SG-YOLO: A Multispectral Small-Object Detector for UAV Imagery Based on YOLO

Highlights

What are the main findings?

• SG-YOLO, a lightweight multispectral detector for UAVs, is proposed, featuring novel modules (SGDS, SSIAF, SCSF) for efficiently extracting and fusing cross-band features while suppressing redundant information.
• Tests on the MODA dataset show SG-YOLO achieves 72.4% mAP₅₀, outperforming both RGB baselines and other methods, balancing accuracy and efficiency.

What are the implications of the main findings?

• Multispectral data (8 bands) is validated to enhance UAV detection robustness, overcoming limitations of RGB images.
• The framework provides a practical, low-complexity solution for real-time multispectral perception on resource-constrained UAV platforms.

Abstract

Object detection in unmanned aerial vehicle (UAV) imagery remains a crucial yet challenging task due to complex backgrounds, large scale variations, and the prevalence of small objects. Visible-spectrum images lack robustness under all-weather and all-illumination conditions; by contrast, multispectral sensing provides complementary cues (e.g., thermal signatures) that improve detection robustness. However, existing multispectral solutions often incur high computational costs and are therefore difficult to deploy on resource-constrained UAV platforms. To address these issues, SG-YOLO is proposed, a lightweight and efficient multispectral object detection framework that aims to balance accuracy and efficiency. First, a Spectral Gated Downsampling Stem (SGDS) is designed, in which grouped convolutions and a gating mechanism are employed at the early stage of the network to extract band-specific features, thereby maximizing spectral complementarity while minimizing redundancy. Second, a Spectral–Spatial Iterative Attention Fusion (SSIAF) module is introduced, in which spectral-wise (channel) attention and spatial-wise attention are iteratively coupled and cascaded in a multi-scale manner to jointly model cross-band dependencies and spatial saliency, thereby aggregating high-level semantic information while suppressing redundant spectral responses. Finally, a Spatial–Channel Synergistic Fusion (SCSF) module is designed to enhance multi-scale and cross-channel feature integration in the neck. Experiments on the MODA dataset show that SG-YOLOs achieves 72.4% mAP₅₀, outperforming the baseline by 3.2%. Moreover, compared with a range of mainstream one-stage detectors and multispectral detection methods, SG-YOLO delivers the best overall performance, providing an effective solution for UAV object detection while maintaining a favorable trade-off between model size and detection accuracy.

1. Introduction

In recent years, benefiting from the high maneuverability of aerial platforms [1] and advances in sensor miniaturization [2], unmanned aerial vehicles (UAVs) equipped with vision-based perception systems have become indispensable in emergency response and disaster relief [3,4], precision agriculture [5], and intelligent security surveillance [6]. Compared with satellite remote sensing and conventional ground-based observation, low-altitude remote sensing via UAVs offers unique advantages in terms of timeliness and spatial resolution. However, UAV-based visual detection still faces substantial challenges. Targets of interest are often extremely small, densely distributed, and easily confused with complex backgrounds [7]. Moreover, dynamic changes in aerial viewpoints cause the same object to exhibit pronounced variations in scale and appearance across scenes, placing stringent demands on the feature representation capability of detection algorithms [8].

Although deep-learning-based methods [9,10,11] have achieved remarkable progress in UAV remote-sensing detection, most studies are still limited to the visible (RGB) modality. Under low-visibility conditions such as low illumination, haze, or rainfall, the imaging quality of visible-spectrum data can degrade sharply due to insufficient ambient light and atmospheric scattering, leading to a substantial drop in detection performance [12,13,14]. In contrast, multispectral imaging can simultaneously capture the spectral reflectance characteristics of targets in the visible and selected bands such as near-infrared. By exploiting non-visible bands, it improves robustness against environmental interference [15,16,17] and provides discriminative information beyond human perception, often referred to as “spectral fingerprints” [18]. Similar multi-domain joint characterization strategies have also proven highly effective for robust target detection in broader remote sensing applications [19,20,21]. As shown in Figure 1, when pedestrians are difficult to distinguish from the background in RGB images, multispectral imagery can still reveal clearer target–background separability in the spectral domain.

For a long time, progress in multispectral object detection has been constrained by the lack of high-quality datasets. Most existing UAV benchmarks contain only two modalities, typically visible and thermal infrared (RGB-T) data [22,23], which has led many methods to focus primarily on bimodal fusion. To alleviate this bottleneck, Han et al. released the MODA dataset in 2025 [16], which, to the best of our knowledge, is the first large-scale UAV multispectral object detection benchmark. It provides a per-frame resolution of 1200 × 900 pixels and covers eight spectral bands ranging from 395 nm to 950 nm. While MODA enables multi-band feature learning, it also introduces new challenges for algorithm design.

With the substantial increase in input bands—from bimodal setups to eight-band inputs—existing fusion strategies face severe challenges. Deep-learning-based multispectral detectors are typically categorized into early-, mid-, and late-fusion schemes [24,25,26]. Although mid- and late-fusion approaches can leverage independent branches to extract modality-specific features, their parameter count grows roughly linearly (or multiplicatively depending on design) with the number of bands, making them difficult to meet real-time requirements on UAV edge devices [27,28]. For full-band inputs, the abundant spectral information can improve recognition robustness, but adjacent bands often exhibit high spectral correlation [29]. Without effective in-network feature selection or attentive mechanisms, treating all bands equally with vanilla convolution results in large computational redundancy and may allow useful signals to be overwhelmed by noise [30]. Therefore, achieving efficient de-redundancy and feature refinement within the network—while preserving the full-band information—is key to designing lightweight, real-time multispectral detectors for UAV platforms.

Accordingly, effectively exploiting rich multispectral information for robust detection, while suppressing inter-band redundancy and reducing model complexity, has become a pressing problem. To this end, a lightweight, real-time multispectral object detector is developed that fully utilizes spatial and spectral information for UAV applications. Our main contributions are as follows:

(1)

We propose a novel downsampling Stem to mitigate the early-stage over-compression of spectral information into a single embedding. The Stem integrates grouped downsampling and a spectral-gating mechanism to selectively screen and reweight multispectral cues within the first two layers, achieving a favorable balance between efficiency and de-redundancy.

(2)

We introduce an SSIAF module to address redundant fusion and insufficient feature representation in multispectral integration. By incorporating the iAFF attention mechanism, SSIAF adaptively enhances informative features and suppresses redundant responses in both channel and spatial dimensions.

(3)

To improve multi-scale detection in UAV scenarios, we design an SCSF module in the neck. SCSF combines multi-scale convolutions with lightweight ECA (Efficient Channel Attention) to guide cross-scale feature aggregation, thereby improving robustness to scale variation and small-object detection.

2. Related Work

2.1. Deep Learning-Based Object Detection

The core challenge of object detection lies in designing efficient and discriminative feature representation mechanisms. In recent years, the field has largely shifted to end-to-end representation learning based on deep neural networks [31]. From an architectural perspective, modern detectors can be broadly grouped into two paradigms: convolutional neural networks (CNNs) and Transformers.

Within the CNN family, detectors are commonly categorized into two-stage and one-stage pipelines. Two-stage detectors, represented by the R-CNN line, typically achieve strong accuracy but incur substantial computational and memory costs, which hinders real-time deployment on UAV edge platforms [32,33,34]. In contrast, one-stage detectors such as YOLO [35] and SSD [36] perform direct regression for classification and localization, offering higher throughput and better suitability for real-time UAV applications.

With the -development of Transformers, DETR-style detectors employ global self-attention to build an end-to-end framework that eliminates non-maximum suppression (NMS) [37]. Hierarchical designs such as Swin Transformer further enhance representation capability in complex scenes [38]. However, compared with CNNs, Transformers lack inherent inductive biases (e.g., locality and translation equivariance), which often leads to optimization difficulties and unstable generalization in remote-sensing scenarios with limited training data or distribution shifts [39].

In UAV remote sensing, characterized by dense object distribution, severe background clutter, and a predominance of small targets, researchers have proposed task-oriented improvements to generic detection frameworks. The drive for efficiency has spurred the development of fundamental lightweight convolutional operators, most notably the depthwise separable convolution introduced in MobileNet [40,41]. For efficiency, lightweight operators (e.g., GhostNet and GSConv) and model compression techniques have been explored to reduce parameters while retaining sensitivity to small targets [8,42]. Regarding feature interaction, recent progress emphasizes long-range dependency modeling (e.g., Mamba-based designs) and global feature fusion to mitigate feature degradation in cluttered backgrounds and improve localization accuracy [43]. In addition, enhanced multi-scale feature pyramids (FPN) and improved bounding-box regression losses have improved robustness to dramatic scale variations [44].

Despite these advances in deep learning-based object detection, most UAV remote-sensing detectors are still developed primarily for visible-light imagery and thus inherit its sensitivity to illumination variations and adverse weather (e.g., fog and rain). Moreover, relying on a single visible modality is subject to inherent physical limitations in extremely complex environments, which makes stable and reliable detection difficult. In contrast, multispectral detection integrates complementary information across spectral bands and leverages object “spectral fingerprints,” offering a promising route to more robust detection under challenging conditions.

2.2. Multispectral/Multi-Modal Fusion for Object Detection

Multispectral fusion of visible and thermal imagery (RGB–T) has demonstrated clear advantages for object detection in challenging environments. Current research primarily focuses on extracting and integrating complementary cues across modalities. Based on where fusion occurs, methods are commonly categorized as early-, mid-, or late-fusion [24]. Mid-fusion approaches typically employ dual-stream backbones to model modality-specific deep features. For instance, ICAFusion [45] proposes a dual-cross-attention transformer to model global feature interactions and capture cross-modal complementarities, while MMI-Det [46] designs a fusion-focusing module to extract discriminative cues from visible and thermal streams and emphasize object regions. These methods achieve state-of-the-art performance on RGB–T benchmarks such as M3FD [47] and Drone-Vehicle [23].

With the advancement of aerial remote sensing, higher spectral-resolution multispectral and hyperspectral sensors (e.g., 8+ bands) are increasingly used in precision agriculture and disaster monitoring. The MODA dataset [16], released by BIT in October 2025, is the first large-scale UAV dataset captured with an 8-band multispectral camera. Extending multi-branch fusion paradigms to such high-band inputs leads to linear or super-linear increases in parameters and computation, severely limiting real-time deployment on UAV edge platforms. To balance accuracy and efficiency, some studies adopt early-fusion strategies [27,48,49], stacking bands along the input channel dimension. Although early fusion maintains a lightweight architecture, standard convolutional operators (e.g., conventional detector Stems) apply uniform weighting and spatial aggregation across all bands, failing to account for heterogeneous physical characteristics of different spectral bands and thus risking that salient spectral fingerprints are masked by background noise or redundancy [17].

In summary, existing multispectral detectors still struggle to reconcile band scalability with computational cost and lack specialized perceptual operators for high-spectral-resolution features. To address these gaps, we propose a lightweight detection framework that (1) reconstructs the Stem with grouped spectral-aware processing and gating to perform differentiated feature extraction at the input stage and (2) integrates an iterative attention feature fusion (iAFF) module to deeply merge multispectral cues and suppress redundancy at minimal computational overhead.

3. Methodology

3.1. Overview of Our Method

The overall architecture of the proposed SG-YOLO is shown in Figure 2. SG-YOLO is a lightweight, high-efficiency detector specifically designed for multispectral small-object detection in UAV aerial imagery.

Input is an N × H × W multispectral image (where N denotes the number of spectral bands). The input first passes through our proposed SGDS (Spectral Gated Downsampling Stem). SGDS decouples different bands during the first two downsampling stages and allocates adaptive weights to each band via a gating mechanism, thereby enhancing responses of informative bands and suppressing redundant noise to provide a more discriminative low-level representation for the backbone.

In the backbone, we employ the SSIAF Block (Spectral–Spatial Iterative Attentional Fusion Block) to strengthen feature extraction. SSIAF adopts a cascaded attention scheme that alternately (iteratively) models channel-wise and spatial-wise interactions, enabling deep fusion of spectral and spatial information and mitigating redundancy accumulation when multispectral features are stacked in deep layers.

The neck handles cross-layer interaction and fusion of multi-scale features. To improve multi-scale fusion quality and channel-wise semantic interaction, we introduce the SCSF Block (Spatial–Channel Synergy Fusion Block) in the neck: it captures contextual information at different receptive fields using multi-scale convolutional kernels and dynamically reinforces salient channels via channel attention, thereby enhancing representation for targets at various scales, especially small and medium objects.

Finally, the multi-scale features produced by the neck are fed into three detection heads for classification and bounding-box regression, yielding robust detection across scales. The overall design emphasizes effective utilization and fusion of multispectral information while remaining parameter- and computation-efficient to facilitate deployment on resource-constrained UAV platforms.

3.2. Spectral Gated Downsampling Stem

Multispectral imagery extends the spectral dimension by incorporating bands beyond visible light (e.g., near-infrared and short-wave infrared). These additional bands provide critical complementary cues under complex backgrounds, low-contrast conditions, and dim illumination, thereby significantly enhancing a detector’s discriminative capability. Conventional RGB-based object detectors are designed to process ${\mathrm{I}}_{\mathrm{R}\mathrm{G}\mathrm{B}}\in {ℝ}^{\mathrm{H}\times \mathrm{W}\times 3}$, whereas multispectral inputs are represented as ${\mathrm{I}}_{\mathrm{M}\mathrm{S}\mathrm{I}}\in {ℝ}^{\mathrm{H}\times \mathrm{W}\times \mathrm{N}}$. The simplest adaptation is to change the input channels of the first convolution to support N-channel inputs. However, this design forces spectral features to be directly compressed into a shared mapping through a single convolution layer, leading to overly early fusion and potential dilution of band-specific information.

Based on this observation, we follow the principle of “late fusion, spectral decoupling, and adaptive recalibration” in our network design: we maintain band-wise independent spatial feature streams in early layers to preserve spectral fingerprints and then use a lightweight gating mechanism to emphasize informative spectral signals while suppressing redundant noise. This produces cleaner and more discriminative low-level representations for subsequent deep spectral–spatial fusion.

As illustrated in Figure 3, SGDS consists of two band-grouped downsampling convolutions and a lightweight channel-gating unit. Given a multispectral input ${\mathrm{I}}_{\mathrm{M}\mathrm{S}\mathrm{I}}$ with N spectral bands, the module first performs two grouped convolution downsampling operations. Unlike standard convolutions that jointly process all input channels, we set the number of groups to N to ensure that each band remains independent during spatial feature extraction.

In the first downsampling stage, the number of channels is expanded from N to an intermediate channel dimension C_mid (subject to C_mid % N = 0), while the spatial resolution is reduced by half. The computation can be expressed as:

$${\mathrm{F}}_1=\delta (\mathrm{B}(\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{3\times 3,\mathrm{s}=2,\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}\mathrm{s}=\mathrm{N}}({\mathrm{I}}_{\mathrm{M}\mathrm{S}\mathrm{I}})))\in {\mathrm{R}}^{{\mathrm{C}}_{\mathrm{m}\mathrm{i}\mathrm{d}}\times {\displaystyle \frac{\mathrm{H}}{2}}\times {\displaystyle \frac{\mathrm{W}}{2}}}$$

(1)

where $\delta (\cdot )$ denotes the SiLU activation function, and $\mathrm{B}(\cdot )$ denotes batch normalization (BatchNorm2d). Subsequently, the second downsampling stage further increases the number of feature channels to the final output dimension C_out (with the constraint C_out % N = 0), while the spatial resolution is halved again:

$${\mathrm{F}}_2=\delta (\mathrm{B}(\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{3\times 3,\mathrm{s}=2,\mathrm{g}\mathrm{r}\mathrm{o}\mathrm{u}\mathrm{p}\mathrm{s}=\mathrm{N}}({\mathrm{X}}^{(1)})))\in {\mathrm{R}}^{{\mathrm{C}}_{\mathrm{o}\mathrm{u}\mathrm{t}}\times {\displaystyle \frac{\mathrm{H}}{4}}\times {\displaystyle \frac{\mathrm{W}}{4}}}$$

(2)

Through these two downsampling operations, each spectral band is independently mapped to C_out/N feature channels. This design effectively avoids forced mixing of spectral characteristics at the network entrance, thus preserving valuable band-discriminative information for subsequent processing.

After the grouped downsampling, we introduce a lightweight spectral gating unit to adaptively recalibrate the importance of each channel. The unit follows a “squeeze-and-excitation” mechanism: it first applies global average pooling to squeeze the feature map ${\mathrm{F}}_2$ along the spatial dimensions, producing a channel descriptor:

$$\mathrm{z}=\mathrm{G}\mathrm{A}\mathrm{P}({\mathrm{F}}_2)$$

(3)

Then, a simple two-layer bottleneck fully connected module (implemented as $1\times 1$ convolutions in code) is used to capture nonlinear inter-channel dependencies and generate per-channel weights:

$$\mathrm{w}=\sigma ({\mathrm{W}}_2\cdot \delta ({\mathrm{W}}_1\cdot \mathrm{z}))$$

(4)

where ${\mathrm{W}}_1$ and ${\mathrm{W}}_2$ are the weights of the convolution layers, $\delta (\cdot )$ denotes the SiLU activation function, and $\sigma (\cdot )$ denotes the Sigmoid function. The resulting gating weights $\mathrm{w}$ lie in the range [0, 1]. The final output of the module is the feature map modulated by these gating weights:

$${{\mathrm{F}}_2}^{\prime }=\mathrm{w}\otimes {\mathrm{F}}_2$$

(5)

where $\otimes$ denotes element-wise multiplication. This gating mechanism can dynamically enhance informative band features and suppress redundant or noisy band responses according to the input content, thereby providing the backbone network with a more discriminative low-level spectral–spatial representation.

In summary, the SGDS module, through the coordinated design of grouped convolution and the gating mechanism, achieves efficient downsampling while markedly improving the model’s selective perception of multispectral information, laying a solid foundation for subsequent deep feature extraction.

3.3. Spectral–Spatial Iterative Attentional Fusion

In UAV-acquired multispectral imagery, the feature representation of small targets is extremely weak. Traditional feature extraction modules (e.g., the standard C3K2 module in YOLO11) mainly rely on simple linear stacking and residual connections. When dealing with multispectral data that exhibit high spectral correlation and complex spatial structures, this design can easily introduce feature redundancy and overwhelm critical small-target cues. To enable more refined feature selection and enhancement, we propose a Spectral–Spatial Iterative Attention Fusion (SSIAF) module.

SSIAF is a substantial improvement over the C3K2 structure in YOLO11, and its overall architecture is shown in Figure 4. The module consists of two residual units. In each unit, CBS and depthwise separable convolution (DWConv) are used for efficient feature extraction, and the residual connection is then implemented via the iAFF module [50].

The core of SSIAF lies in the iterative Attentional Feature Fusion (iAFF) unit. Through a dual-branch attention mechanism, iAFF captures both local spatial context and global spectral correlation. Different from conventional one-shot attention, iAFF adopts an iterative strategy: the first fusion produces an initial fused feature, which then guides a second-stage re-allocation of attention weights, enabling progressive feature refinement.

The specific computation is illustrated in Figure 5. Let the input feature map be $\mathrm{F}$. In the SSIAF branch, the feature is processed by convolution to obtain an intermediate feature ${\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}$. $\mathrm{F}$ and ${\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}$ are first added element-wise and then fused via Multi-Scale Channel Attention (MS-CAM). Next, the MS-CAM module is applied iteratively to perform a second attention-based fusion on the already fused features, ultimately producing a high-quality fused representation.

The MS-CAM module used in this work is shown in Figure 6. To encourage the network to attend to both global information and local details, MS-CAM contains two branches: a global channel-attention branch and a local-attention branch.

The detailed computation of iAFF is as follows:

$$\begin{array}{ll}\mathrm{Y}& =\mathrm{i}\mathrm{A}\mathrm{F}\mathrm{F}(\mathrm{F},{\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}})\\ & =(\mathrm{F}\otimes \phi ({\mathrm{F}}^{″}))\oplus ({\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}\otimes (1-\phi ({\mathrm{F}}^{″})))\end{array}$$

(6)

Among them, $\mathrm{F}$ and ${\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}$ denote the original feature map and the intermediate feature obtained after convolution, respectively. The symbol $\phi (\cdot )$ indicates that the feature map is processed by the MS-CAM module. The input feature after the first integration is denoted as ${\mathrm{F}}^{″}$, which can be expressed as:

$${\mathrm{F}}^{″}=(\mathrm{F}\otimes \phi ({\mathrm{F}}^{\prime }))\oplus ({\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}\otimes (1-\phi ({\mathrm{F}}^{\prime })))$$

(7)

Here, ${\mathrm{F}}^{\prime }=\mathrm{F}\oplus {\mathrm{F}}_{\mathrm{c}\mathrm{o}\mathrm{n}\mathrm{v}}$. Assuming the feature map fed into MS-CAM is ${\mathrm{F}}_{\mathrm{z}}$, the output feature map ${{\mathrm{F}}_{\mathrm{z}}}^{\prime }$ can be expressed as:

$${{\mathrm{F}}_{\mathrm{z}}}^{\prime }=\sigma (\mathrm{g}({\mathrm{F}}_{\mathrm{z}})\oplus \mathrm{l}({\mathrm{F}}_{\mathrm{z}}))$$

(8)

Here, $\mathrm{B}(\cdot )$ denotes batch normalization, and $\mathrm{G}\mathrm{A}\mathrm{P}(\cdot )$ denotes global average pooling.

3.4. Spatial-Channel Synergy Fusion

In the Neck part of an object detection network, the core task is to efficiently fuse feature maps from different levels of the Backbone so as to aggregate low-level spatial details and high-level semantic information. This is crucial for UAV imagery scenarios where scale variation is severe and small objects are abundant. Traditional feature pyramid networks typically perform multi-scale feature fusion using simple upsampling followed by addition or concatenation, without fully considering the differences among features at different scales in terms of spatial detail and channel semantics. As a result, the fused features often lack sufficient discriminative power. To address this issue, we design a Spatial-Channel Synergistic Fusion module. The key innovation of this module lies in constructing a parallel multi-branch structure, which enables adaptive and fine-grained fusion of Neck features through the synergy between multi-scale spatial context capture and dynamic channel-attention re-calibration, thereby significantly improving the model’s perception of small and medium-sized objects.

As shown in Figure 7, the overall architecture of the SCSF module adopts a C3-like CSP (Cross Stage Partial) design. While preserving lightweight characteristics, it achieves joint optimization of multispectral features in the spatial and channel dimensions through the synergy between a multi-scale spatial branch and an enhanced channel-attention branch. The SCSF module follows the CSP philosophy to balance feature reuse and enriched gradient flow. The input feature is processed by two 1 × 1 convolutions, Conv₁ and Conv₂, which compress the channel dimension from $\mathrm{C}$ to ${\mathrm{C}}_{\mathrm{h}}=\mathrm{C}\times \mathrm{e}$ (where $\mathrm{e}$ is the channel scaling factor, with a default value of 0.5), thereby reducing feature dimensionality and computational cost. The output of Conv₁ is fed into the SCAB module for deeper feature enhancement, while the output of Conv₂ is retained as a shortcut branch to preserve the original feature information. Finally, the features from the two branches are concatenated along the channel dimension to recover the original number of channels and then fused by Conv₃ to produce the output feature $\mathrm{Y}$.

SCAB is the core of the SCSF module. It is designed to accommodate the spatial-context and channel-semantic characteristics of multispectral data, and its structure is illustrated in Figure 8. It mainly consists of two parallel branches: an MS branch (multi-scale spatial feature extraction) and an ECA branch (enhanced channel attention).

3.4.1. Multi-Scale Spatial Feature Extraction Branch

To accommodate the wide range of target scales in UAV imagery, this branch employs parallel large-kernel depthwise separable convolutions to capture spatial context with different receptive fields. The detailed computation process is shown in Figure 8. The input feature is first split into three parallel branches (with kernel sizes K= {3, 5, 7}). In each branch, the corresponding scale features are extracted using a “depthwise separable convolution + pointwise convolution” combination:

$${\mathrm{F}}_{\mathrm{k}}=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}(\mathrm{D}\mathrm{W}\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{k}}(\mathrm{F})),\mathrm{k}\in \mathrm{K}$$

(9)

where $\mathrm{F}$ denotes the input feature map, $\mathrm{D}\mathrm{W}\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{\mathrm{k}}(\mathrm{F})$ represents a depthwise convolution with kernel size $\mathrm{k}$, which expands the receptive field while keeping channels independent; and $\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}^{\mathrm{k}}(\cdot )$ denotes a 1 × 1 pointwise convolution that compresses the output channels of each branch to ${\mathrm{C}}_{\mathrm{h}}/3$.

After concatenating the channel outputs of all branches ${\left\{{\mathrm{F}}_{\mathrm{k}}\right\}}_{\mathrm{k}\in \mathrm{K}}$ convolution is applied to fuse them into a unified spatially enhanced feature:

$${\mathrm{F}}_{\mathrm{s}}=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}^{\mathrm{f}\mathrm{u}\mathrm{s}\mathrm{e}}(\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{c}\mathrm{a}\mathrm{t}({\mathrm{F}}_3,{\mathrm{F}}_5,{\mathrm{F}}_7))$$

(10)

After fusion, the spatial contextual information of targets at different scales is retained.

3.4.2. Enhanced Channel-Attention Branch

To address the issue that different spectral bands in multispectral data contribute unequally to semantic representation, this branch is optimized based on the Efficient Channel Attention (ECA) mechanism. Specifically, a 1D convolution with a dynamically determined kernel size is introduced to capture inter-channel dependencies. The detailed pipeline is illustrated in Figure 8.

The input feature is compressed into a channel-wise statistical vector via Global Average Pooling:

$$\mathrm{Z}=\mathrm{G}\mathrm{A}\mathrm{P}(\mathrm{X})$$

(11)

Here, $\mathrm{Z}\in {ℝ}^{{\mathrm{C}}_{\mathrm{h}}\times 1\times 1}$ reflects the global semantic information of each channel. The compressed vector $\mathrm{Z}$ is then fed into a 1D convolution to capture inter-channel correlations, and a Sigmoid activation is applied to generate the channel-attention weights:

$$\alpha =\sigma (\mathrm{C}\mathrm{o}\mathrm{n}\mathrm{v}1\mathrm{d}(\mathrm{Z}))$$

(12)

where $\sigma (\cdot )$ denotes the Sigmoid function and $\alpha \in {\left[0,1\right]}^{{\mathrm{C}}_{\mathrm{h}}\times 1\times 1}$, enabling the enhancement of informative channels and suppression of redundant ones.

Finally, the channel-attention weights are multiplied element-wise with the original feature, followed by a 1 × 1 convolution to adjust the channel number and obtain the channel-enhanced feature:

$${\mathrm{F}}_{\mathrm{c}}=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}(\mathrm{F}\otimes \alpha )$$

(13)

where $\otimes$ denotes element-wise multiplication, and ${\mathrm{F}}_{\mathrm{c}}\in {ℝ}^{{\mathrm{C}}_{\mathrm{h}}\times \mathrm{H}\times \mathrm{W}}$.

3.4.3. Spatial–Channel Feature Fusion and Modulation

The spatial-enhanced feature ${\mathrm{F}}_{\mathrm{s}}$ and the channel-enhanced feature ${\mathrm{F}}_{\mathrm{c}}$ are concatenated along the channel dimension to obtain a fused feature. The fused feature is then passed through a 1 × 1 convolution to restore the number of channels to ${\mathrm{C}}_{\mathrm{h}}$, and a Sigmoid activation is applied to generate the spatial–channel collaborative attention map:

$$\mathrm{M}=\sigma (\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}({\mathrm{F}}^{\prime }))$$

(14)

where $\mathrm{M}\in {\left[0,1\right]}^{{\mathrm{C}}_{\mathrm{h}}\times \mathrm{H}\times \mathrm{W}}$, enabling dual-dimensional modulation of the features in both spatial and channel domains. Finally, the attention map and the original input feature are modulated via element-wise multiplication, followed by a 1 × 1 convolution to match the channel dimension, and a residual connection is introduced to alleviate gradient vanishing:

$$\mathrm{Y}=\mathrm{C}\mathrm{o}\mathrm{n}{\mathrm{v}}_{1\times 1}(\mathrm{F}\otimes \mathrm{M})+\mathrm{F}$$

(15)

4. Experimental Validation and Analysis

In this section, we conduct experiments to verify the effectiveness of the proposed method and perform a comprehensive evaluation and comparison with state-of-the-art approaches. We adopt standard evaluation metrics for object detection, including Precision (P), Recall (R), F1-score, mean Average Precision (mAP), as well as efficiency metrics GFLOPs and Parameters. The definitions and calculations of these metrics follow the common protocols in the field. The formulas are as follows:

$$\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{P}}}$$

(16)

$$\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}={\displaystyle \frac{\mathrm{T}\mathrm{P}}{\mathrm{T}\mathrm{P}+\mathrm{F}\mathrm{N}}}$$

(17)

$$\mathrm{F}1\text{-}\mathrm{s}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2\times {\displaystyle \frac{\mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}}$$

(18)

$$\mathrm{AP}={\displaystyle {\sum }_{\mathrm{n}=1}^{\mathrm{N}}({\mathrm{R}}_{\mathrm{n}}-{\mathrm{R}}_{\mathrm{n}-1}){\mathrm{P}}_{\mathrm{n}}}$$

(19)

$$\mathrm{mAP}={\displaystyle \frac{1}{\mathrm{N}}}{\displaystyle {\sum }_{\mathrm{i}=1}^{\mathrm{N}}{\mathrm{AP}}_{\mathrm{i}}}$$

(20)

$$\mathrm{G}\mathrm{F}\mathrm{L}\mathrm{O}\mathrm{P}\mathrm{s}={\displaystyle \frac{{\displaystyle {\sum }_{\mathrm{l}=1}^{\mathrm{L}}\mathrm{M}\mathrm{A}\mathrm{C}{\mathrm{s}}_{\mathrm{l}}}}{{10}^9}}$$

(21)

Here, $\mathrm{T}\mathrm{P}$, $\mathrm{F}\mathrm{P}$, and $\mathrm{F}\mathrm{N}$ denote the number of true positive, false positive, and false negative detections, respectively. ${\mathrm{P}}_{\mathrm{n}}$ and ${\mathrm{R}}_{\mathrm{n}}$ denote the Precision and Recall at the sampling point $\mathrm{n}$, respectively, and $\mathrm{N}$ is the total number of points at which the Recall value changes. ${\mathrm{AP}}_{\mathrm{i}}$ represents the Average Precision for the class of $\mathrm{i}$. $\mathrm{M}\mathrm{A}\mathrm{C}{\mathrm{s}}_{\mathrm{l}}$ indicates the number of Multiply–Accumulate Operations in the layer $\mathrm{l}$.

4.1. Data Set Used for Experiments

To evaluate the effectiveness of SG-YOLO, extensive experiments were conducted on the MODA dataset [16]. Proposed by Beijing Institute of Technology, MODA is the first large-scale multispectral UAV object detection dataset. It contains 14,041 multispectral images, covering 8 categories with a total of 330,191 oriented annotations. Each MSI image has a resolution of 1200 × 900 pixels and includes 8 spectral bands spanning 395–950 nm. The dataset was collected across diverse urban scenes from different periods.

As shown in Figure 9a, MODA covers 8 typical ground-object categories with more than 300,000 annotated instances in total. Among them, small vehicles (car) and pedestrians (pedestrian) account for a large proportion, which is consistent with the natural target distribution in real remote-sensing observation scenarios. Notably, the class distribution ratios in the training and test sets are highly consistent (approximately 7:3). This rigorous split ensures the fairness and reliability of subsequent experimental evaluations. Although the dataset exhibits a certain long-tail distribution, even the categories with fewer samples (e.g., bus and tricycle) still contain more than 1000 instances each, which is sufficient to support deep learning models in learning and generalizing features of rare classes.

Figure 9b presents a scatter plot of object width and height distributions, showing that a large number of targets are concentrated in the range below 200 × 200 pixels. Such small objects are particularly challenging for feature extraction. Meanwhile, target sizes span from only a few pixels to several hundred pixels, forming a rich feature space and placing high demands on the model’s multi-scale feature representation capability. In addition, the spatial distribution heatmap in Figure 10 confirms that the dataset does not suffer from center bias: objects are randomly distributed across images, ensuring realistic evaluation and improving the assessment of model robustness in complex scenarios.

4.2. Experimental Settings

Computer configurations and environment: All experiments in this study were performed on a computing workstation running the Ubuntu 20.04 LTS operating system. The hardware configuration comprised an Intel Core i9-10850K CPU (Intel Corporation, Santa Clara, CA, USA) @3.60 GHz 32 GB of RAM, and an NVIDIA GeForce RTX 3090 GPU (NVIDIA Corporation, Santa Clara, CA, USA). The software stack was built on Python 3.10.16, with deep learning frameworks supported by PyTorch 2.5.1 and CUDA 12.4.

Implementation Details: We employed the Stochastic Gradient Descent (SGD) optimizer to train the network, with the hyperparameters set as follows: an initial learning rate of 0.01, a weight decay of 5 × 10⁻⁴, and a momentum of 0.937. The model was trained for 300 epochs with a batch size of 32, and the input image size was set to 640. Data augmentation techniques included Mosaic (with a probability of 1.0), Mixup (with a probability of 0.0, i.e., disabled), and horizontal flipping (with a probability of 0.5). Due to the unique characteristics of multispectral data, HSV color space adjustment was not utilized in this study. During the inference phase, Non-Maximum Suppression (NMS) was applied with an Intersection over Union (IoU) threshold set to 0.7.

4.3. Ablation Experiments

4.3.1. Ablation Study on Input Spectral Bands

To verify the effectiveness of multispectral image inputs, we trained and tested eight detectors using both the visible version of the MODA dataset and its multispectral counterpart. We only modified the detectors by changing the number of input channels in the first layer to match the number of multispectral bands; the MODA dataset used in this work contains 8 channels. The experimental results are summarized in

Table 1. Comparison of visible and multispectral inputs.

Model	Images	Pre	Rec	F1-Score	mAP50	mAP50:95	GFLOPs	Para (M)
YOLOv5n	RGB	0.626	0.592	0.609	0.608	0.443	7.3	2.58
YOLOv5n	MSI	0.648	0.602	0.624	0.632	0.468	7.9	2.58
YOLOv5s	RGB	0.674	0.621	0.646	0.657	0.487	24.8	9.40
YOLOv5s	MSI	0.705	0.635	0.668	0.692	0.523	26	9.40
YOLOv8n	RGB	0.636	0.576	0.605	0.604	0.442	8.3	3.08
YOLOv8n	MSI	0.654	0.607	0.630	0.639	0.474	8.5	3.08
YOLOv8s	RGB	0.664	0.622	0.642	0.655	0.488	29.4	11.41
YOLOv8s	MSI	0.686	0.642	0.663	0.684	0.52	29.7	11.42
YOLO11n	RGB	0.652	0.587	0.618	0.625	0.458	6.6	2.66
YOLO11n	MSI	0.671	0.608	0.638	0.65	0.482	6.7	2.66
YOLO11s	RGB	0.675	0.632	0.653	0.663	0.486	22.3	9.70
YOLO11s	MSI	0.698	0.644	0.670	0.692	0.529	22.6	9.70
YOLOv12n	RGB	0.632	0.588	0.609	0.611	0.444	6.6	2.63
YOLOv12n	MSI	0.662	0.599	0.629	0.638	0.478	6.7	2.63
YOLOv12s	RGB	0.685	0.616	0.649	0.666	0.493	22.2	9.52
YOLOv12s	MSI	0.704	0.643	0.672	0.695	0.53	22.5	9.52

Note: The best-performing results highlighted in bold.

, where the best results are highlighted in bold. Multispectral input demonstrates clear effectiveness in the object detection task.

Compared with visible-light input, models using multispectral images achieve consistent improvements across five key metrics: Precision, Recall, F1-Score, mAP₅₀, and mAP_50:95. On average, mAP₅₀ increases by 2.9 percentage points and mAP_50:95 increases by 3.3 percentage points. This trend remains consistent across different model scales and detector variants, demonstrating the general benefit of multispectral information for improving detection performance. Notably, the additional computational cost introduced by multispectral input is small: the largest increase in computation occurs on YOLOv5n, but it is only 0.6 GFLOPs, while the parameter count remains essentially unchanged. This indicates that multispectral inputs improve detection accuracy while maintaining good computational efficiency. Overall, these results confirm that the richer spectral information provided by multispectral images can effectively enhance the model’s representation capability, thereby improving the accuracy and robustness of object detection.

4.3.2. Module Ablation Study

To evaluate the effectiveness of the proposed modules, we conducted comprehensive ablation experiments using YOLO11n and YOLO11s as baselines.

Table 2. Ablation study of SG-YOLOn on MODA Dataset.

Model	SGDS	SSIAF	SCSF	Para (M)	GFLOPs	Pre	Rec	mAP50	mAP50:95
A	×	×	×	2.66	6.7	0.671	0.608	0.65	0.482
B	√	×	×	2.65	6.7	0.674	0.624	0.662	0.488
C	√	√	×	2.73	7.7	0.681	0.633	0.669	0.493
D	√	√	√	2.73	7.6	0.689	0.634	0.679	0.501

Note: The best-performing results highlighted in bold. “√” indicates inclusion, while “×” denotes exclusion.

and

Table 3. Ablation study of SG-YOLOs on MODA Dataset.

Model	SGDS	SSIAF	SCSF	Para (M)	GFLOPs	Pre	Rec	mAP50	mAP50:95
A	×	×	×	9.70	22.6	0.698	0.644	0.692	0.529
B	√	×	×	9.68	21.7	0.722	0.66	0.711	0.537
C	√	√	×	9.71	22	0.713	0.675	0.72	0.547
D	√	√	√	9.99	25.2	0.723	0.675	0.724	0.553

Note: The best-performing results highlighted in bold. “√” indicates inclusion, while “×” denotes exclusion.

present two extensive ablation studies that assess the individual and combined effects of the proposed SGDS, SSIAF, and SCSF modules on the performance of the SG-YOLO model.

A, B, C, and D denote the baseline network and the networks obtained by progressively adding the SGDS, SSIAF, and SCSF modules, respectively. First, we compare the baseline (Group A) with Group B, which contains only the SGDS module. For SG-YOLOn, mAP₅₀ increases from 0.65 to 0.662 and mAP_50:95 rises from 0.482 to 0.488, while the parameter count and computational cost remain essentially unchanged (Parameters: 2.66 M to 2.65 M; GFLOPs: unchanged at 6.7). The improvement is more pronounced for SG-YOLOs: mAP₅₀ increases by 2.75% and Precision by 3.44%, with slight reductions in parameters (9.70 M to 9.68 M) and computation (22.6 to 21.7 GFLOPs). These results indicate that SGDS effectively enhances multispectral object detection capability without increasing model size or computational burden.

Next, after progressively introducing SSIAF on top of SGDS (Group C), SG-YOLOn achieves mAP₅₀ = 0.669 and Recall improves by 1.2% (0.624 to 0.635). For SG-YOLOs, mAP₅₀ increases to 0.720 and Recall improves by 1.5% (0.66 to 0.675), with parameters and computation remaining largely stable. With the further addition of SCSF (Group D), SG-YOLOn reaches mAP₅₀ = 0.679 and mAP_50:95 = 0.501, with Precision improving accordingly, while computation drops back to 7.6 GFLOPs. SG-YOLOs further improves to mAP₅₀ = 0.724 and mAP_50:95 = 0.553, with Precision rising to 0.723. Although parameters and computation increase slightly, overall performance continues to improve, confirming that the three modules work synergistically without redundancy.

To evaluate the detection performance of the proposed network on multi-scale targets, we categorize the objects according to the standard COCO evaluation metrics [51] based on their pixel areas: small objects ($\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}<32^2$), medium objects ($32^2\le \mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}<96^2$), and large objects ($\mathrm{a}\mathrm{r}\mathrm{e}\mathrm{a}\ge 96^2$). We compared SG-YOLO with the baseline network, and the results are presented in

Table 4. Comparison of detection accuracy between SG-YOLO and baseline models for objects at different scales.

Model	APs	APm	APl	MaxDets
YOLO11n	0.139	0.373	0.410	100
SG-YOLOn	0.163 (+2.4%)	0.412 (+3.9%)	0.462 (+5.2%)	100
YOLO11s	0.166	0.413	0.471	100
SG-YOLOs	0.203 (+3.7%)	0.442 (+2.9%)	0.512 (+4.1%)	100

Note: The best-performing results highlighted in bold.

. The experimental results indicate that, whether for small, medium, or large objects, SG-YOLO achieves varying degrees of improvement in average precision compared to the baseline. This demonstrates that the proposed network exhibits more robust and superior detection capabilities across all object scales.

As shown in Figure 11 and Figure 12, we compare AP across different categories for SG-YOLOn and SG-YOLOs. For both the n and s models, the Group D model achieves the best detection accuracy in the majority of categories. This suggests that our method does not merely improve detection for a single class but delivers consistent gains across all categories to varying degrees. Overall, the proposed approach achieves substantial improvements in detection performance with only a modest increase in parameters and computation, fully validating the rationale and synergistic effectiveness of the proposed module combination.

4.3.3. Visualization Results

To better illustrate the improvements of SG-YOLO over the baseline network, we selected several test images from diverse scenarios and conducted a qualitative comparison with the baseline. As shown in Figure 13, we present the multispectral detection results of YOLO11s and SG-YOLO under different scenes. In the first set of images, due to the direct compression of multispectral data, the baseline model mistakenly detects flowers on a tree as pedestrians, whereas SG-YOLO avoids this false detection by filtering band information at an early stage. In the second set (top-right), the shadow cast by an electric scooter leads the baseline model to misclassify it as a car based on its shape. In the third set, patterns on the ground cause the baseline model to incorrectly recognize markings on the basketball court as pedestrians. In contrast, SG-YOLO makes full use of spectral information and effectively reduces false positives through early band filtering and subsequent spectral-attention fusion.

To further investigate the mechanism of background suppression and feature enhancement in SG−YOLO, we employed the Grad−CAM technique to visualize the activation maps of both the baseline model and our proposed SG−YOLOs using the cases from Figure 13. As illustrated in Figure 14, SG−YOLOs consistently exhibits superior background noise suppression across various scenarios. For instance, in the first row, the response intensity on the terrain surrounding the road is markedly lower than that of the baseline, while the activation on the target itself is significantly higher. A similar trend is observed in the subsequent image sets: the baseline model displays excessive attention toward background elements such as forests, which inevitably degrades detection precision. In contrast, SG−YOLOs demonstrates higher responsiveness in target regions and effectively suppresses large-scale background interference (e.g., forests and rooftops). Given that these background elements often share similar spectral signatures, these results validate that our proposed network effectively leverages spectral information to distinguish targets from complex backgrounds, thereby enhancing target saliency while minimizing false positives.

4.4. Comparison of Experiment with Previous Methods

We compared the performance of several mainstream detectors—YOLOv5, YOLOv8, YOLO11, and YOLOv12—on the MODA dataset, and further included two methods specifically designed for multispectral object detection, S2ADet [52] and OSSDet [16], the experimental results are summarized in

Table 5. Comparison results with other detection algorithms.

Method	Car	Bike	Pedestrian	Van	Truck	Bus	Tricycle	Aw-Bike	mAP50	mAP50:95	GFLOPs	Para (M)
StripRCNN [53]	0.905	0.452	0.307	0.761	0.671	0.890	0.565	0.736	0.661	0.410	231.8	45.2
LSKNet-S2Anet [54]	0.905	0.389	0.431	0.761	0.686	0.895	0.518	0.725	0.664	0.386	194.9	29.9
S2ADet [52]	0.903	0.350	0.408	0.721	0.572	0.866	0.541	0.713	0.635	0.389	406.0	65.2
OSSDet [16]	0.905	0.453	0.457	0.792	0.697	0.899	0.588	0.727	0.690	0.427	263.1	36.5
YOLOv5n	0.964	0.462	0.539	0.665	0.535	0.871	0.325	0.693	0.632	0.468	7.9	2.58
YOLOv5s	0.972	0.55	0.591	0.725	0.621	0.904	0.435	0.736	0.692	0.523	26.0	9.41
YOLOv8n	0.965	0.47	0.543	0.683	0.521	0.901	0.352	0.965	0.639	0.474	8.5	3.08
YOLOv8s	0.971	0.558	0.601	0.727	0.593	0.895	0.388	0.74	0.684	0.52	29.7	11.42
YOLO11n	0.941	0.459	0.525	0.695	0.556	0.91	0.416	0.679	0.65	0.482	6.7	2.66
YOLO11s	0.968	0.55	0.614	0.698	0.607	0.907	0.433	0.761	0.692	0.529	22.6	9.70
SG-YOLOn	0.969	0.499	0.529	0.734	0.6	0.938	0.462	0.703	0.679	0.501	7.6	2.73
SG-YOLOs	0.973	0.58	0.619	0.735	0.647	0.94	0.51	0.785	0.724	0.553	25.2	9.99

Note: The best-performing results highlighted in bold.

. The results show that the proposed SG-YOLOs achieves the best detection accuracy for most categories. Although it underperforms StripRCNN and LSKNet-S2ANet on a few classes, our method requires substantially fewer parameters and far less computation than these two heavy models (231.8 GFLOPs and 194.9 GFLOPs, respectively). Moreover, SG-YOLOs attains the top performance on both mAP₅₀ and mAP_50:95, indicating the best overall detection accuracy among the compared approaches. In addition, SG-YOLOn delivers the best detection results within the n-scale model family. It is worth noting that our method slightly increases the parameter count and computational complexity compared with YOLO11; however, the increase is marginal and does not hinder deployment on edge devices such as UAV platforms.

Figure 15 presents qualitative comparisons of different detectors across various scenes. Red circles indicate false positives, while green circles highlight missed detections. As can be observed, in cluttered backgrounds, UAV targets often occupy only a very small portion of the image, placing stringent demands on a detector’s multi-scale capability. Moreover, although spectral information can provide additional cues, it may also introduce noise and interference.

As shown in the second and third rows of Figure 14, some detectors directly compress the spectral data for detection, making them susceptible to background noise and resulting in false positives—e.g., rooftop clutter, a flag-raising platform, and road stripes under trees are mistakenly detected as targets. In contrast, our method effectively mitigates these false alarms by selectively filtering band information. Meanwhile, for small-object detection in complex backgrounds, other detectors tend to suffer from a large number of missed detections. Although our method does not achieve perfect detection, it substantially reduces both missed targets and false positives, further underscoring the high level of difficulty of object detection from a UAV viewpoint.

To verify the generalization capability of the proposed SG-YOLO, we conducted comparative evaluations against other detectors on the public MMOT dataset [17]. This dataset also contains multispectral UAV imagery with targets at various scales and provides eight spectral bands. As shown in

Table 6. Detection results of different detectors on the MMOT dataset.

Method	Car	Bike	Pedestrian	Van	Truck	Bus	Tricycle	Aw-Bike	mAP50	mAP50:95	GFLOPs	Para (M)
YOLOv5n	0.936	0.284	0.61	0.627	0.305	0.847	0.135	0.557	0.538	0.395	7.9	2.58
YOLOv5s	0.955	0.371	0.66	0.716	0.38	0.851	0.239	0.593	0.596	0.457	26.0	9.41
YOLOv8n	0.944	0.313	0.589	0.629	0.412	0.797	0.133	0.581	0.55	0.411	8.5	3.08
YOLOv8s	0.948	0.422	0.677	0.662	0.378	0.877	0.263	0.619	0.606	0.468	29.7	11.42
YOLO11n	0.943	0.33	0.612	0.621	0.45	0.781	0.169	0.563	0.559	0.423	6.7	2.66
YOLO11s	0.948	0.393	0.672	0.687	0.455	0.869	0.213	0.615	0.607	0.470	22.6	9.70
SG-YOLOn	0.955	0.396	0.622	0.644	0.365	0.91	0.27	0.576	0.592	0.443	7.6	2.73
SG-YOLOs	0.955	0.479	0.664	0.665	0.425	0.942	0.339	0.619	0.636	0.482	25.2	9.99

Note: The best-performing results highlighted in bold.

, SG-YOLO consistently achieves the best detection performance across multiple categories and attains the highest scores on both mAP₅₀ and mAP_50:95, demonstrating the superiority of the proposed method.

4.5. Inference Speed Analysis and Deployment Potential

To substantiate the claim that SG-YOLO is suitable for real-time perception on resource-constrained UAV platforms, we conducted a comprehensive analysis of its inference efficiency and deployment potential. First, we measured the actual inference speed of the proposed SG-YOLO variants on a desktop-grade GPU (NVIDIA GeForce RTX 4060, NVIDIA Corporation, Santa Clara, CA, USA) as a performance baseline. All models were tested with a batch size of 1 and an input resolution of 640 × 640 pixels under the PyTorch framework. The results, including the number of parameters, computational complexity (GFLOPs), and the critical metric of actual measured frames per second (FPS), are summarized in

Table 7. Inference speed comparison of different models on an RTX 4060 GPU.

Model	Params (M)	GFLOPs	Infrence Speed (FPS)
SG-YOLOn	2.73	7.6	641.7
SG-YOLOs	9.99	25.2	374.3

.

Even the more capable SG-YOLOs variant achieves a remarkable inference speed of 374.3 FPS on the desktop GPU, while the ultra-lightweight SG-YOLOn exceeds 640 FPS. This provides concrete, hardware-validated evidence of the model’s low-latency characteristics. To translate this efficiency into the context of edge deployment, we analyze the computational requirements against the capabilities of modern embedded processors commonly used in UAVs. The most computationally intensive variant, SG-YOLOs, requires 25.2 GFLOPs per inference. For a standard real-time perception requirement of 30 FPS, this translates to a sustained computational demand of approximately 0.756 TFLOPs (25.2 GFLOPs × 30). This demand is well within the theoretical peak performance of prevalent UAV-embedded AI platforms, such as the NVIDIA Jetson Orin Nano 8 GB (up to 20 TFLOPs FP16) and the Jetson AGX Xavier (up to 11 TFLOPs FP16). A conservative estimation, which accounts for practical factors like memory bandwidth and typical accelerator utilization rates, strongly suggests that SG-YOLOs can reliably achieve frame rates well above 30 FPS on such platforms. The SG-YOLOn variant, with its minimal computational footprint of only 7.6 GFLOPs, is projected to operate at extremely high frame rates, thereby freeing substantial computational headroom for other critical onboard tasks such as sensor fusion, path planning, or communications.

In conclusion, the combination of high measured FPS on standard hardware and a favorable theoretical performance profile on target edge devices confirms that the lightweight design of SG-YOLO is not merely a theoretical advantage but a practical enabler for efficient deployment.

5. Conclusions

This paper proposes the SG-YOLO framework, an efficient small-object detection solution tailored for multispectral UAV aerial imagery. The framework employs the proposed SGDS module to perform early-stage feature extraction across different spectral bands and gated band selection, enabling effective discrimination of spectral feature importance. The SSIAF module leverages cascaded feature-attention fusion to integrate high-level semantic information across spectra, enhancing both local and global channel interactions while suppressing channel redundancy. In addition, SCSF strengthens multi-scale feature fusion, preserving channel interaction while enabling effective cross-scale aggregation.

Extensive experiments on the MODA dataset demonstrate that using multispectral imagery yields higher detection accuracy than using visible-light images alone. Moreover, SG-YOLO achieves strong performance in both accuracy and efficiency, validating its suitability for real-time multispectral UAV perception.

Beyond its direct application in UAV remote sensing, the design principles of SG-YOLO offer transferable insights for broader research areas. The SGDS module presents a paradigm for early-stage, input-aware gating and de-redundancy in multi-modal data fusion, which is applicable to other scenarios involving high-dimensional, correlated inputs, such as autonomous driving sensor fusion (camera, LiDAR, radar) or multi-modal medical image analysis. Nevertheless, object detection from UAV viewpoints remains highly challenging. In extreme scenarios with severe background interference and minimal spectral differences among categories, the proposed method still requires improved feature extraction and background suppression. Future work will focus on two key directions: enhancing the model’s performance in complex backgrounds and conducting direct deployment and performance validation on specific embedded platforms (e.g., the Jetson series). These efforts aim to further improve the model’s generalizability across different UAV platforms and challenging environments, thereby broadening its applicability in UAV remote sensing tasks.