high-efficiency detection, providing technical support for intelligent power inspection.

(2) To tackle the challenges of capturing both global structural and local detailed features of power towers, as well as suppressing complex background interference, we design a multi-head attention feature fusion (MAFF) module. By integrating the Mamba mechanism, multi-head self-attention, and gated recurrent units, this module enhances feature representation capability, and realizes both background noise suppression and precise target feature extraction.

(3) We design an edge-preserving dynamic upsampling (EPDU) module and embed it into the neck network. Through adaptive scale-aware upsampling, it preserves effective features of power towers at different scales and reduces information loss of small targets.

(4) We evaluate the performance of the proposed method on two public datasets, Duke and Aerial-Power. Experimental results demonstrate that the proposed method achieves state-of-the-art detection performance compared with mainstream baseline detectors.

The rest of this article is organized as follows. The related work is described in Section 2. Section 3 introduces the proposed power tower detection method and describes the structure and the improved module in detail. Section 4 describes the experimental process and gives the analysis results. Section 5 draws the conclusion.

2 Related work

2.1 YOLO-based object detection in remote sensing imagery

The YOLO series (Redmon et al., 2016) has become a mainstream single-stage framework for detecting power facilities in remote sensing images, owing to its continuous evolution in real-time efficiency, multi-scale representation, and robustness to complex scenes. Early versions established the grid-based prediction paradigm but were quickly found inadequate for power towers, whose slender profiles and large scale variation led to high miss rates on small distribution towers. Subsequent generations have gradually addressed these challenges along three main directions: scale adaptation, background resilience, and computational efficiency.

For scale adaptation, YOLOv2 introduced anchor clustering and high-resolution classifier pre-training (Redmon and Farhadi, 2017), which stabilized the detection of towers with relatively fixed proportions. YOLOv3 (Redmon and Farhadi, 2018) incorporated a feature pyramid network with top-down fusion and multi-scale prediction heads, providing a stronger multi-scale representation that benefited small tower detection—though its fixed upsampling still risks smoothing away the fine edges of slender structures. YOLOv4 and YOLOv5 (Bochkovskiy et al., 2020; Liu et al., 2023) improved feature extraction in cluttered backgrounds and introduced lightweight scaling strategies, making them more suitable for environments where towers are surrounded by vegetation or buildings. YOLOv8 (Hussain, 2024) redesigned the neck with the C2f module and adopted a task-alignment learning strategy, which enhances the coordination between classification and bounding box regression, thereby improving localization accuracy for elongated towers. The latest YOLO11 further simplifies the neck fusion and refines the loss design, delivering a parameter-efficient model well-suited for batch processing of high-resolution imagery. Variants such as Gold-YOLO (Wang et al., 2023) and TPH-YOLO (Zhu et al., 2021) push small-object performance through optimized feature interaction and Transformer-based prediction heads, respectively, and both have informed the development of more capable tower detectors.

Despite these advances, a core limitation remains: the feature fusion stages of these detectors rely on generic, fixed-kernel upsampling, which functions as a static low-pass filter that cannot adaptively recover the high-frequency edge information essential for differentiating slender towers from similarly textured backgrounds. Moreover, the standard backbone progressively downsamples feature maps, naturally suppressing the axial continuity that characterizes power towers. Customizing feature extraction and designing dynamic upsampling that explicitly account for tower morphology are thus necessary steps to overcome this bottleneck.

2.2 Mamba-based attention mechanisms and sequence modeling

Beyond architectural improvements in detectors, handling the long-range dependencies and complex background clutter typical of remote sensing scenes calls for effective attention and sequence modeling mechanisms. Mamba (Gu and Dao, 2024) has emerged as a compelling alternative to self-attention by performing selective state-space modeling with linear time complexity, which is particularly attractive for high-resolution overhead imagery. Its ability to extract and update key information from long sequences without quadratic cost makes it a natural fit for large-scale power tower inspection (Bao et al., 2025).

Early Mamba models used uniform scanning that poorly matched the slender geometry of power towers and tended to overlook small instances (Yang et al., 2025). Vision Mamba (Liu et al., 2024) addressed this by introducing bidirectional state-space modeling for 2D data, enhancing feature interaction across spatial directions and improving structural representation for elongated objects. Building on this, Multi-Scale Mamba (Rong et al., 2025) constructs a multi-granularity representation through hierarchical state updates and adaptive receptive fields, enabling differentiated feature support for towers of various sizes. Selective Scan Mamba (Dong et al., 2024) further refines the representation by adaptively filtering key regions to suppress background noise from vegetation and buildings. In deployment-constrained settings, Sparse Mamba (Xing et al., 2025) reduces computational redundancy via sparse sequence sampling while preserving the model’s global modeling capacity.

These variants demonstrate clear potential, yet their state-space dynamics are not explicitly shaped to preserve the structural continuity of power towers—neither the directional edges nor the slender contours that are critical for precise localization. Furthermore, they have not been integrated with content-aware upsampling strategies, leaving a gap in the reconstruction of fine edge details during multi-scale fusion. Addressing this gap by combining linear-time global modeling with attention-driven feature enhancement and dynamic sampling offers a meaningful path toward more accurate and robust power tower detection in remote sensing imagery.

3 Proposed methods

This paper proposes the multi-head attention dynamic sampling fusion network (MAFD-Net) for remote sensing power tower detection based on YOLO11 framework, addressing the challenges of large scale variation and complex background. The overall architecture of the model is presented in Figure 1. The core innovations of MAFD-Net lie in the design of the MAFF and EPDU modules. The proposed model architecture consists of a backbone, neck, and detection head. The backbone performs basic feature extraction and enhancement of power towers through a combination of convolutional layers, C3k2 modules, and MAFF modules. Specifically, the C3k2 module, an advanced variant of the Cross Stage Partial (CSP) bottleneck, is utilized to optimize gradient flow and reduce computational redundancy while preserving rich semantic representation. In particular, the MAFF module integrates the long-sequence dependency modeling capability of the Mamba mechanism with parallel self-attention mechanism to efficiently mine local texture details of power towers while suppressing complex background interference. Additionally, the Spatial Pyramid Pooling-Fast (SPPF) module is employed at the end of the backbone. By utilizing sequential max-pooling operations, it rapidly aggregates multi-scale contextual information and significantly enlarges the network’s receptive field. The neck employs the EPDU module for dynamic sampling and concatenation of multi-scale features, and further optimizes feature fusion by incorporating the MAFF module. Finally, the detection head achieves accurate detection of power towers based on the fused features.

3.1 Multi-head attention feature fusion module

To overcome severe background interference in high-resolution optical remote sensing images, a novel MAFF module is proposed in this paper (see Figure 2). In practical grid inspections, power towers are frequently embedded within highly complex surroundings, such as intricate urban textures or dense vegetation. Standard convolutional layers operate exclusively with local receptive fields, making them highly susceptible to localized background noise that structurally resembles tower components. To address this, the core approach of the MAFF module combines the long-sequence dependency capture capability of the bidirectional Mamba (BM) block with the global feature modeling capability of the parallel self-attention mechanism (PSAM) block. By incorporating this mechanism, the module establishes global dependencies across the entire feature space. This global perspective allows the network to effectively capture the long-range topological structures of power towers while adaptively suppressing irrelevant background clutter. Finally, the concatenated features undergo channel shuffling and convolutional fusion to achieve cross-dimensional feature enhancement, providing a strongly motivated and robust solution to the challenges of complex background interference.

1) PSAM Block: The traditional single-head self-attention mechanism generates only a single attention weight distribution in its processing of features. Such a distribution tends to focus on the most salient local features within an image, thereby disregarding other potential structural and semantic cues. To address this issue, the upper branch of the MAFF module employs a Parallel Self-Attention Mechanism (PSAM), wherein the input features are simultaneously fed into three attention heads. Within each attention head, the features undergo independent linear projection layers to produce the head-specific query Q, key K, and value V, followed by independent self-attention computation, the calculation method is shown in Equation 1.where W i Q, W i K and W i V are the learnable weight matrices for the i-th head.

Q i = F W i Q , K i = F W i K , V i = F W i V (1)

To prevent excessively large dot-product results from causing gradient vanishing, the self-attention computation in this branch first calculates the inner product of Q i, K i while incorporating a scaling factor d k. The result is then normalized via the Softmax activation function and finally aggregated with V i to produce the subspace’s projected output. The calculation method of output is shown in Equation 2:

h e a d i = Softmax Q i K i T d k V i (2)

This mechanism maps features into distinct subspaces, thereby extracting mutually complementary local cues. The outputs of the three heads, denoted as h e a d i, are concatenated along the channel dimension and subsequently fed into a feed-forward neural network for deep feature integration and nonlinear transformation. The result is then reshaped back to the original spatial dimensions, yielding the upper branch output Y 1 ∈ R C × H × W.

2) BM Block: Although the upper branch can effectively extract rich local fine structures and mid-range contextual information through parallel subspaces, it remains insufficient for capturing ultra-long-range dependencies that span the entire scene in remote sensing images. To address this limitation, the lower branch of the module introduces a bidirectional Mamba (BM) Block (BMB), aiming to achieve global perception with linear complexity. The core of BMB relies on the state space model. For each time step x t of the input sequence, the model first transforms the continuous parameters A, B into discrete parameters A and B through a discretization rule. Subsequently, the evolution of the hidden state h t is driven by the state matrix and mapped to the output y t, the calculation method is shown in Equations 3, 4:

h t = A ̄ h t − 1 + B x t (3)

y t = C h t (4)

As two-dimensional remote sensing images do not inherently exhibit the strict causality of one-dimensional text sequences in the spatial domain, this paper devises a bidirectional scanning mechanism. Through the synchronous state evolution of forward and reverse sequences, a complete feature sequence y = [ y 1 , y 2 , … , y L ] is extracted. This mechanism ensures that every spatial location in the image fully aggregates global contextual cues, thereby eliminating the information blind spots introduced by unidirectional serialization. Finally, the sequence is reshaped back to the spatial dimensions to produce the lower branch output Y 2 ∈ R C × H × W.

To prevent heterogeneous features from remaining isolated due to simple channel concatenation, the module performs an interleaved reorganization upon entering the feature fusion stage, thereby compelling local cues and global representations to engage in deep cross-branch interaction. Thereafter, a 1 × 1 convolution is employed to conduct a secondary linear combination of cross-channel information and to compress the channel count back to the original dimension C, controlling the model parameter count. Finally, a residual connection is introduced to add the fused features to the original input F, stabilizing deep gradient flow and supplementing the original spatial localization information. The complete fusion process described above can be formalized as shown in Equation 5:

Z = C o n v 1 × 1 C S Y 1 ; Y 2 + F (5)

The MAFF module enables large-range feature perception without introducing large-kernel convolutions. The long-sequence modeling ability of the Mamba mechanism is combined with the global modeling capacity of hierarchical multi-head attention and the detailed feature extraction of the gated recurrent structure. This design reduces computational overhead while maintaining strong feature representation ability, so that both global structural information and local texture features of power towers can be accurately extracted from remote sensing images. It also alleviates the interference caused by complex backgrounds, making the module well suited for batch processing of high-resolution remote sensing images.

3.2 Edge-preserving dynamic upsampling module

In complex remote sensing imagery and power infrastructure scenarios, targets exhibit extreme scale variations, ranging from massive transmission structures to tiny distribution poles. Traditional upsampling methods, such as nearest neighbor interpolation and transposed convolution, adopt a globally uniform sampling strategy relying on fixed, predefined rules (as illustrated in Figure 3). These rigid methods frequently blur fine-grained spatial details and introduce checkerboard aliasing artifacts, inevitably leading to the loss of critical structural information for small and slender targets like power towers. To explicitly address this limitation and preserve morphological traits across varying scales, the Edge-Preserving Dynamic Upsampling (EPDU) module is proposed. This module purposefully abandons traditional zero-padding operations and uniform weight mechanisms. Instead, it adaptively adjusts spatial weights based on the specific local features of each target, achieving high-fidelity upscaling of edge details through kernel prediction and content-aware reassembly. This algorithmic flexibility ensures that critical structural boundaries remain intact, substantially enhancing multi-scale detection performance.

The EPDU module proposed in this study achieves higher-precision image upsampling (as illustrated in Figure 4). An input feature map X ∈ R C × 32 × 32 (where C denotes the number of input channels) has its data flow first entering the kernel predictor. This predictor performs channel reduction and interaction through a 1 × 1 convolution, subsequently injects non-linearity via the ReLU activation function, and aggregates local spatial context using a 3 × 3 convolution, thereby generating an intermediate feature tensor F i n ∈ R C × H × W, the calculation method is shown in Equation 6:

F = C o n v 3 × 3 R e L U C o n v 1 × 1 X (6)

To spatially match the target resolution, tensor F undergoes spatial rearrangement via a Pixel Shuffle operation, which periodically unfolds the information from the channel dimension onto a high-resolution grid, yielding a tensor of size 9 × 64 × 64. A softmax function is then applied along the channel dimension, producing the normalized dynamic parameter tensor P ̂ ∈ R 9 × 64 × 64. The calculation method of tensor is shown in Equation 7:

P ̂ = S o f t m a x P i x e l S h u ffl e F (7)

Tensor P ̂ stores a parameter vector of size 1 × 1 × 9 at each target coordinate point. Before performing the operation, this vector is reshaped into a 3 × 3 spatially adaptive weight matrix (e.g., W edge for edges or W b g for background). Meanwhile, to preserve fine-grained spatial information from the original features, the module employs a skip connection to feed the original input X into a patch extraction unit. This unit reversely extracts the corresponding 3 × 3 local neighborhood patches from the original dense feature map according to the target coordinate mapping, and reshapes them into an unfolded tensor X ̃ ∈ R C × 9 × 64 × 64. During the reassembly stage, the generated dynamic, location-specific weights and the corresponding local feature patches are combined via a local pointwise multiplication operator. At this logical node, the response value of the output feature map Y at an arbitrary coordinate ( x ′ , y ′ ) is determined by the inner product of the location-specific kernel and the feature patch. The calculation procedure is illustrated in Equation 8:

Y x ′ , y ′ = ∑ i , j ∈ 1,2,3 P ̂ x ′ , y ′ i , j ⋅ X patch i , j (8)

Through this mechanism, the resulting high-resolution feature map of size C × 64 × 64 not only fundamentally eliminates the periodic grid artifacts caused by uneven overlapping receptive fields in conventional transposed convolution, but also effectively alleviates the loss of fine-grained spatial information induced by deep semantic abstraction via pixel-level dynamic reassembly.

4 Experiment

4.1 Dataset

Experiments in this paper are conducted on the Duke power transmission and distribution infrastructure dataset (Bradbury et al., 2018) and the aerial-power power facility dataset (Savva et al., 2023). The Duke dataset consists of high-resolution optical remote sensing images with a spatial resolution of approximately 0.8 m. It contains two categories of targets, namely, transmission towers and distribution towers, and the scenes are predominantly suburban open areas. The aerial-power dataset comprises aerial remote sensing imagery at 0.5 m resolution, covering complex scenes with small objects such as urban and forested regions. The two datasets effectively complement each other in terms of scene distribution and target scale. A detailed comparison of their core characteristics is presented in Table 1.

TABLE 1. Statistics of experimental datasets.

To ensure consistent model inputs and reliable experimental results, uniform preprocessing is applied to both datasets. All images are resized to 640 × 640 pixels, annotation formats are standardized, and invalid samples such as blurred or target-free images are removed to generate standardized experimental data. Samples from the Duke and aerial-power datasets after preprocessing are illustrated in Figures 5, 6. To improve generalization, we use data augmentation including random cropping, rotation, brightness variation and occlusion simulation during training, which enhances the model’s adaptability to diverse scenes and target shapes.

4.2 Evaluation metrics

We evaluate the model’s detection performance using five core metrics: Precision ( P ), Recall ( R ), F1-score, mean Average Precision ( mAP 0.5 ), and mean Intersection over Union (mIoU). All results are reported as percentages, with the corresponding calculation formulas given in Equations 9–13, respectively. where, T P represents the number of correctly detected power tower samples, F P is the number of false positive background samples, F N denotes the number of missed power tower samples. N is the number of power tower categories (transmission towers and distribution tower), and M is the total number of power tower instances in the test set. mAP 0.5 refers to the mean average precision at an IoU threshold of 0.5, following the standard evaluation protocol for remote sensing object detection.

P = T P T P + F P × 100 % (9)

R = T P T P + F N × 100 % (10)

F 1 = 2 T P 2 T P + F P + F N × 100 % (11)

mAP 0.5 = 1 N ∑ i = 1 N ∫ 0 1 P R d R × 100 % (12)

m I o U = 1 M ∑ j = 1 M T P T P + F P + F N × 100 % (13)

4.3 Experimental details

In the training process, input images are resized to a resolution of 640 × 640 pixels, with a batch size of 16 and a total of 100 training epochs. The initial learning rate is set to 0.01 and dynamically adjusted through cosine annealing decay, which gradually decreases in later training stages to promote stable convergence. The weight decay coefficient is fixed at 0.0005 to further reduce parameter overfitting. All experiments are carried out on the experimental platform detailed in Table 2.

TABLE 2. Experimental environment Configuration.

4.4 Ablation experiment

To quantitatively verify the contributions of the proposed MAFB and EPDU modules to power tower detection in remote sensing imagery and clarify their synergistic mechanism, we design ablation experiments with YOLO11 as the baseline model. Precision, recall, F1-score, and mAP 0.5 are selected as the core evaluation metrics. All ablation experiments are conducted on the uniformly preprocessed and standardized combined dataset, with consistent training parameters and experimental environments to guarantee the objectivity and comparability of results, as shown in Table 3. Following the single-variable principle for fair comparison, the model generalization performance is further validated independently on the Duke and aerial-power datasets.

TABLE 3. Ablation experiment results.

The results of the ablation experiments indicate that the introduction of either module alone effectively improves the detection performance of the YOLO11 baseline model. The model’s performance achieves an additive improvement with the collaborative fusion of the two modules. When only the MAFB module was incorporated, the model achieved an average precision of 76.60%. This represents an improvement of 2.37% over the baseline model with a significant increase in recall. This module enhances the efficiency of power tower feature extraction, effectively suppresses background interference, and improves the detection performance for small-scale distribution towers. When only the EPDU module was incorporated, both the model’s recall rate and mean average precision were improved. Its scale-adaptive sampling strategy meets the detection needs of power towers of different scales, avoids feature loss in small targets. The complete MAFD-NET model integrates both modules and optimizes the feature transfer logic between them. It achieves an average precision of 78.26%, an increase of 4.03% over the baseline model with a significant increase in F1-score. This result verifies the significant synergistic effect between the two modules. The MAFB module provides more discriminative target features and the EPDU module ensures accurate focus on small-scale power tower targets. Together, they form a complementary and reinforcing effect, significantly improving the model’s detection accuracy and robustness in complex, multi-scale backgrounds.

We use Grad-CAM (Gradient-Weighted Class Activation Mapping) for model interpretability analysis, generating detection heatmaps for the YOLO11 baseline and our proposed MAFD-NET on both datasets. Visualization results (see Figures 7, 8) reveal that on the Duke dataset, MAFD-NET can focus more precisely on the main regions of power towers and exhibits more significant suppression of background interference. For the more complex aerial-power dataset, the baseline model is susceptible to interference from ground features such as forests and buildings, leading to scattered feature responses. In contrast, MAFD-NET stably focuses on power tower targets, pays greater attention to small-scale distribution towers, and yields more concentrated and accurate feature responses. These two sets of heatmaps collectively verify the synergistic effects of the MAFB and EPDU modules in feature enhancement, small-target preservation, and background suppression.

4.5 Comparison of different target detection networks

To comprehensively evaluate the overall performance and generalization capability of our proposed MAFD-NET for power tower detection in remote sensing imagery, comparative experiments are conducted on both the standardized combined dataset and the independent aerial-power test set. The compared models cover state-of-the-art remote sensing-specific optimized models. Precision, recall, F1-score, mAP 0.5, and mIoU are adopted as the evaluation metrics, with the results presented in Tables 4, 5.

TABLE 4. Performance comparison between MAFD-NET and state-of-the-art models on the duke dataset.

Bold values indicate the optimal performance.

TABLE 5. Performance comparison between MAFD-NET and state-of-the-art models on the aerial-power dataset.

Bold values indicate the optimal performance.

The experimental results demonstrate that MAFD-Net achieves state-of-the-art performance on both the standardized Duke dataset and the Aerial-power dataset, outperforming not only general-purpose object detection frameworks but also task-specific power tower detection models including SCOPE-YOLO and RCCF-YOLO. The Aerial-power dataset presents more challenging conditions with a higher density of small and occluded targets against complex backgrounds. Under these conditions, all compared models suffer noticeable performance degradation, yet MAFD-Net maintains a high m A P 0.5 value with the smallest drop, reflecting stronger generalization in realistic inspection environments. Much of this advantage can be attributed to the MAFB module’s ability to suppress background clutter and the EPDU module’s capacity to preserve fine features of small slender targets during upsampling. Working together, these two modules allow the model to sustain stable detection even in complex scenes not encountered during training.

Additionally, since the input resolution remains consistent across evaluations, the inherent complexity metrics of each model, the number of parameters (Params) and floating-point operations (FLOPs) are consolidated and presented in Table 6. As observed, our MAFD-NET achieved advanced detection accuracy while maintaining a highly competitive parameter count of 16.3 M and a computational cost of 33.7G, proving its excellent balance between precision and efficiency.

TABLE 6. Comparison of model complexity and parameters.

Figures 9, 10 illustrate the detection performance of different models on the two datasets, respectively. We can see that MAFD-NET accurately identifies power towers of various scales even in more challenging scenarios, such as residential areas, forested regions, and densely occluded environments. It yields significantly fewer missed and false detections, with more precise bounding box localization, fully demonstrating the model’s robustness and practical applicability.

Comprehensive analysis of the experimental results verifies that the MAFD-NET model exhibits outstanding overall detection performance, following extensive validation against a diverse set of comparative models with distinct core architectures. The model achieves state-of-the-art detection accuracy across all experiments, yielding optimal results for all five core evaluation metrics, and delivers consistent performance gains over current models. These results fully validate the practical application value of the core improved modules proposed in this study. MAFD-NET performs exceptionally well in bounding box regression for multi-scale targets, generating precise and reliable localization results that satisfy the high-precision requirements for power tower positioning in power line inspections. Furthermore, the model features strong scene adaptability, effectively addressing the practical challenges of significant scale variations in power towers and severe background interference in remote sensing imagery, and maintains low miss and false detection rates across various complex scenarios. Leveraging the collaborative design of multi-head attention feature fusion and dynamic sampling modules, MAFD-NET constructs an efficient network architecture tailored for power tower detection in remote sensing imagery. Its superior comprehensive detection performance outperforms traditional and mainstream detection models, thus providing an efficient and reliable technical solution and theoretical support for the intelligent remote sensing inspection of power facilities in large-scale, complex scenarios.

4.6 Cross-validation results

Furthermore, to comprehensively eliminate the potential bias introduced by a single random dataset split and to rigorously verify the statistical reliability of MAFD-NET, we conducted a 5-fold cross-validation on both the Duke and Aerial-power datasets. As shown in Table 7, our model demonstrates exceptional stability across different data distributions. Specifically, MAFD-NET achieves an average m A P 0.5 of 78.26% with a minimal standard deviation of ± 0.15% on the Duke dataset, and an average m A P 0.5 of 76.74% ± 0.05% on the Aerial-power dataset. This extremely low variance statistically confirms that the superior performance of MAFD-NET is robust and independent of specific training sample selections, fully satisfying the stringent reliability requirements for intelligent power grid inspection.

TABLE 7. 5-Fold cross-validation results of MAFD-NET on duke and aerial-power datasets.

Bold values indicate the optimal performance.

5 Conclusion and future work

This paper proposes the MAFD-Net detection network based on YOLO11, which enhances the feature extraction and multi-scale adaptation capabilities for power towers via the MAFB and EPDU modules. The MAFB module fuses the Mamba architecture with a multi-head self-attention mechanism to realize long-range feature perception without relying on large-kernel convolutions, effectively enhancing target features and mitigating background interference simultaneously. The EPDU module adopts adaptive scale perception and a dual-branch sampling strategy to extract global features for large-scale targets and detailed features for small-scale targets, respectively, which alleviates the small-target feature loss introduced by downsampling operations. These two modules collaborate synergistically to achieve the dual goals of feature enhancement and scale adaptation, which provide core support for the improvement of the model’s overall performance.

Extensive experiments are conducted on two public datasets, Duke and Aerial-Power, to verify the effectiveness of our design. Ablation studies demonstrate that both the MAFB and EPDU modules significantly boost the performance of the YOLO11 baseline; their fusion enables the model to achieve an mAP 0.5 of 78.26%, a 4.03 percentage point improvement over the YOLO11 baseline. Compared with state-of-the-art mainstream detection models, the proposed MAFD-Net yields optimal results across multiple metrics, including precision, recall, F1-score, and mIoU, and effectively balances detection accuracy and computational efficiency. The experimental results reveal that the model exhibits more concentrated feature responses in target regions, allocates greater attention to small-scale towers, and demonstrates enhanced anti-interference robustness in complex backgrounds. These capabilities directly contribute to reliable geospatial power asset inventory construction and the elimination of inspection blind spots, thereby supporting equipment condition monitoring and risk-oriented maintenance planning for smart grid systems.

In future work, we will continue to refine the model architecture by addressing several inherent limitations of the proposed MAFD-Net. One direction is to introduce super-resolution reconstruction modules that can strengthen feature representation for extremely small towers, helping to overcome the persistent detection bottleneck of ultra-small distribution towers. We also intend to improve the Mamba sequence sampling strategy and the multi-head attention mechanism, which should further lower the miss rate for small towers and sharpen the capture of their slender structural contours. Beyond module-level improvements, power line topological priors can be incorporated into a joint detection framework; geometric constraints between transmission towers and lines would then help suppress false alarms in complex scenes where background textures closely resemble tower features. To boost robustness under diverse environmental conditions, we will explore multi-spectral feature fusion. In parallel, model lightweight compression techniques will be investigated to satisfy the low-latency deployment requirements of edge devices carried by UAVs. Through these efforts, we aim to enhance the practicality and engineering applicability of the proposed algorithm and accelerate its large-scale deployment in real-world power transmission line inspection systems.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Feng Liu, Nanjing Tech University, China

Reviewed by: Shengdong Zhang, Wenzhou University, China

Dachuan XU, Beijing Normal University, China