DBF-YOLO: a lightweight dual-branch fusion method for Visible-SAR cross-modal remote sensing target detection

1 Introduction

With the rapid development of remote sensing platforms and multi-source sensor technologies, object detection has become a key task in the intelligent interpretation of remote sensing images. It has demonstrated extensive application value in areas such as ocean monitoring, port supervision, disaster assessment, land resource surveys, and military reconnaissance. Compared with natural-scene object detection, remote sensing object detection usually faces challenges such as large scale variations, complex back-grounds, dense target distributions, and blurred object boundaries, which impose higher requirements on feature representation and localization accuracy (Li K. et al., 2020). At the same time, remote sensing interpretation is evolving from single-modal analysis toward multimodal imagery interpretation, indicating that multi-source sensing has become an important direction for robust scene understanding (Sun et al., 2023). Furthermore, with the gradual deployment of remote sensing intelligent interpretation algorithms on spaceborne, airborne, and other edge devices, models must also consider parameter scale, computational overhead, and real-time inference capability. This makes lightweight design not only a technical objective for model compression but also a key demand for practical applications. In recent years, single-stage detectors represented by the YOLO family have become important baseline frameworks for real-time object detection because of their favorable trade-off between accuracy and efficiency (Redmon et al., 2016). Single-shot detection further demonstrates that efficient detection can be achieved without the heavy proposal-generation process used in two-stage frameworks (Liu et al., 2016). More recently, YOLOv10 has provided an end-to-end real-time detection framework, making it a suitable lightweight baseline for this study (Wang et al., 2024).

However, in complex remote sensing scenarios, a single modality often struggles to provide stable and complete target information. Visible images offer rich texture, color, and appearance details, which provide semantic support for object recognition but are susceptible to weak light, cloud cover, and complex weather conditions. In contrast, SAR images provide all-weather and day-and-night imaging capability and can represent the scattering structure and geometric contours of targets, but they usually suffer from speckle noise and limited texture expression. Optical-SAR fusion studies indicate that the two modalities are complementary in texture description, spectral information, geometric structure, and observation conditions (Zhang R. et al., 2026). Deep learning-based multimodal remote sensing fusion further shows that heterogeneous sensors can improve scene interpretation when their complementary representations are effectively integrated (Li et al., 2022). For object detection, cross-modal attention has also been shown to strengthen target representation by enabling adaptive interaction between modalities (Zhao P. et al., 2024). Nevertheless, the structural information provided by SAR is not always directly usable: local-statistical filtering is often needed to suppress noise while preserving local structures (Lee, 1980), and speckle noise remains a typical factor that weakens the readability and structural stability of SAR images (Lee, 1981). Therefore, how to effectively utilize visible texture details and SAR structural information has become an important issue in cross-modal remote sensing target detection.

Existing visible-SAR and optical-SAR cross-modal detection methods have improved detection performance to some extent, but there are still several shortcomings when deployed in real-world remote sensing scenarios. First, the two modalities have significant differences in statistical distribution, texture organization, and semantic expression. Simple concatenation, fixed weighting, or shallow fusion strategies may introduce feature conflicts and information redundancy rather than stable collaboration. Second, the speckle noise and edge degradation in SAR images weaken their structural expression ability, thus limiting the effect of cross-modal fusion. Furthermore, practical visible-SAR observations may suffer from modality-quality fluctuations and local degradation. Reliable geometric correspondence is important for stable cross-modal feature interaction, and related studies have shown that registration quality can affect the effectiveness of optical-SAR feature interaction (Chen et al., 2025; Zhang W. et al., 2026). The datasets adopted in this study provide accurately registered visible-SAR image pairs; accordingly, the experiments reported in this paper evaluate modality complementarity under registered conditions, with particular emphasis on modality-quality variation and heterogeneous feature representations. Under degraded optical conditions such as cloud occlusion, the reliability of the two modalities may change substantially (Ding et al., 2025). These variations make fixed fusion strategies more likely to propagate unreliable information and weaken detection stability. Therefore, improving the structural usability of SAR information under degradation, adapting cross-modal fusion to changes in modality quality and heterogeneous representations, and controlling model complexity are key issues in visible-SAR target detection. These factors make fixed fusion strategies more likely to propagate unreliable information and weaken the stability of the detection process. Therefore, improving the structural usability of the SAR modality under degradation conditions, enhancing the adaptability of cross-modal fusion to modality quality changes and heterogeneous modality representations, and controlling model complexity have become key issues in visible-SAR target detection.

To address these challenges, we propose a lightweight dual-branch fusion method for visible-SAR cross-modal remote sensing target detection, DBF-YOLO, aimed at improving detection accuracy while maintaining model efficiency. The main contributions and technical innovations of this paper are summarized as follows.

• This paper introduces DBF-YOLO, a lightweight dual-branch fusion framework based on YOLOv10. Different from conventional single-stream or simple concatenation-based visible-SAR fusion strategies, DBF-YOLO preserves modality-specific feature extraction and establishes progressive cross-modal interaction through multi-scale fusion paths. The model introduces a Multi-scale Cross-modal Fusion Path (MCFP) at three levels (P3, P4, and P5) to enable progressive cross-modal feature interaction.
• A SAR Gradient Enhancement Module (SGM) is designed to improve the usability of SAR structural information before feature extraction. Unlike directly feeding noisy SAR inputs into fusion networks, SGM introduces denoised intensity, gradient magnitude, and gradient direction priors to provide more stable structural representations for cross-modal learning.
• An Adaptive Gated Dual-Modal Fusion Module (AGD) is proposed to perform adaptive modality weighting. Unlike fixed fusion or static weighting strategies, AGD dynamically adjusts visible and SAR contributions according to spatial and channel-level reliability differences.

The rest of the paper is structured as follows. Section 2 reviews related work. Section 3 details the proposed DBF-YOLO model, explaining its structural framework and operation. Section 4 presents experiments that validate the effectiveness and generalization of the proposed method. Finally, Section 5 summarizes the conclusions and outlines potential future research directions.

2 Related work

2.1 Visible-SAR cross-modal remote sensing target detection and fusion methods

Visible and SAR modalities have significant complementary advantages in remote sensing object detection. Visible images can provide rich texture, color, and appearance details, but are prone to information degradation under low light, cloud cover, and complex weather conditions. SAR images, on the other hand, possess all-weather, all-day imaging advantages and are more sensitive to target geometric structures and scattering features. However, SAR images also suffer from strong speckle noise, sparse texture, and insufficient semantic readability. Therefore, visible-SAR cross-modal object detection has become an important research direction in multimodal remote sensing object detection in recent years. Relevant survey studies indicate that the complementary advantages of visible and SAR modalities mainly lie in texture-structure collaboration, and current performance bottlenecks are focused on modality differences, registration errors, and efficient fusion (Gomez-Chova et al., 2015).

Existing cross-modal detection and fusion methods can be broadly categorized into data-level fusion, feature-level fusion, and task-coupled fusion. Data-level fusion directly concatenates or transforms the two modalities at the input stage, preserving more raw information but remaining sensitive to distribution differences and spatial misalignment. Feature-level fusion, which usually uses dual-branch encoding and intermediate feature interaction, has become a mainstream solution because it can model texture-structure complementarity during representation learning. Zhao et al. proposed MMFDet with multi-branch feature extraction and multimodal difference-complement modeling, showing that difference-aware representation is beneficial for multimodal remote sensing object detection (Zhao et al., 2025).

Recent studies have also shifted from general fusion toward task-driven fusion. Xu and Yang proposed JFDet and showed that coupling image fusion with detection loss can make multimodal fusion more consistent with downstream object detection requirements (Xu and Yang, 2026). Yang et al. investigated long-tailed object detection for multimodal remote sensing images, indicating that class imbalance and target distribution differences also affect multimodal detection robustness (Yang et al., 2023). Tang et al. introduced a semantic-aware fusion strategy in the loop of high-level vision tasks, suggesting that fusion quality should be evaluated by its contribution to semantic tasks rather than visual appearance alone (Tang et al., 2022). Liu et al. proposed a target-aware dual adversarial learning framework and a multi-scenario benchmark for infrared-visible object detection, further demonstrating the importance of detection-oriented multimodal fusion evaluation (Liu et al., 2022).

In summary, visible-SAR cross-modal detection is developing along three directions: explicitly modeling modality differences rather than relying on simple concatenation; optimizing fusion for downstream detection rather than visual fusion quality alone; and balancing robustness, localization accuracy, and deployment efficiency. Bahaduri et al. introduced a multimodal transformer with cross-channel attention for remote sensing object detection, showing that channel-wise cross-modal interaction can improve heterogeneous feature modeling (Bahaduri et al., 2023) (Bahaduri et al., 2023). Motivated by these studies, this paper adopts a dual-branch YOLOv10 framework and introduces SAR structural enhancement and adaptive gated fusion to improve cross-modal collaborative detection while keeping the model lightweight. Different from existing methods that mainly focus on fusion architecture complexity or attention-based interaction, this study focuses on lightweight cross-modal collaboration by jointly considering SAR structural enhancement, progressive multi-scale interaction, and adaptive modality weighting.

2.2 Structural enhancement and dynamic feature fusion methods

Beyond fusion location and overall architecture, structural enhancement and dynamic feature fusion are key factors that influence visible-SAR cross-modal detection performance. In real remote sensing scenes, modality quality fluctuation, local degradation, and spatial misalignment can enlarge representation differences between heterogeneous modalities. Therefore, fusion must not only decide where to combine modalities, but also determine how to suppress unreliable features and emphasize complementary cues under local quality changes.

For visible-SAR detection, dynamic feature fusion is important because the reliability of the two modalities is not fixed. Optical images are usually more informative in texture, color, and appearance details, whereas SAR images provide more stable structural and scattering cues under low illumination, occlusion, and cloud cover. Sun et al. used uncertainty-aware learning for RGB-infrared cross-modality vehicle detection, showing that modality reliability can vary across scenes and should be explicitly considered during fusion (Sun et al., 2022). Fixed concatenation, simple addition, or static weighting can therefore introduce redundant or degraded information, causing feature conflict and reducing detection performance.

The quality of single-modality features also determines whether dynamic fusion can be effective. For SAR images, the main advantage lies in stable edges, contours, scattering structures, and geometric information rather than rich texture. However, these features are easily affected by multiplicative speckle noise and local degradation. If SAR structural information is not enhanced before fusion, a gating or attention module may still operate on noisy and unstable features. This observation motivates the SGM design in this paper, which enhances SAR structural cues before they enter the du-al-branch backbone.

Dynamic fusion has consequently become a core mechanism for multimodal object detection. Vaswani et al. established the self-attention mechanism as a general way to model long-range dependency and adaptive feature interaction (Vaswani et al., 2017). Woo et al. proposed CBAM, showing that channel and spatial attention can refine feature responses with limited additional complexity (Woo et al., 2018). Guo et al. designed background-aware cross-attention multiscale fusion for multispectral object detection, indicating that attention can help suppress background interference during multimodal fusion (Guo et al., 2024). Chen et al. used probabilistic ensembling for multimodal object detection, showing that uncertainty-aware combination can improve reliability under modality variation (Chen et al., 2022). Wang et al. proposed MRLF for optical and SAR images and demonstrated that multi-resolution layered fusion can strengthen cross-modal representation (Wang et al., 2025). Hao et al. introduced LASFNet, suggesting that lightweight attention-guided self-modulation can improve multimodal object detection without excessive computational burden (Hao et al., 2025).

3 Methodology

3.1 Overall framework

The DBF-YOLO (Dual-Branch Fusion YOLO) model proposed in this paper adopts a dual-branch cross-modal detection framework based on YOLOv10. The overall architecture, as detailed in Figure 1, consists of four components: the SAR input enhancement module, dual-branch backbone network, multiscale cross-modal fusion structure, and the detection head. For the input visible and SAR images, the model first performs unified scaling. The visible image is then directly fed into the RGB branch, while the SAR image undergoes denoising and structural enhancement through the SGM module before being passed to the SAR branch. Both branches extract multiscale features and perform intermediate fusion at the P3, P4, and P5 levels using the AGD module, ultimately passing through the detection head for target classification and bounding box regression. This framework preserves independent encoding capabilities for both modalities, while allowing for progressive cross-modal feature interaction from shallow to deep layers.

At the input stage, the original visible and SAR images undergo unified scaling. For the visible data, it is directly fed into the RGB branch of the backbone network. For the SAR data, it first undergoes enhancement via the SGM module, where the denoised intensity, gradient magnitude, and gradient direction information are jointly encoded and then fed into the SAR branch, providing richer structural priors for the subsequent network.

In the feature extraction stage, the network uses a dual-branch structure. The visible branch primarily learns target texture, shape, and spectral detail features, while the SAR branch focuses on stable scattering structures and edge contour representations. Both branches progressively downsample along their respective hierarchical backbones, obtaining mid-to-high-level feature representations at the P3, P4, and P5 scales. Since the imaging mechanisms of the two modalities are different, directly mixing them at the input stage can lead to feature coupling confusion. Thus, the dual-branch independent encoding structure is preserved to enhance the specificity of feature expression.

In the feature fusion stage, instead of using single-point fusion or end-to-end decision fusion, this method performs intermediate fusion at three levels: P3, P4, and P5. The P3 fusion focuses more on shallow local texture and edge responses, which is crucial for small target detection; the P4 fusion balances resolution and semantic information, serving as an important bridge between shallow details and deep semantics; the P5 fusion targets high-level abstract semantics and large target structural expressions, helping to enhance discriminability in complex backgrounds. The three-layer fusion forms a progressive cross-modal information interaction path, enabling the net-work to collaboratively model visible and SAR features at different scales.

In the detection stage, the fused multiscale features are fed into the standard YOLO detection head for further aggregation and prediction output. The detection head performs upsampling, concatenation, and convolution operations to facilitate cross-scale information flow, ultimately completing target classification and bounding box regression at multiple scales. This creates a complete closed-loop network, from input enhancement, feature extraction in both branches, and hierarchical fusion to multiscale detection.

3.2 Multi-scale cross-modal fusion design

Existing multimodal detection methods often rely on single-layer fusion or late decision fusion, which can struggle to balance local details, transitional semantics, and high-level discriminative information (Fang et al., 2021). To address this problem, this study introduces cross-modal fusion paths at the P3, P4, and P5 levels, enabling progressive information exchange across multiple scales. The structure of these fusion paths is illustrated in Figure 2.

Specifically, fusion nodes are set at the P3, P4, and P5 levels. The design rationale is that different scale features carry different semantic responsibilities, and thus, the fusion requirements vary. The P3 feature map has a higher spatial resolution, preserving more edge, texture, and local geometric information, which is suitable for complementing small target and fine-grained structural discrimination features. The P4 features, located in the middle layer, balance spatial resolution and semantic abstraction, making them better suited for coupling local features of the two modalities and mid-level semantics. The P5 feature map has the lowest resolution but strongest semantics, making it ideal for fusing global context and overall target structural information, thus enhancing the model’s ability to identify large targets and fuzzy targets in complex scenes.

Therefore, the cross-modal fusion path structure is not a simple repetition of the same fusion operation but a progressively deepening cross-modal collaborative mechanism. P3 fusion emphasizes detail supplementation, P4 fusion emphasizes semantic transition, and P5 fusion emphasizes high-level discriminative enhancement. This layered fusion method effectively avoids two extremes: first, premature fusion that directly couples noise and statistical differences between modalities, weakening feature independence; second, fusion that occurs only at the decision layer, which fails to fully utilize complementary information in intermediate representations. By establishing fusion paths at three critical nodes, the network can efficiently improve multimodal in-formation utilization while maintaining clear structural integrity.

From an overall perspective, the cross-modal fusion path structure allows the model to not only learn the complementary relationship between different modalities but also learn how this relationship varies at different scales. This makes it better suited to address common challenges in remote sensing target detection, such as large-scale variations, complex backgrounds, and blurred target boundaries.

3.3 SAR Gradient Enhancement Module

SAR images often contain multiplicative speckle noise, and target edges and local structures can be degraded by both noise amplification and sparse texture, which weakens subsequent feature extraction and cross-modal fusion (Dalal and Triggs, 2005). In visible-SAR cross-modal detection, the main advantage of SAR imagery lies in its stable structural information, including target contours, scattering boundaries, and geometric patterns, rather than rich texture details. However, speckle noise and local intensity fluctuations may reduce the reliability of these structural representations. Therefore, enhancing the structural usability of SAR inputs before cross-modal interaction is important for improving the effectiveness of multimodal feature fusion.

To address this issue, we propose the SAR Gradient Enhancement Module (SGM) to improve the structural usability of SAR inputs before cross-modal feature interaction. Rather than serving as an independent preprocessing operation, SGM provides enhanced structural priors for the SAR branch by jointly encoding denoised intensity, gradient magnitude, and gradient direction information. As shown in Figure 3, SGM consists of local-statistical filtering, gradient extraction under smoothing constraints, and multi-channel structural encoding.

Given the input SAR feature map X s a r ∈ R B × C × H × W, where B represents the batch size, C represents the number of input channels, and H and W represent the spatial height and width of the feature map, respectively. The output after processing by the SGM is denoted as F s g m. For the single-channel SAR input used in this paper, the final output of the SGM typically consists of the denoised intensity map, gradient magnitude map, and gradient direction map. Therefore, the number of output channels is expanded from 1 to 3.

3.3.1 Local statistical filtering and structural enhancement modeling

For the multiplicative speckle noise in SAR images, we use an adaptive filtering strategy based on local statistics. This strategy enhances the smoothing ability in uniform regions and retains the original scattering information in edge and structural mutation areas, achieving a balance between denoising and edge preservation. The local statistical filtering process is formulated in Equations 1–4.

Let the pixel value at position x , y in the input image be denoted as I x , y, and the local neighborhood window centered at x , y is denoted as W x , y, with a window size of N. The pixel position within the window is denoted as x i , y i. The local mean and variance at position x , y are defined as:

Let the pixel value at position p in the input image be X s a r p, and the local neighborhood window centered at p be denoted as Ω p, with a window size of k × k. The position of each pixel in window Ω p is denoted as. The local mean at position μ p and the local variance at position σ p 2 are defined as: where μ p represents the average scattering intensity within the neighborhood at position p, σ p 2 denotes the local fluctuation intensity within the neighborhood, and Ω p = k 2 represents the number of pixels within the window. Additionally, to estimate the overall noise level, the global noise variance estimate is denoted as σ n 2 .

μ p = 1 Ω p ∑ q ∈ Ω p X s a r q (1)

σ p 2 = 1 Ω p ∑ q ∈ Ω p X s a r q 2 − μ p 2 (2)

Based on this, the adaptive weight at position p is defined as w p, and is given by:

w p = σ p 2 σ p 2 + σ n 2 (3)

Furthermore, the filtered output X ∼ s a r can be expressed as:where w p ∈ 0 , 1 reflects whether the current position should retain the original input or tend more towards the local mean. Specifically, when σ p 2 is small, it indicates that the local region is flatter, and w p is small, causing the output to be closer to μ p, thereby enhancing smoothing and denoising. When σ p 2 is large, it indicates that the local region is more likely to contain edges, contours, or scattering mutations, and w p increases, causing the output to be closer to the original input X s a r p, thereby pre-serving key structural information.

X ∼ s a r p = μ p + w p X s a r p − μ p (4)

Thus, the first stage of the SGM achieves an adaptive balance between denoising and edge preservation, providing more stable input for the subsequent extraction of structural priors.

3.3.2 Gradient-aided structural encoding

After obtaining the denoised result X ∼ s a r, the module further extracts its edge gradient information. Considering that residual noise may still be amplified by the gradient operator, X ∼ s a r is first subjected to lightweight Gaussian smoothing, and the smoothed result is denoted as: The gradient-aided structural encoding and multi-channel output are described in Equations 5–10.where * represents the 2D convolution operation, and G represents the Gaussian smoothing kernel.

X ∼ s a r = X ∼ s a r * G (5)

Next, the local gradient responses are computed using the S o b e l convolution kernels in the horizontal and vertical directions: where S x and S y represent the S o b e l convolution kernels in the horizontal and vertical directions, respectively; G x and G y represent the local variation rates of the image in the two orthogonal directions.

G x = X ∼ s a r * S x (6)

G y = X ∼ s a r * S y (7)

According to the gradient responses in the two directions, the gradient magnitude map and the gradient direction map can be further calculated as: where M denotes the gradient magnitude map, which is used to characterize the edge response intensity; Θ denotes the gradient direction map, which is used to characterize the dominant direction of the local structure; and arctan ⁡ 2 · denotes the four-quadrant inverse tangent function, whose output range is typically ( − π , π. Compared with using intensity information alone, M and Θ can explicitly complement structural priors such as geometric contours and directional distributions, and are therefore more suitable for providing stable structural representations for the sub-sequent backbone network.

M = G x 2 + G y 2 (8)

Θ = arctan ⁡ 2 G x , G y (9)

3.3.3 Multi-Channel Structural Encoding Output.

After completing local statistical filtering and gradient-assisted structural extraction, the SGM concatenates the denoised intensity map, the gradient magnitude map, and the gradient direction map along the channel dimension to obtain the final enhanced feature F s g m:where F s g m denotes the enhanced feature output by the SGM module; C o n c a t · denotes the concatenation operation along the channel dimension; and X ∼ s a r, M, and Θ denote the denoised intensity map, the gradient magnitude map, and the gradient direction map, respectively. For single-channel input, the number of output channels is expanded from 1 to 3.

F s g m = C o n c a t X ∼ s a r , M , Θ (10)

This design does not change the spatial resolution, but instead jointly encodes scattering intensity, edge strength, and directional priors at the input level, enabling the SAR branch to obtain a more stable and structurally discriminative representation before entering the dual-branch backbone network. As illustrated in Figure 3, the output of SGM is not used independently for prediction, but rather serves as the input to the SAR backbone branch, which then interacts with the visible branch features at the P3, P4, and P5 scales through intermediate fusion. Therefore, its essential role is to provide a higher-quality structural foundation for subsequent cross-modal fusion.

The three components generated by SGM provide complementary structural descriptions of SAR imagery. Specifically, the denoised intensity map preserves the overall scattering distribution and suppresses the influence of speckle noise, the gradient magnitude map emphasizes target boundaries and local structural changes, and the gradient direction map captures geometric orientation information. By jointly encoding these complementary representations, SGM provides more reliable structural priors for the SAR branch and facilitates more effective interaction with visible features during cross-modal fusion.

3.4 Adaptive Gated Dual-Modal Fusion Module

Visible and SAR features differ substantially in statistical distribution, response pattern, and discriminative capability. Therefore, fixed concatenation or direct weighting is insufficient to capture the spatially varying contributions of the two modalities (Perez et al., 2018). To address this issue, this paper proposes an Adaptive Gated Dual-Modal Fusion (AGD) module. The primary purpose of AGD is to perform adaptive modality weighting rather than generating explicit attention visualization. By learning spatial- and channel-dependent gating weights, AGD dynamically adjusts the contribution of visible and SAR features during fusion according to their relative reliability. The module first applies independent projections to the features of the two modalities to reduce distribution discrepancies and achieve alignment in a latent space. The aligned features are then concatenated, and a lightweight gating network generates modality weighting coefficients for adaptive feature integration. Finally, the gating weights are used to perform adaptive weighted fusion of the dual-modal features, while convolutional refinement and residual connections are further introduced to enhance the discriminability and stability of the output features. The overall process is illustrated in Figure 4. By dynamically adjusting the fusion ratio between texture information and structural information according to scene content, this module improves the collaborative effect of cross-modal feature fusion.

Given the input visible feature map F v i s ∈ R B × C × H × W and SAR feature map F s a r ∈ R B × C × H × W, where B represents the batch size, C represents the number of channels, and H and W represent the spatial height and width of the feature maps, respectively; the fused features output by the module are denoted as F o u t ∈ R B × C × H × W. Since AGD is embedded at the P3, P4, and P5 levels, the above notation applies to the intermediate fusion nodes at different scales.

3.4.1 Feature alignment and latent space mapping

Due to the significant differences in the imaging mechanisms of visible and SAR, the two modalities often exhibit inconsistencies in statistical distribution, response patterns, and feature magnitude ranges. To improve the stability of subsequent gating learning, independent projection mappings are first performed for each modality: The complete formulation of the AGD module is presented in Equations 11–18. where F v i s ′ and F s a r ′ represent the aligned features of visible and SAR after projection, respectively, and ϕ v i s and ϕ s a r denote the projection mapping functions for the two modalities. Based on the actual implementation, ϕ · can be understood as a lightweight projection subnet composed of 1 × 1 convolution, batch normalization, and activation functions. This step does not change the spatial resolution; its purpose is to map the two modalities into a more comparable and fusion-friendly latent feature space.

F v i s ′ = ϕ v i s F v i s (11)

F s a r ′ = ϕ s a r F s a r (12)

3.4.2 Joint Representation and Gating Weight Generation.

After obtaining the aligned dual-modal features, they are first concatenated along the channel dimension to obtain the joint representation:where F C ∈ R B × 2 C × H × W denotes the joint dual-modal feature, and C o n c a t · denotes the concatenation operation along the channel dimension. This joint representation enables the gating network to simultaneously observe visible texture information and SAR structural information.

F C = C o n c a t F v i s ′ , F s a r ′ (13)

On this basis, the gating generation network ψ · is used to perform local context modeling and channel compression on the joint representation, generating modality-specific gating weights for each channel and spatial location:where W ∈ 0 , 1 B × C × H × W denotes the modality gating weight representation, ψ · denotes the gating generation network, and σ · denotes the S i g m o i d activation function. The gating generation network is composed of a depthwise separable convolution, batch normalization, an activation function, and two 1 × 1 convolutions. Specifically, the depthwise convolution is used to introduce local spatial context information, after which channel compression is performed to form a more compact intermediate gating representation, and finally a weight map with the same number of channels as the input is generated. Each element in W represents the retention ratio of the SAR modality at the corresponding channel and spatial location, whereas 1 − W represents the retention ratio of the visible modality.

W = σ ψ F C (14)

3.4.3 Adaptive weighted fusion

Based on the generated gating weights, the dual-modal features are fused through element-wise weighted fusion:where F f u s e d denotes the weighted fused feature, ⊙ denotes element-wise multiplication, W denotes the gating weight corresponding to the SAR modality, and 1 − W denotes the complementary weight corresponding to the visible modality.

F f u s e d = W ⊙ F s a r ′ + 1 − W ⊙ F v i s ′ (15)

This formulation indicates that, at any spatial location and in any channel, the fused result is jointly composed of SAR features and visible features, with their proportions dynamically determined by the gating weights. When the visible modality is affected by low illumination, occlusion, or texture degradation, the model tends to increase, thereby enhancing the structural contribution of SAR. Conversely, in regions with clear texture, 1 − W correspondingly increases to preserve more visible details. Since AGD is deployed at the intermediate fusion nodes of the P3, P4, and P5 scales, this mechanism is able to learn the modality allocation strategies required for detail supplementation, semantic transition, and high-level discriminative enhancement at different semantic levels.

3.4.4 Feature refinement and residual output

To further enhance the discriminability of the fused features and reduce information loss, AGD introduces two branches: residual projection and convolutional refinement, after adaptive weighting. First, the joint feature F C undergoes residual projection:where F r denotes the residual branch output, and ρ · denotes the residual projection function, typically consisting of 1 × 1 convolution, which maps the joint feature from the channel to C channels.

F r = ρ F C (16)

Next, the weighted fused features undergo convolutional refinement:Where F e denotes the refinement branch output, and R · represents the feature refinement mapping, composed of 3 × 3 convolution, normalization, activation, and 1 × 1 convolution.

F e = R F f u s e d (17)

Finally, the output feature representation of AGD is denoted as.where F o u t represents the final output of the AGD module. This design allows the residual branch to retain the direct information flow from the original joint representation, while the refinement branch further models the local nonlinear relationships in the fused features. The combination of the two helps enhance the stability and robustness of the output features. As shown in Figure 4, the output of AGD is not directly used for final prediction but serves as the intermediate fusion result for the corresponding layer, which is then passed to the subsequent Neck and Detect heads. Therefore, its primary function is to improve the quality of cross-modal information interaction at the P3, P4, and P5 key nodes of the dual-branch backbone in a lightweight manner.

F o u t = F e + F r (18)

3.5 Loss function

The training objective of this detection framework adopts a joint optimization form used in object detection tasks, consisting of three components: bounding box regression loss, object classification loss, and distribution focal loss, denoted as L b o x, L c l s, and L d f l, respectively (Li X. et al., 2020; Zheng et al., 2020; Lin et al., 2017). The overall loss L t o t a l function can be expressed as The overall training objective and its three loss components are defined in Equations 19–23.where L t o t a l denotes the total loss; L b o x, L c l s and L d f l represent the weight coefficients for bounding box regression loss, classification loss, and distribution focal loss, respectively. In the experiments of this paper, these values are set to 7.5, 0.5, and 1.5, respectively, following common YOLO-style detectionloss balancing practice and the localization/classification/distribution modeling losses used in dense object detection.

L t o t a l = λ b o x L b o x + λ c l s L c l s + λ d f l L d f l (19)

3.5.1 Bounding box regression loss

In the localization branch, the bounding box regression loss based on Complete-IoU (CIoU) is used to simultaneously constrain the overlap, center point distance, and aspect ratio consistency between the predicted box and the ground truth box. Let the predicted box and the ground truth box be denoted as B p and B g, respectively. Then, the bounding box regression loss can be expressed as:where F represents the foreground positive sample set; ω i represents the weight corresponding to the i-th positive sample; S is the normalization factor corresponding to the total target score; C I o U · denotes the Complete-IoU measure between the predicted and ground truth boxes.

L b o x = 1 S ∑ i ∈ F ω i 1 − C I o U B p i , B g i (20)

Compared to the traditional IoU loss, which only considers overlap area, CIoU comprehensively measures the quality of the predicted box from three aspects: overlap, center point distance, and aspect ratio consistency. This makes it more effective in improving the accuracy of target boundary localization, especially for precise target identification tasks under strict IoU thresholds.

3.5.2 Object classification loss

In the classification branch, binary cross-entropy loss is used to supervise the category predictions at each candidate position. Let the prediction value for the i-th candidate position in the c-th class be s i c, and the corresponding supervised target score be s ^ i c. Then the classification loss can be written as: where B C E · represents the binary cross-entropy loss; s i c represents the category prediction value; s ^ i c represents the supervised target score assigned to that position; S is the normalization factor, used to mitigate the impact of variations in the number of positive samples on the stability of the loss value. This loss term is used to constrain the network’s ability to discriminate between target categories, allowing the model to more accurately distinguish between different categories of targets, such as bridges, ports, and oil tanks, in complex backgrounds.

L c l s = 1 S ∑ i ∑ c B C E s i c , s ^ i c (21)

S = max ∑ i ∑ c s ^ i c , 1 (22)

3.5.3 Distribution focal loss

To further improve the fine-grained expression capability of bounding box regression, this paper introduces the Distribution Focal Loss (DFL) in the localization branch, which models the discrete distribution of boundary positions. Let the true continuous position of a boundary be denoted as, and its left and right adjacent discrete integer positions be y l = y and y r = y l + 1, with corresponding interpolation weights W l = y l − y and W r = 1 − W l. Then, the DFL can be expressed as:where: C E · represents the cross-entropy loss; p i represents the predicted distribution of the i-th positive sample at the discrete position; W l and W r represent the linear interpolation weights of the true position relative to the left and right discrete points, respectively.

L d f l = 1 S ∑ i ∈ F ω i W l C E p i , y l + W r C E p i , y r (23)

This loss models the discrete probability distribution of continuous boundary positions, allowing the network to learn the target boundary distribution more precisely, thereby improving the regression accuracy for small targets and boundary-ambiguous targets.

3.5.4 Joint optimization analysis

The SGM and AGD modules proposed in this paper do not introduce separate auxiliary supervision terms but are instead jointly optimized end-to-end under the detection loss constraints mentioned above. Specifically, the SGM enhances the denoising intensity, edge magnitude, and directional priors in the SAR input, improving the structural representation capability of the SAR branch. The AGD, on the other hand, performs adaptive gated fusion of the visible and SAR features at the P3, P4, and P5 levels, allowing the contributions of each modality at different scales and spatial locations to be dynamically adjusted around the detection target. As a result, structural enhancement, hierarchical fusion, and detection prediction are all incorporated into the same optimization framework, and parameter updates are ultimately completed through the joint backpropagation of λ b o x, λ c l s and λ d f l. This loss design not only maintains the training stability of the detection framework but also avoids the complexity introduced by additional auxiliary supervision, allowing the method to improve the performance of visible-SAR cross-modal object detection without significantly increasing the training burden.

In summary, the loss function in this paper adopts a classification-localization-distribution regression joint optimization strategy, jointly constraining category discrimination, bounding box localization, and discrete distribution modeling under a unified detection goal. This design is consistent with the dual-branch structural enhancement and dynamic fusion mechanism proposed in this paper and provides an optimized foundation for achieving precise and stable cross-modal object detection in complex environments.

4 Experiment

4.1 Dataset

To validate the effectiveness of the proposed model under both standard cross-modal detection conditions and complex imaging scenarios, this paper utilizes two publicly available datasets, OGSOD 1.0 and OSPRC, for experimentation (Wang et al., 2023). These two datasets address the two core challenges the method aims to tackle: the former is used to evaluate the performance of the dual-branch intermediate fusion framework under standard visible-SAR detection conditions, while the latter is used to verify the structural compensation ability and dynamic fusion effectiveness of the SGM and AGD modules in complex degraded scenes.

4.1.1 OGSOD 1.0 dataset

OGSOD 1.0 is a publicly available benchmark dataset for visible-SAR multi-modal remote sensing object detection. It includes accurately registered pairs of visible and SAR images, with labeled categories such as bridges, ports, and oil tanks. The dataset has two main characteristics: first, the modality registration quality is high, as shown in Figure 5, where the visible and SAR images are well aligned in the same scene, effectively removing interference caused by significant spatial misalignment. This makes it suitable for evaluating the effectiveness of the proposed dual-branch feature extraction and P3/P4/P5 multi-level intermediate fusion structure. Second, the dataset contains many small-to medium-sized targets, with complex relationships be-tween the target boundaries and the background, making it ideal for evaluating the model’s ability to perform high-precision localization at high IoU thresholds. Based on these characteristics, OGSOD 1.0 is mainly used to validate the proposed method’s detection accuracy, localization quality, and its balance between accuracy and complexity.

4.1.2 OSPRC dataset

The OSPRC dataset also consists of paired visible and SAR images but focuses on verifying robustness under complex observation conditions. Figure 6 shows representative images from the dataset, which differs from OGSOD 1.0 by including disturbances such as resolution changes, polarization variations, and cloud cover. This dataset is particularly relevant for testing the model’s adaptability to modality degradation and scene variations. For the dual-branch fusion framework, this dataset is highly useful as the resolution and polarization variations test the stability of the feature extraction and multiscale intermediate fusion, while the cloud cover scenes directly validate the SAR structural enhancement provided by the SGM module and the dynamic weight allocation ability of the AGD module under conditions of modality reliability changes. Therefore, OSPRC is primarily used to verify the robustness of the proposed method, especially its structural compensation and dynamic fusion effectiveness in complex environments.

4.2 Experimental platform and parameters

To evaluate the effectiveness of DBF-YOLO in visible-SAR cross-modal image detection, comparative experiments, robustness experiments, and ablation experiments were designed based on the OGSOD 1.0 and OSPRC datasets. Tables 1, 2 provide the version and configuration information of the experimental environment, as well as the relevant parameter settings used during training, to ensure the reproducibility and fairness of the experiments. The floating-point operations (FLOPs) of all models were calculated using the THOP library. During the calculation, the input image size was set to 640 × 640 pixels, and the complete inference process was considered, including the backbone network, cross-modal fusion modules, neck, and detection head. For DBF-YOLO, both the visible branch and the SAR branch were included in the FLOPs calculation to ensure that the reported computational complexity reflects the actual deployment cost of the complete dual-branch framework.

TABLE 1. Configuration experimental environment.

TABLE 2. Training parameters setting.

4.3 Evaluation metrics

To quantitatively evaluate the detection performance of different methods, this experiment assesses model performance using evaluation metrics including mean average precision (mAP), the number of model parameters, and floating-point operations (FLOPs).

When the intersection over union (IoU) between a predicted bounding box and the ground-truth bounding box (GT) exceeds a given threshold, the predicted bounding box is regarded as a true positive (TP); otherwise, it is classified as a false positive (FP). Any ground-truth object that is not detected is counted as a false negative (FN). The IoU is defined as follows: The evaluation metrics used in this study are defined in Equations 24–28.where B p and B g represent the predicted bounding box and its corresponding ground-truth bounding box, respectively.

I o U B p , B g = B p ∩ B g B p ∪ B g (24)

The precision and recall at threshold t are calculated as follows:

P r e c i s i o n = T P T P + F P (25)

R e c a l l t = T P T P + F N (26)

For each object class, the average precision A P i is defined as the integral of the precision–recall curve:where P i r represents the precision of class i as a function of recall.

A P i t = ∫ 0 1 P i r d r = ∫ 0 1 P r e c i s i o n R e c a l l d R e c a l l (27)

Then, the mAP is obtained by averaging over all target classes:

m A P t = 1 n ∑ i = 1 n A P i t (28)

This study adopts the following two widely used evaluation metrics:

mAP50: The mAP calculated at an IoU threshold of 0.5, which reflects the detection capability of the model under standard overlap conditions.

mAP50-95: The mAP evaluated over IoU thresholds ranging from 0.5 to 0.95, with a step size of 0.05. This metric provides a comprehensive and effective measure of localization accuracy and model robustness.

4.4 Comparison experiment

To systematically evaluate the detection performance of the proposed model, this paper compares it with representative single-modal detectors and mainstream multi-modal methods on the OGSOD 1.0 dataset. The single-modal comparison methods include the two-stage detector Faster R-CNN (Ren et al., 2017), Transformer-based RT-DETR-L and RT-DETR-ResNet50 (Zhao Y. et al., 2024), and the lightweight one-stage baseline YOLOv10n, which are tested using visible and SAR input modalities, respectively. The multimodal comparison methods include SuperYOLO (Zhang et al., 2023), ICAFusion (Shen et al., 2024), CFT, and YOLO-CMFM (Zhao et al., 2026), which is designed for visible-SAR detection. Among them, Faster R-CNN and the RT-DETR series provide single-modal performance references under traditional detection and Transformer-based detection frameworks, respectively. YOLOv10n is used as the lightweight single-modal baseline for the proposed dual-branch framework. SuperYOLO, ICAFusion, and CFT represent recently published multimodal fusion detection methods, while YOLO-CMFM represents a recent method that models heterogeneous modality differences for visible-SAR detection.

As shown in Table 3, the experimental results demonstrate that DBF-YOLO achieves 94.4% mAP50% and 70.1% mAP50-95 under dual-modal input. Compared with the single-modal YOLOv10n baseline, the model achieves improvements in both mAP50 and mAP50-95, indicating that dual-modal collaboration not only enhances target recognition capability but also improves localization performance under stricter IoU thresholds. Compared with multimodal methods, the proposed model further outperforms SuperYOLO, ICAFusion, CFT, and YOLO-CMFM in terms of mAP50-95, while using only 5.0 M parameters and 20.5 GFLOPs, achieving a better balance between detection accuracy and model complexity.

TABLE 3. Performance comparison between DBF-YOLO and different algorithms.

Bold values indicate the best results.

From the single-modal results shown in Figure 7, YOLOv10n already demonstrates strong baseline performance, achieving a detection accuracy (mAP50) of 92.4% with visible input and 92.0% with SAR input. In comparison, after introducing the du-al-modal collaborative mechanism, DBF-YOLO improves the detection accuracy to 94.4%, while also increasing the comprehensive localization accuracy (mAP50-95) to 70.1%. Specifically, compared with the visible baseline, the two metrics are improved by 2.0 and 9.0 percentage points, respectively; compared with the SAR baseline, they are improved by 2.4 and 7.3 percentage points, respectively. These results indicate that dual-modal collaboration not only improves detection recall capability but also enhances target detection accuracy under stricter IoU thresholds to some extent.

Compared with multimodal methods, although CFT achieves a detection accuracy of 95.7%, which is slightly higher than that of the proposed DBF-YOLO (94.4%), DBF-YOLO shows a certain advantage in comprehensive localization accuracy, which better reflects bounding-box localization quality. Specifically, DBF-YOLO achieves a comprehensive localization accuracy of 70.1%, outperforming SuperYOLO (53.8%), ICAFusion (64.1%), CFT (62.9%), and YOLO-CMFM (66.0%) by 16.3%, 6.0%, 7.2%, and 4.1%, respectively. Considering the strict requirements for localization accuracy in practical detection tasks, this result indicates that the advantage of DBF-YOLO lies in achieving higher-standard localization accuracy rather than merely pursuing detection rates under loose thresholds, thereby verifying the effectiveness of the proposed model in high-quality target localization.

From the perspective of efficiency, DBF-YOLO has only 5.0 M parameters and 20.5 GFLOPs, with an overall complexity significantly lower than ICAFusion with 20.2 M parameters and 29.8 GFLOPs, and CFT with 44.8 M parameters and 35.8 GFLOPs. Specifically, compared with ICAFusion, the proposed method reduces the number of parameters by approximately 75.2% and computational cost by approximately 61.4%, while improving comprehensive localization accuracy by 6.0%. Compared with CFT, the number of parameters is reduced by approximately 88.8%, GFLOPs by approximately 67.9%, and comprehensive localization accuracy is improved by 7.2%. Even compared with the lightweight method SuperYOLO, the proposed method has only 0.2 M more parameters, but reduces GFLOPs substantially from 56.2 to 20.5, while still improving comprehensive localization accuracy by 16.3%. These results demonstrate that DBF-YOLO achieves better detection accuracy while maintaining low model complexity, resulting in a better balance between accuracy and efficiency.

For different target categories, the detection accuracy of the proposed model reaches 96.6% and 99.5% for bridges and ports, respectively, which is comparable to the performance of the compared multimodal methods. For oil tanks, the detection accuracy is 87.3%, slightly lower than the 90.7% achieved by CFT. However, benefiting from its lightweight design with only 5.0 M parameters and 20.5 GFLOPs, DBF-YOLO still shows a clear overall advantage. Therefore, the advantage of DBF-YOLO shown in Table 3 lies in its comprehensive effectiveness. While maintaining a lightweight architecture, the proposed method effectively balances multi-class recognition accuracy and high-threshold localization accuracy in complex scenarios, thereby improving its potential for practical engineering applications.

The above results show that the proposed method has strong overall advantages in improving localization performance under stricter IoU thresholds and controlling model complexity. On the one hand, the SGM module enhances the structural usability of the SAR modality at the input stage, which helps provide more stable geometric priors for subsequent feature extraction. On the other hand, the AGD module adaptively weights dual-modal features at the three intermediate fusion levels of P3, P4, and P5, which helps improve the effectiveness of cross-modal information interaction at different scales. Overall, the experimental results demonstrate that the proposed method effectively reduces computational complexity through input-level structural compensation and multi-scale adaptive fusion design, while maintaining high localization accuracy and feature utilization efficiency. This design not only optimizes the use of computational resources, but also lays a foundation for efficient cross-modal target detection in practical resource-constrained environments such as airborne and spaceborne platforms.

4.5 Robustness experiments under different imaging conditions

To further verify the robustness of the model under different resolutions, polarization modes, and cloud occlusion conditions, this paper conducts tests on six typical imaging scenarios in the OSPRC dataset. The results are shown in Table 4. The experimental results indicate that DBF-YOLO maintains relatively stable performance under different imaging conditions, with particularly significant improvements in cloud occlusion scenarios.

TABLE 4. Robustness validation under different imaging conditions.

Under normal imaging conditions, the dual-modal model consistently achieves positive performance gains compared with the visible single-modal model. Taking the OSPRC-10-VH dataset as an example, DBF-YOLO achieves mAP50 and mAP50-95 values of 77.3% and 54.5%, respectively, under visible-SAR dual-modal input, which are 2.1% and 4.0% higher than those of the visible single-modal input at 75.2% and 50.5%. On the OSPRC-10-VV dataset, these improvements further increase to 4.5% and 5.7%, respectively. This indicates that when polarization changes lead to differences in SAR representation, dual-modal fusion can still provide stable performance gains.

Under lower spatial resolution conditions, the model also shows good stability. In the OSPRC-20-VH and OSPRC-20-VV scenarios, DBF-YOLO improves mAP50-95 by 4.8% and 4.3%, respectively, compared with the visible single-modal model. This indicates that even when texture information is degraded due to reduced resolution, the structural compensation provided by SAR can still improve target localization quality. In particular, the stable improvement in the stricter mAP50-95 metric shows that the model not only improves coarse-grained recognition, but also maintains advantages in bounding-box regression and localization accuracy.

The most significant improvements are observed in cloud occlusion scenarios. Under the OSPRC-Cloud-VH condition, the visible single-modal model achieves only 42.3% mAP50% and 18.4% mAP50-95, whereas DBF-YOLO reaches 76.2% and 47.2%, corresponding to substantial improvements of 33.9% and 28.8%, respectively. Under the OSPRC-Cloud-VV condition, this advantage becomes even more pronounced: compared with visible input, mAP50 increases from 42.2% to 77.7%, and mAP50-95 increases from 18.2% to 48.8%, with improvements of 35.5% and 30.6%, respectively. These results clearly demonstrate that when visible images are severely occluded and texture information is substantially degraded, the all-day and all-weather structural information provided by the SAR modality plays a decisive compensatory role in detection performance.

It is worth noting that, compared with SAR single-modal input, DBF-YOLO also maintains consistent improvements across all scenarios. Taking the two cloud occlusion scenarios as examples, mAP50 and mAP50-95 increase by 5.1% and 0.9%, respectively, under Cloud-VH, and by 7.7% and 2.7%, respectively, under Cloud-VV. This suggests that the proposed method does not simply replace RGB information with SAR information. Instead, it further integrates the still-effective local texture and category details from visible images on the basis of reliable SAR structural information, thereby achieving more robust cross-modal complementarity.

Overall, the results in Table 4 demonstrate two main points. First, DBF-YOLO has strong scene adaptability and can maintain stable performance improvements under changes in resolution, polarization mode, and visible-image degradation. Second, these gains are not incidental, but are consistent with the design logic of the proposed model, in which SGM improves the usability of SAR input and AGD adaptively allocates modality weights at multiple scales. In other words, as scene degradation becomes more severe, the model can automatically increase its reliance on the SAR modality, thereby maintaining stable detection performance.

4.6 Ablation experiment

To verify the independent contributions and synergistic effects of the multi-scale cross-modal fusion structure, the SGM module, and the AGD module, this paper con-ducts ablation experiments on the OGSOD 1.0 dataset using YOLOv10n as the baseline.

To clearly evaluate the contribution of each proposed component, the ablation experiments are conducted using Index A as the baseline model. Starting from this baseline, the subsequent variants are constructed by progressively introducing the MCFP, SGM, and AGD modules. Specifically, Index B evaluates the effect of MCFP, while Index C and Index D investigate the individual contributions of SGM and AGD, respectively. The remaining variants combine different modules to analyze their complementary effects, and Index H represents the complete DBF-YOLO model with all proposed components.

From the overall results in Table 5, the baseline model achieves mAP50 and mAP50-95 values of 93.6% and 69.8%, respectively. After adding the proposed modules, each comparison model outperforms the baseline in at least one metric, indicating that the designed components are not simply stacked together, but can improve detection performance from different perspectives. The final complete model improves the two metrics to 94.4% and 70.1%, corresponding to gains of 0.8% and 0.3% over the baseline, respectively.

TABLE 5. Comparison of ablation experiment results.

Bold values indicate the best results. MCFP, SGM, and AGD, represent the multiscale cross-modal fusion structure, the SGM, module, and the AGD, module, respectively.

From the effects of individual modules shown in Figure 8, some improvements are more inclined to enhance detection performance under a loose threshold, while others are more focused on improving localization quality under a high threshold. Among different ablation combinations, the highest mAP50 reaches 94.4%, indicating that the multi-scale cross-modal fusion structure and gated fusion mechanism are effective in enhancing overall recognition capability. When SGM is introduced, mAP50-95 is further improved to 69.9%, suggesting that structural enhancement at the SAR input stage has a more direct effect on improving bounding-box quality under precise target recognition thresholds. This result is consistent with the functional role of SGM: its main function is not to enhance shallow visible texture responses, but to strengthen the edge and directional priors of the SAR modality, thereby improving the geometric constraints of subsequent fused features.

When the multi-scale cross-modal fusion structure, SGM, and AGD are introduced simultaneously, the model achieves the best overall performance. This indicates a clear synergistic relationship among the three components. The multi-scale cross-modal fusion structure provides a multi-scale interaction space, SGM first improves the input quality of the SAR branch, and AGD then performs adaptive gated allocation at the P3, P4, and P5 fusion nodes. In other words, the performance improvement of the complete model does not come from any single module alone, but from the joint effect of the complete pipeline of input enhancement, hierarchical interaction, and adaptive selection.

Therefore, the improvement of the proposed method is mainly reflected in the cooperation among modules, rather than an extreme gain from a single component. Especially under a strong baseline, the complete model can stably improve both metrics simultaneously, demonstrating the structural rationality of the proposed multi-scale cross-modal fusion framework and dual-module mechanism.

4.7 Visual experiment

To intuitively demonstrate the detection performance of DBF-YOLO in complex scenarios, visualization experiments were conducted on the OGSOD 1.0 and OSPRC datasets. The experiments selected typical scenes containing densely distributed and occluded targets, and a qualitative analysis was performed on the same test sets using the recently published multimodal fusion detection method ICAFusion as the comparison model. Figure 9 presents the visualization results under different scenarios of DBF-YOLO on the OGSOD 1.0 dataset, where green, blue, and red rectangles indicate detected bridges, harbors, and oil tanks, respectively. As observed in the figure, targets of different types in remote sensing images face challenges such as local detail loss due to limited resolution and easy confusion with the background. Nevertheless, the model still demonstrates high detection accuracy.

In densely distributed scenarios, as shown in the red dashed rectangle in Figure 10a, the detection results of ICAFusion shown in Figure 10b are limited due to the relatively low detail resolution of remote sensing images, which makes detection in densely populated target areas less effective. Furthermore, because these targets have little contrast in grayscale or brightness relative to the background, missed detections are likely to occur. In contrast, the detection results of DBF-YOLO shown in Figure 10c demonstrate that, even though the dense arrangement of oil tanks within the red dashed box leads to similar textures and would generally cause confusion between similar targets, DBF-YOLO is still able to accurately detect most small targets. This indicates that, by incorporating the SGM and AGD modules, DBF-YOLO effectively reduces false detections while the bounding boxes remain closely aligned with the ground-truth targets, further demonstrating the improvement in precise target recognition under strict IoU thresholds.

Figure 11 is the comparison of visualization results on the OSPRC dataset under cloudy conditions. When targets are partially occluded by clouds, as shown in Figure 11, the advantages of DBF-YOLO become even more apparent. First, in cases of partial occlusion, DBF-YOLO is able to detect all visible portions of the targets, while in the left image of (b), some parts are missed. Second, due to the correction provided by the SAR modality features, DBF-YOLO can correctly detect bridge targets that are partially obscured by clouds, whereas in the right image of (b), the bridge targets are misidentified as harbor targets due to the cloud occlusion. This demonstrates that, in cloud-covered or lowresolution regions, DBF-YOLO can leverage the structural information from the SAR branch and the adaptive weighted allocation of the AGD module to effectively compensate for the missing visible features, ensuring robust target recognition.

Through visual analysis, DBF-YOLO exhibits more concentrated and continuous activation responses in target regions, while background interference is significantly reduced. The visible branch provides rich texture information, whereas the SAR branch strengthens edge and structural features. The two are dynamically weighted and fused in the AGD module, thereby ensuring stable detection performance in complex back-grounds.

5 Discussion

The experimental results show that, in visible-SAR cross-modal remote sensing target detection, the key to improving model performance does not lie in simply increasing the network scale, but in whether the model can be specifically designed around modality differences, the utilization of structural priors, and fusion efficiency. DBF-YOLO achieves an mAP50-95 of 70.1% on the OGSOD 1.0 dataset, outperforming the multimodal methods compared in this paper. Meanwhile, it uses only 5.0 M parameters and 20.5 GFLOPs, indicating that the proposed method achieves a more reasonable balance between detection accuracy and model complexity. In the cloud occlusion scenarios of the OSPRC dataset, compared with the visible single-modal model, DBF-YOLO improves mAP50-95 by 28.8% and 30.6% on the Cloud-VH and Cloud-VV subsets, respectively, demonstrating good stability and environmental adaptability under complex degraded conditions.

From the perspective of the methodological mechanism, the results of this paper are consistent with the recent trend in multimodal object detection, which has shifted from complex fusion toward effective fusion and task-oriented fusion. Existing studies generally indicate that performance improvement in multimodal detection does not depend on the stacking complexity of fusion structures, but rather on whether modality differences, structural degradation, and detection task requirements can be effectively modeled. In this paper, SGM enhances the structural usability of the SAR modality through local statistical filtering, Gaussian smoothing, and gradient-assisted encoding. By jointly providing denoised intensity, gradient magnitude, and gradient direction information, SGM reduces the influence of speckle noise while preserving structural priors that are beneficial for subsequent cross-modal fusion. MCFP establishes progressive cross-modal interaction paths at the P3, P4, and P5 levels, enabling detail supplementation, semantic transition, and high-level discriminative enhancement to be collaboratively achieved at different scales. AGD realizes dynamic modality allocation across channels and spatial locations through lightweight projection, gating weight generation, and residual refinement. Together, these designs form a technical pipeline of input enhancement, hierarchical interaction, and dynamic selection, enabling the model to achieve favorable detection performance while maintaining a lightweight structure.

Furthermore, the experimental results under resolution variation, polarization variation, and cloud occlusion conditions in the OSPRC dataset demonstrate that the proposed method can dynamically utilize the stable structural information of SAR according to the degree of scene degradation, thereby compensating for the representational deficiencies of the visible modality in complex environments. This indicates that, in visible-SAR target detection tasks, robustness improvement depends more on the effective utilization of the reliability of different modalities than on simply replacing one modality with another. Especially when visible information is significantly degraded, the geometric contours and scattering structural information provided by the SAR modality can offer effective compensation, supporting the model in maintaining favorable detection performance.

6 Conclusion

This paper proposes DBF-YOLO, a lightweight dual-branch fusion model for visible-SAR cross-modal remote sensing target detection. Based on the YOLOv10 framework, the model adopts a dual-branch backbone network to separately extract visible and SAR features, and establishes intermediate fusion paths at the P3, P4, and P5 levels. Meanwhile, the SGM module is used to enhance the structural usability of SAR input, and the AGD module is introduced to achieve adaptive weighted fusion of dual-modal features, thereby improving cross-modal collaborative detection capability at different scales. Experimental results show that the proposed method on the OGSOD 1.0 dataset achieves a favorable balance between detection accuracy and deployment complexity. Under complex imaging conditions in the OSPRC dataset, including different resolutions, polarization modes, and cloud occlusion, the model also maintains stable performance gains, especially showing a more significant structural compensation advantage in scenarios where visible information is severely degraded.

Overall, this study demonstrates that lightweight design centered on SAR structural prior enhancement, multiscale intermediate interaction, and dynamic modality weighting is an effective approach for improving detection performance, localization quality, and complex-scene adaptability in visible-SAR target detection. The experiments conducted in this study are based on accurately registered visible-SAR image pairs and mainly validate the effectiveness of DBF-YOLO under reliable geometric correspondence. Future work will further explore the robustness of the proposed framework under more challenging conditions, including geometric inconsistencies, modality degradation, and resource-constrained deployment scenarios.

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.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frsen.2026.1919576/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Yaoming Cai, Zhongnan University of Economics and Law, China

Reviewed by: Hongbo Xiang, China University of Geosciences Wuhan, China

Guochao Wang, National University of Defense Technology, China