A lightweight deep learning framework for early fire and smoke detection in forest fire monitoring

Efficient early forest fire detection is important for mitigating wildfire spread and related ecological and socio-economic losses. However, vision-based detection remains challenging because fire and smoke targets often exhibit large-scale variation, irregular morphology, low contrast, blurred boundaries, and strong interference from complex forest backgrounds. This study proposes FRW-YOLOv8, a lightweight YOLOv8-based model for fire and smoke detection in UAV-perspective forest scenes. FasterNeXt is introduced into the backbone to reduce redundant computation while preserving feature representation. A reparameterized feature pyramid network, composed of SimConv, transposed convolution, and RepBlock, is designed to enhance cross-scale feature interaction and semantic fusion. Wise-IoU is further adopted to improve bounding-box regression for small and irregular fire-related targets. The model was trained and internally evaluated on the M4SFWD dataset, which provides diverse UAV-perspective forest fire scenes. On M4SFWD, FRW-YOLOv8 improved Fire AP50 and Smoke AP50 by 1.0 and 2.1 percentage points, respectively, while mAP50 and mAP50-95 increased by 1.6 and 1.7 percentage points compared with YOLOv8n. Parameters and FLOPs were reduced by 20.67 and 23.46%, respectively. Although M4SFWD provides diverse UAV-perspective forest fire scenes for model evaluation, its synthetic nature still introduces certain limitations; therefore, two independent real-world external datasets were further evaluated to assess real-scene generalization. FRW-YOLOv8 achieved a Fire AP50 of 88.3% and a Fire AP50-95 of 56.8%. It also achieved a Smoke AP50 of 93.5% and a Smoke AP50-95 of 57.9%, outperforming YOLOv8n by 1.2 and 1.3 percentage points for fire, and by 1.7 and 0.2 percentage points for smoke, respectively. These results indicate an improved trade-off between detection accuracy and model complexity and suggest potential for resource-constrained forest fire monitoring.

As a cornerstone of the Earth’s biosphere, forest ecosystems harbor indispensable phytodiversity, orchestrate the global carbon cycle, and function as critical terrestrial carbon sinks. By anchoring soil substrates through dense vegetative root systems, forests effectively mitigate surface runoff and edaphic erosion, thereby playing a pivotal role in regulating hydrological cycles (Lin et al., 2024; Singh, 2022). Consequently, these ecosystems are irreplaceable (Carta et al., 2023) in sustaining biodiversity, buffering the global climate, and maintaining ecological equilibrium. Although wildfire is traditionally conceptualized as a destructive disturbance, moderate fire regimes can stimulate pyrophyte seed germination, facilitate adaptive ecological succession, and modulate community structures. However, increasingly frequent wildfires driven by anthropogenic activity and climate change can exceed the evolutionary carrying capacity and self-regulatory resilience of these ecosystems, inflicting irreversible degradation on forest flora, community dynamics, and landscape integrity (Saffre et al., 2022; Kantarcioglu et al., 2023). Furthermore, catastrophic wildfires mobilize vast pyrogenic carbon emissions, shifting forests from carbon sinks to net sources, which feed back into global warming trends (Pang et al., 2022). Exacerbated by climate change, the increasing prevalence of extreme meteorological anomalies has intensified the frequency and severity of forest fires, positioning them as a major threat to forest ecological security, human phytogenetic resources, and global ecological stability (Guede-Fernández et al., 2021; Barmpoutis et al., 2020). The socio-economic and ecological repercussions are profound; beyond degrading habitats and disrupting plant–animal co-evolutionary networks (Naderpour et al., 2021; Sathishkumar et al., 2023), these disasters damage critical infrastructure, crops, and forestry assets (Abid, 2021). A representative example is the late 2019 Australian megafires, which incinerated approximately 20% of the continental forest cover, destroyed over 3,000 homes, claimed 33 lives, and decimated vast populations of wildlife, driving at least 34 endemic species toward the brink of extinction while incurring billions of dollars in losses (Luan et al., 2024). Therefore, the rapid and precise detection of initial ignition points is paramount to safeguarding forest ecosystems and advancing proactive fire mitigation strategies (Rubí et al., 2023).

Forest fires typically erupt in geographically remote or topographically complex terrain, characterized by stochastic outbreaks, rapid propagation, and substantial challenges in post-ignition suppression (Zhang et al., 2021). Timely early warning during the nascent phase is critical to arresting fire spread and maximizing management efficiency. Crucially, early ignitions often manifest as smoldering behavior or faint smoke plumes that evade conventional visual detection; by the time open canopy fires ignite, the optimal window for ecological intervention has passed (Sun et al., 2022). This challenge is exacerbated by sparse communication infrastructure, limited forestry personnel, and complex wilderness terrain, which impede prompt data feedback (Khan et al., 2022). Furthermore, heterogeneous forest backdrops, fluctuating illumination, and variable microclimatic conditions introduce severe noise into optical signals, yielding high rates of missed detections and false alarms (Xu et al., 2021). Traditional monitoring frameworks rely heavily on manual patrols and fixed watchtowers. Although intuitive, these methods are constrained by adverse weather, visual dead zones, and human fatigue, making them inadequate for vast, contiguous forest landscapes (Deng and Zhang, 2021; Lei et al., 2024). While localized point sensors (e.g., infrared and smoke detectors) improve responsiveness, they remain prone to environmental cross-sensitivity and delayed triggers in open-air forest dynamics (Yun et al., 2025; Wang et al., 2024a). Satellite remote sensing provides synoptic coverage but suffers from coarse spatial resolution and cloud attenuation, limiting continuous, real-time micro-scale monitoring (Chuvieco et al., 2020). To bridge this gap, Unmanned Aerial Vehicles (UAVs) have emerged as a promising low-altitude remote sensing platform, offering high maneuverability, flexible deployment, and real-time acquisition of high-resolution canopy imagery. Nevertheless, standard UAV-based vision monitoring still faces technical bottlenecks regarding target recognition accuracy and background false alarms in structurally complex forest environments (Liu et al., 2023; Srivastava et al., 2022).

Recent advances in UAV remote sensing, LiDAR, and computer vision, particularly deep-learning-based object detection have propelled forest monitoring toward an automated, high-precision paradigm. These algorithms can effectively extract fine-grained spatial features of microscale flames and early-stage smoke, circumventing human visual limitations and providing technical support for precision forestry conservation (Liu et al., 2026). To contextualize these developments, several representative architectures warrant evaluation. Sung et al. (2025) developed the FSP-YOLO model, embedding GhostNet and coordinate attention to balance accuracy and model parameterization, yet its robustness under extreme forest illumination remains suboptimal. Yu et al. (2025) integrated deformable convolutions and HIoU loss to enhance small-target localization, though at the expense of computational overhead. For the FFYOLO framework, Yun et al. (2024) designed a CPDA attention module to bolster feature extraction alongside GSConv modules for parameter reduction; however, such lightweighting strategies occasionally compromise fine-textured smoke features. Zhou et al. (2025) introduced the YOLOv8n-SMMP model utilizing a SlimNeck structure, which reduced computational complexity but remained insufficiently sensitive to low-contrast smoke against complex vegetative backgrounds. Chen et al. (2025) proposed SRoFF-Yolover, replacing standard convolutions with dilated convolutions to expand the receptive field, though it occasionally suffered from feature sparsity. Mo et al. (2025) formulated the YOLO-MFD model, addressing smoke and flame morphing via spatial-deformable and frequency-domain Fourier fusion (SAPM), but the associated multi-scale modules (FAWM) increased inference latency and hardware deployment costs. Han Y. et al. (2024) presented LUFFD-YOLO, leveraging GhostNetV2 and multi-head self-attention (ESDC2f) for small-target detection, yet the structural synergy between its submodules remained unoptimized. Zhang et al. (2024) engineered the YOLOv8-FEP model featuring a PAN-Bag structure to optimize multi-level feature fusion, which improved accuracy but introduced computational redundancy. Wang H. et al. (2025) developed DSS-YOLO by substituting backbone components with DynamicConv and adding a SPPELAN module, yielding only marginal precision gains. Li et al. (2025) balanced speed and accuracy via a Gather-Distribute feature mechanism, SimAM attention, and Focal-DIoU loss, though module complexity remained high for resource-constrained edge devices. Sheng et al. (2024) proposed FireYOLO-Lite, a lightweight forest fire detection model based on YOLOv8. The model employs GhostNetV2 for lightweight feature extraction, introduces the CPDCA attention mechanism to enhance multi-scale spatial feature representation, and adopts an Inner-IoU-based loss to improve the localization of small fire targets. Finally, Han X. et al. (2024) developed a task-specific Attentive Fire and Smoke Detection Model, termed a-FSDM, which redesigns the detection head using an Attentive Transparency Detection Head to enhance transparent fire and smoke features across multiple scales and channels. In addition, a relative burning intensity representation was introduced to support downstream fire-risk assessment. Wang Z. et al. (2025) further proposed FFD-YOLO for forest fire detection by modifying both the backbone and feature-fusion network of YOLOv8. A Light Edge Information Extraction Module was embedded in the backbone to capture flame-boundary and spatial information more effectively, while Strip Large Kernel Spatial Attention was introduced to strengthen the representation of flames with substantial scale variations. This spatial attention mechanism was further combined with Efficient Channel Attention to construct a Dynamic Fusion module, which replaced the original upsampling components and adaptively integrated multi-level features. In addition, a pseudo-fire dataset containing flame-like interference sources was constructed to improve the model’s ability to distinguish real flames from visually similar objects. Beyond the fire and smoke detection domain, Fındıkçı et al. (2026) developed Melan-YOLO for underwater object detection using a dual-strategy framework that separately addresses high-accuracy detection and resource-constrained deployment. Melan-YOLO-Pro incorporates asymmetric convolutions and RepNCSPELAN4 to enhance feature representation under degraded underwater imaging conditions, whereas Melan-YOLO-Edge combines GhostConv, CBAM, and ShuffleAttention to reduce computational requirements while retaining discriminative features. Both variants also employ an underwater image-enhancement pipeline and SIoU loss to improve the localization of irregularly shaped marine debris. Although Melan-YOLO addresses a different application domain, its customized architectural design and explicit treatment of the accuracy–efficiency trade-off provide relevant methodological insights for lightweight fire and smoke detection. Collectively, these studies demonstrate that coordinated optimization of edge representation, multi-scale feature fusion, localization loss, and lightweight architecture is important for detecting irregular targets in complex environments.

Lightweight model development is particularly relevant to forest fire monitoring because UAV-based and other edge visual monitoring platforms are often constrained by onboard computational resources, memory capacity, energy supply, and communication conditions. Compared with larger detection models, lightweight architectures can reduce model storage and computational requirements, thereby providing a more practical basis for local and low-latency fire and smoke detection. This is particularly important when high-resolution imagery must be processed continuously and communication with remote computing resources may be limited or unstable. However, lightweighting is not an end in itself, because excessive structural simplification may potentially weaken the representation of extremely small flames, diffuse smoke, and low-contrast boundaries. Therefore, this study aims to improve the trade-off among detection accuracy, localization reliability, model complexity, and inference throughput, rather than merely minimizing the parameter count or computational cost. To address the above challenges, this study proposes FRW-YOLOv8, a lightweight fire and smoke detection model based on YOLOv8n. In the backbone, the standard C2f module is replaced with FasterNeXt to reduce redundant computation while maintaining feature representation capability for flame and smoke targets. This modification decreases the number of parameters and floating-point operations, making the model more suitable for resource-constrained visual detection scenarios. In addition, a reparameterized feature pyramid network, termed RepFPN, is introduced to improve multi-scale feature fusion. RepFPN uses transposed convolutions for upsampling and SimConv for efficient feature transformation, thereby enhancing the interaction between shallow spatial details and deep semantic features. During training, multi-branch RepBlocks are used to strengthen feature representation. During inference, these branches can be reparameterized into single convolutional layers, which simplifies the inference structure without introducing additional computational branches. This design is helpful for detecting small flames and diffuse smoke regions whose features may be weakened during deep feature extraction. Furthermore, Wise-IoU is adopted as the bounding-box regression loss. By dynamically adjusting the gradient gain according to anchor-box quality, Wise-IoU guides the model to focus more on samples with moderate localization quality, which improves the stability of bounding-box regression for irregular flame and smoke targets. The proposed model was evaluated on M4SFWD, a UAV-perspective forest fire dataset constructed using Unreal Engine 5 that provides diverse forest scenes, target scales, weather conditions, and illumination settings for controlled model assessment. In addition, two independent real-world external datasets, FLAME dataset and the Wildfire Smoke dataset, were introduced to further examine the model’s generalization ability on real fire and smoke images. Through this evaluation strategy, this study aims to develop a lightweight and effective visual detection approach for early fire and smoke recognition, providing technical support for forest fire monitoring, early warning, fire management, and forest ecosystem protection.

Datasets

This study used M4SFWD dataset for model training and internal evaluation, together with two independent real-world datasets for external testing. The M4SFWD dataset was used to assess the detection performance of the model under controlled multi-scene conditions, whereas the real datasets were introduced to examine the generalization ability of the model from synthetic UAV-perspective images to real flame and smoke images. Representative examples from the M4SFWD dataset and the two external fire and smoke datasets are shown in Figure 1, illustrating the diversity of scene backgrounds, target appearances, and imaging conditions considered in this study.

M4SFWD dataset

To evaluate the detection performance of the proposed model in structurally complex and ecologically diverse environments, the M4SFWD dataset was used in this study. This large-scale remote sensing dataset was constructed using Unreal Engine 5 to simulate multi-scale forest fire scenarios from an unmanned aerial vehicle perspective. To reproduce the complex backgrounds and structural variations commonly observed in real-world forest ecosystems, the dataset contains eight representative geographic scenes, including dense forest canopies, subalpine mountainous areas, and volcanic landscapes. These scenes are further combined with three meteorological conditions, namely sunny, foggy, and rainy or snowy weather, as well as three illumination profiles, including daytime, dusk, and night. Such a combination effectively simulates the variations in lighting, visibility, and atmospheric interference that commonly occur in practical forest fire monitoring. In total, the dataset contains 3,974 high-resolution images with 17,763 manually annotated bounding boxes of fire and smoke targets. The original images preserve realistic resolution variations, such as 1,480 × 684 and 1,280 × 720, reflecting the non-uniform image acquisition characteristics of unmanned aerial vehicle sensors under complex field conditions. However, images with different aspect ratios and resolutions may introduce scale inconsistency during model training and affect convergence stability. Therefore, all images were resized to a uniform resolution of 640 × 640 pixels during both the training and testing stages to ensure experimental consistency. The dataset was randomly partitioned at the individual-image level into fixed training, validation, and test subsets containing 2,820, 763, and 391 images, respectively, with 12,666, 3,381, and 1,716 annotated instances. No explicit stratification by geographic scene, meteorological condition, or illumination condition was imposed. Given the limited number of available source scenarios and the uneven distribution of terrain, weather, illumination, and target configurations among them, a strict scene-exclusive split could substantially reduce environmental-condition coverage in one or more subsets. The fixed image-level partition was therefore retained as a practical compromise to maintain broad and comparable coverage of the available scenarios and environmental conditions across the three subsets. The same fixed subsets were used for all baseline, ablation, comparison, and proposed-model experiments to ensure controlled and reproducible comparisons. The dataset defines two target categories: “fire” and “smoke.”

FLAME dataset and Wildfire Smoke dataset

To further evaluate the cross-domain generalization ability of the proposed model, the FLAME dataset and the Wildfire Smoke dataset were introduced as independent external datasets. The FLAME dataset was originally introduced by Shamsoshoara et al. (2021), whereas the Wildfire Smoke dataset used in this study was obtained from the public dataset released by AI for Mankind in collaboration with HPWREN. Detailed information regarding the repositories, dataset versions, licenses, and access links is provided in the Data Availability Statement. The FLAME dataset was used for fire-specific evaluation, whereas the Wildfire Smoke dataset was used for smoke-specific evaluation. Unlike M4SFWD, which was used for model training, validation, and internal testing, the two external datasets were used exclusively for independent testing and were not involved in model training, fine-tuning, model selection, domain adaptation, or threshold optimization.

The downloaded version of the FLAME dataset used in this study contains 2,000 images and 10,312 annotated fire instances, with an image resolution of 512 × 512 pixels. The Wildfire Smoke dataset contains 737 images and 737 annotated smoke instances, with an image resolution of 640 × 480 pixels. Both datasets are single-class object detection datasets. Specifically, flame contains only the fire category, whereas Wildfire Smoke contains only the smoke category. The annotations of both datasets are provided in YOLO bounding-box format, in which each target is represented by a class index and normalized bounding-box center coordinates, width, and height.

All available images and valid annotations in the downloaded dataset archives were included in the external evaluation. No samples were selected or excluded according to model predictions, target size, scene difficulty, image quality, or detection performance. Before evaluation, the correspondence between the image and annotation files was checked, and no empty, invalid, or out-of-range annotations were identified. All images were resized to 640 × 640 pixels during model evaluation.

Compared with the synthetic M4SFWD dataset, the two external datasets contain real-world fire and smoke images with variations in target scale, illumination, background distribution, imaging conditions, and scene characteristics. The FLAME dataset includes flame targets with irregular shapes, nonuniform brightness distributions, blurred boundaries, and varying scales. The Wildfire Smoke dataset contains smoke targets characterized by weak texture information, low contrast, diffuse contours, and visual similarities to clouds, fog, dust, and complex backgrounds. Therefore, the two datasets provide complementary evidence for assessing the generalization performance of the proposed model in unseen real-world scenarios.

The semantic definitions of fire and smoke in the two external datasets are consistent with the corresponding categories in M4SFWD, although their numerical class indices are not fully identical. In M4SFWD, fire and smoke are encoded as class 0 and class 1, respectively. In the FLAME dataset, the single fire category is encoded as class 0. Therefore, class-0 fire predictions generated by the M4SFWD-trained detector were directly evaluated against the class-0 fire annotations, whereas class-1 smoke predictions were excluded from the flame evaluation.

In the Wildfire Smoke dataset, the single smoke category is encoded as class 0, whereas the M4SFWD-trained detector outputs smoke as class 1. Before metric computation, class-1 smoke predictions were selected from the detector outputs and their class labels were remapped from class 1 to class 0 to match the annotation index of the Wildfire Smoke dataset. The predicted bounding-box coordinates and confidence scores remained unchanged during this operation. Class-0 fire predictions were excluded because no fire annotations were available in the Wildfire Smoke dataset. This class-index remapping was performed only within the external-evaluation pipeline and did not alter model training, model parameters, or the semantic definitions of the detector outputs. Accordingly, Fire AP50 and Fire AP50–95 were reported for the FLAME dataset, whereas Smoke AP50 and Smoke AP50–95 were reported for the Wildfire Smoke dataset.

The final YOLOv8n and FRW-YOLOv8 checkpoints trained on M4SFWD were directly evaluated on the two external datasets without retraining or fine-tuning. The same image-preprocessing procedure, confidence threshold, NMS configuration, category-selection and class-index-remapping procedure, prediction–ground-truth matching rule, and AP calculation script were applied to both models.

Methods

YOLOv8

YOLOv8 (Varghese and Sambath, 2024) is a new-generation object detection model released by Ultralytics in 2023. It inherits the high efficiency and strong detection accuracy of the YOLO series in real-time detection tasks, while incorporating several architectural improvements to further enhance both speed and accuracy. Its backbone network integrates key optimization modules, including CBS, C2f, and SPPF. These modules improve feature extraction efficiency, enrich gradient flow, and enhance multi-scale receptive field fusion, enabling the model to better adapt to targets with large-scale variations and irregular morphologies, such as flames. In addition, YOLOv8 introduces an important improvement in its detection framework by replacing the traditional anchor-based strategy with an anchor-free structure. This structure directly performs object localization and bounding box prediction through center-point regression, eliminating the need for dataset-specific anchor-box clustering and parameter tuning. As a result, the training process is simplified, and the model becomes more flexible in detecting targets with blurred boundaries and irregular shapes, such as flames and smoke. In outdoor forest fire monitoring scenarios, where complex illumination conditions and severe background interference are common, YOLOv8 can maintain robust feature representation and strong generalization capability. Therefore, it provides an effective technical foundation for real-time forest fire monitoring and early warning.

The proposed FRW-YOLOv8 network

While YOLOv8 has demonstrated competitive performance in flame detection, it remains vulnerable to background interference, missed detections of small targets, and limited localization precision in complex outdoor environments. Consequently, there is still significant potential to enhance the detection accuracy and robustness for fire and smoke targets. To address these limitations, this paper proposes a lightweight fire and smoke detection model based on YOLOv8n. The proposed model integrates the FasterNeXt module and constructs a Reparameterized Feature Pyramid Network (RepFPN) through the efficient integration of three core components: SimConv, Transpose, and RepBlock. Furthermore, the Wise-IoU (WIoU) loss function is employed to optimize the training process. These improvements significantly bolster detection precision and robustness in complex environments while maintaining a high inference speed, thereby providing potential for efficient forest fire monitoring. The specific structure of the proposed model is shown in Figure 2.

The proposed model replaces the original C2f modules in the YOLOv8 backbone with FasterNeXt lightweight units, which, integrated with convolutional layers and Spatial Pyramid Pooling Fast (SPPF), construct a progressive downsampling feature extraction network that significantly reduces parameter count and computational complexity while preserving rich gradient flow for early fire detection. By incorporating SimConv and Reparameterized Blocks (RepBlocks), alongside transpose convolutions for upsampling and feature concatenation, the model implements a bidirectional (top-down and bottom-up) feature fusion mechanism that leverages a multi-branch structure for enhanced feature expression during training and a single-branch mechanism for real-time inference. Furthermore, WIoU is employed as the bounding box regression loss, introducing a dynamic non-monotonic focusing mechanism that assesses anchor quality through outlier degrees and allocates gradient gains accordingly; this approach prioritizes average-quality samples and suppresses harmful gradients from low-quality anchors, effectively resolving localization bias caused by irregular flame morphologies and complex background interference. Ultimately, through these architectural and loss-function modifications, the proposed model reduces model complexity while improving multi-scale feature fusion and bounding-box regression, thereby providing a more favorable accuracy–complexity trade-off for forest fire monitoring.

FasterNeXt module

In the object detection of forest fires and smoke, early-stage smoke exhibits high diffusivity and significant visual ambiguity, while video-acquisition devices are typically deployed in wild environments or on Unmanned Aerial Vehicles (UAVs) characterized by constrained power and computational resources. Although the C2f module in the original YOLOv8 enhances feature flow via the Cross Stage Partial (CSP) structure, its internally stacked standard bottlenecks rely on numerous 3 × 3 convolutions. This leads to substantial channel computational redundancy when processing high-resolution surveillance images, thereby increasing inference latency.

To address these challenges, the FasterNeXt module is introduced as a lightweight and efficient feature extraction structure. Its design philosophy centers on residual grouped stacking and channel split-fusion, with a core architecture consisting of several series-connected FasterNet Blocks. Initially, input features undergo preliminary channel transformation and enhancement, followed by a “Split” operation that bifurcates the feature channels before they enter the core processing unit composed of n FasterNet Blocks. Within each FasterNet Block, a Partial Convolution (PConv) structure is employed: the input channels are divided into two parts, where convolution is performed only on a selected subset of channels, while the remaining channels are preserved through identity mapping. These channels are subsequently concatenated and processed through pointwise convolution (PWConv), Batch Normalization (BN), and activation functions. Finally, residual connections are introduced to fuse the processed features with the initial input. This approach significantly reduces computational overhead while retaining comprehensive feature information. After the final FasterNet Block, the features undergo global channel fusion, followed by integration and dimension alignment to produce the final feature map. By leveraging partial convolutions to minimize redundant calculations, residual connections to mitigate gradient vanishing, and grouped stacking to bolster feature extraction, this structure effectively balances model lightweighting with representational power, making it highly suitable for real-time computer vision tasks requiring high inference efficiency. The specific structure of the proposed model is shown in Figure 3.

To satisfy the dual requirements of real-time performance and lightweight architecture in forest fire monitoring, this study embeds the FasterNeXt module into the backbone of YOLOv8 as a replacement for the original C2f module. Specifically, the standard Bottleneck units within the C2f are reconfigured into highly efficient FasterNet Blocks. Let the input feature map of the module be X ∈ R C × H × W, where C, H, and W represent the number of channels, height, and width, respectively.

Considering that smoke in forest environments often exhibits diffuse distribution and blurred boundaries—characterized by small areas, low concentrations, and high similarity to the background during the early stages of a fire—direct deep convolution is prone to losing subtle edge information. Consequently, the module first performs feature preprocessing to integrate channels and enhance non-linear representations. By passing the input feature X through a 1 × 1 convolutional layer, the model effectively improves its response to low-contrast regions during the initial detection phase. The implementation process is described in Equation 1:

Flames and smoke often cover extensive areas, resulting in high semantic redundancy across feature map channels. To preserve original spatial details without increasing computational complexity, the preprocessed features are bifurcated into two branches via a Channel Split operation, yielding y 0 and y sub, as shown in Equation 2:

Given the irregular morphologies of flames and the variable shapes of smoke in forest environments, a single convolutional layer struggles to capture complex spatial distributions. Therefore, a sequence of non-linear transformations, M(·), consisting of stacked FasterNet Blocks, is employed to process y sub. Each block integrates Partial Convolution (PConv) and Pointwise Convolution (PWConv) to balance computational efficiency with representational power. The implementation is given in Equation 3:

After deep feature extraction, the features from both branches are re-fused to integrate fine-grained spatial details with high-level semantics. For global targets such as smoke, robust inter-channel interaction enhances the model’s overall perception of large-scale diffuse regions. By concatenating y 0 and y feat along the channel dimension, followed by a 1 × 1 convolution, cross-channel information interaction is achieved as shown in Equation 4: where BN(·) denotes Batch Normalization and σ (·) represents the SiLU activation function.

The efficiency gains of the FasterNeXt module are primarily derived from the PConv operator. In forest fire images, a vast number of pixels belong to smooth background regions where spatial information is highly repetitive across channels. While traditional convolutions assume that all channels carry independent spatial information, significant semantic redundancy exists. Based on this, a channel splitting strategy with n div = 2 is adopted in PConv to balance computational cost and feature extraction capability. PConv partitions the input along the channel axis into two symmetric subsets, X 1 and X 2, each containing C/2 channels. A standard 3 × 3 convolution is applied only to X 1 to capture spatial features, while X 2 remains unchanged, as described in Equations 5, 6:

Through this mechanism, the operator performs convolution on only half of the channels. The remaining channels pass through via identity mapping, significantly reducing the memory access cost (MAC) while retaining the primary textures of flames and smoke.

Since PConv only updates a subset of channels, the unconvolved channels lack spatial interaction. To facilitate full-channel information fusion and increase sensitivity to low-contrast smoke edges, a series of Point-wise Convolutions (PWConv) are integrated after PConv. The internal calculation sequence is defined in Equation 7:

To further execute dimension mapping and global interaction, the features pass through a second 1 × 1 PWConv layer, as shown in Equation 8:

The smoothness and non-monotonic properties of the SiLU activation function prevent complete gradient truncation when processing complex, low-contrast smoke features, thereby assisting the model in learning blurred boundaries. This cascaded PWConv structure enforces full-channel fusion, compensating for the limitations of PConv.

In forest fire detection, early-stage flames may occupy only a few pixels, and initial smoke is often extremely thin. As network depth increases, these subtle features are prone to dilution during gradient propagation. Consequently, each FasterNet Block introduces a residual connection, performing element-wise addition between the input and output. This effectively prevents the loss of weak features in deep layers, as expressed in Equation 9:

The FasterNeXt module (Chen et al., 2024) achieves a significant reduction in model complexity while maintaining an effective balance in feature extraction by integrating efficient FasterNet Blocks into the Cross Stage Partial (CSP) architecture. By utilizing PConv to eliminate channel-wise redundancy and employing a balanced splitting strategy, the module compresses the computational load during spatial feature extraction. Simultaneously, through a dual-layer PWConv structure and SiLU-based nonlinear transformations, the module ensures deep interaction across all channels. Ultimately, the residual architecture preserves low-level fine textures while substantially enhancing the discriminative precision for low-contrast smoke and flame features against complex forest backgrounds. In the implemented backbone, the FasterNeXt modules were configured across four stages, containing 1, 2, 2, and 1 FasterNet Blocks, respectively.

Reparameterized feature pyramid network

In the visual recognition of forest fires and smoke, targets exhibit extreme scale variations where early-stage ignitions may occupy only a few pixels, whereas mature wildfires cover extensive areas. Furthermore, smoke is characterized by its diffuse nature, blurred boundaries, and semi-transparency. Single-scale feature maps inherently struggle to simultaneously capture minute fire points and expansive smoke plumes. To address these challenges, the traditional feature fusion structure is refined into a Reparameterized Feature Pyramid Network (RepFPN). Architecturally, RepFPN establishes bidirectional fusion paths consisting of both top-down and bottom-up trajectories, achieving efficient cross-scale information interaction and semantic enhancement through the deep synergy of three core modules: SimConv, Transpose, and RepBlock.

During the scale transformation within the feature pyramid, RepFPN optimizes upsampling and downsampling efficiencies via the Transpose and SimConv modules, respectively. In the top-down path, to mitigate the loss of spatial details caused by continuous downsampling in deep layers, learnable Transpose convolutions are introduced to replace traditional linear interpolation. By learning feature recovery weights, this approach provides precise spatial response and structural support for both minute fire points and diffuse smoke. Conversely, in the bottom-up path, the streamlined SimConv unit is employed for channel transformation and downsampling. This effectively minimizes computational overhead and caters to the hardware constraints of edge monitoring devices while preserving the representation of local flame textures. As the primary functional component of the neck network, RepBlock is strategically deployed at fusion nodes across different scales. During the training phase, a multi-branch structure comprising 3 × 3 convolutions, 1 × 1 convolutions, and identity mappings is utilized to strengthen the extraction of diffuse smoke edges and irregular flame morphologies. During the inference phase, structural reparameterization is leveraged to equivalently fuse these branches into a single-path 3 × 3 convolution, ensuring real-time detection performance without incurring additional computational costs. Through the organic integration of these modules, RepFPN not only effectively bridges the semantic gap in multi-scale feature transmission but also significantly bolsters the model’s detection robustness against complex fire targets in forest environments under a lightweight deployment framework.

In forest fire and smoke detection tasks, the visual characteristics of flames and smoke exhibit significant complexity and stochasticity. Flames are characterized by irregular morphologies, substantial scale variations, and dynamic, ever-changing boundaries. Smoke, conversely, tends to be diffuse and semi-transparent, typically manifesting with low luminance and indistinct textures. In real-world scenarios, these features often blend with complex background interferences, posing a formidable challenge to the model’s feature extraction capabilities. Simultaneously, most edge devices deployed for field monitoring are constrained by limited computational resources, restricted memory footprints, and stringent power consumption requirements. This necessitates detection models that maintain robust feature extraction while achieving architectural lightweighting and high-speed inference. To effectively balance precision with the real-time demands of terminal deployment—while mitigating efficiency losses caused by redundant computations—this study introduces the SimConv module (Yue et al., 2023). As a lightweight basic convolutional unit sequentially composed of two-dimensional convolution, Batch Normalization (BN), and a ReLU activation function, SimConv facilitates an efficient equilibrium between representational capacity and inference speed. Its operation is defined in Equation 10:

where X represents the input feature map, Y denotes the output feature map, Conv2D indicates the standard 2D convolution operation, BN is the Batch Normalization layer, and ReLU refers to the Rectified Linear Unit activation function. The specific structure of the proposed model is shown in Figure 4.

Initially, given that the local edges and textural information of flames and smoke are typically hidden in confined spatial regions, standard convolutions are employed to effectively capture these local dependencies via a sliding window mechanism. Let the input feature map be X ∈ R C in × H × W, with Cout output channels, a kernel size of k × k, and a stride of s. The convolutional operation is defined in Equation 11:

The bias term b is deliberately omitted because the learnable shift parameter β in the subsequent Batch Normalization (BN) layer provides an equivalent functional substitution. Removing the bias term facilitates lightweight deployment on edge devices by reducing the parameter count without compromising the model’s representational capacity. Here, * denotes the convolution operation and W represents the kernel weights.

Since the spatial location information of forest fire and smoke targets is crucial to detection accuracy, it is necessary to keep the feature map size unchanged to preserve location details. The module uses an automatic filling method, as defined in Equation 12: where p is the padding size and k is the kernel size. This ensures that when the stride is 1, the dimensions of the feature map remain constant before and after convolution, effectively preventing the loss of minute targets caused by resolution reduction.

Subsequently, because the forest background contains interference factors such as changes in lighting, shadows, and swaying leaves, the feature distribution of the convolution output is unstable. Therefore, a batch normalization layer standardizes the convolution output to constrain the feature distribution. Its calculation process is given in Equation 13: where μ and σ are the batch mean and standard deviation, respectively, γ and β are learnable scaling and offset parameters, and ε is a small constant to prevent division by zero.

Finally, since the irregular shapes of flames, the semi-transparent edges of the smoke, and complex background interference exhibits nonlinear characteristics, linear convolution and normalization operations cannot adequately fit these complex relationships. Therefore, the module uses the ReLU activation function to introduce a non-linear transformation, enhancing the model’s ability to fit complex scenes. Its mathematical expression is given in Equation 14:

Compared with traditional complex convolutional structures with multiple branches and activations, SimConv offers advantages such as simple structure, low computational cost, and ease of deployment. It can reduce forward inference time while maintaining the receptive field and channel feature representation capabilities, thus improving real-time performance in forest fire monitoring scenarios. This module can be flexibly embedded into backbone networks, feature pyramids, and multi-scale fusion structures, providing stable and efficient foundational feature support for the accurate detection of smoke and flame targets.

The RepBlock (Wang et al., 2024b) serves as the pivotal functional element of the proposed neck feature fusion network, where its performance fundamentally governs the efficacy of feature integration and, by extension, the model’s accuracy and real-time performance in forest fire detection tasks. Built upon the RepVGGBlock unit, this component leverages structural reparameterization to dynamically adjust its topology. By decoupling the training architecture from the inference structure, the module achieves a superior balance between representational power and execution speed, effectively meeting the hardware constraints of edge-deployed forest monitoring systems. The operational paradigm of the RepVGGBlock is characterized by a transition from a multi-branch training configuration to a streamlined single-path inference mechanism. This dual-phase approach ensures that the structural design remains highly specialized for its respective functional goals. In the training stage, the module aims to maximize the extraction of deep semantic features through an augmented representational space. The specific structure of the proposed model is shown in Figure 5. Specifically, it employs three parallel processing branches to transform the input feature map X ∈ R C in × H × W into the output feature map Y ∈ R C out × H × W, a configuration detailed below:

The primary branch employs a 3 × 3 convolution to capture local spatial correlations within the feature map. Since the boundary features of flames and smoke are often subtle and difficult to delineate, this branch is specifically designed to extract fine-grained local details, textures, and edge information. The operation is expressed in Equation 15: where Conv3 × 3(X) denotes the standard 3 × 3 convolution applied to the input X for spatial feature extraction, and BN3(·) signifies the subsequent Batch Normalization (BN) layer used to standardize feature distributions and accelerate network convergence.

To complement spatial extraction, a 1 × 1 convolutional branch is integrated to facilitate dimensionality transformation and channel-wise compression. Given that forest fire scenarios are typically characterized by complex background interference, this branch effectively suppresses redundant noise and concentrates the model’s focus on salient flame and smoke features. By re-weighting channel information, it reduces computational overhead while preserving critical feature representations. Its calculation is defined in Equation 16: where Conv1 × 1(X) achieves dimensionality adjustment through a 1 × 1 kernel, and BN1(·) ensures that the feature distribution of this branch remains consistent with the other parallel paths.

The final component is the identity mapping branch, which serves to preserve the most primitive image information. This is particularly crucial for detecting early-stage fires, where minute flames and wispy smoke may otherwise be lost or diluted during deep feature processing. By directly passing the input features, this path ensures the retention of subtle signals. Notably, the identity branch is conditionally active; it operates only when the input and output channel counts are identical (C in = C out) and the stride is 1. If these conditions are not met, the branch is automatically deactivated, leaving the processing to the two convolutional branches. The logic is represented in Equation 17:

During the training phase, the irregular morphologies and small scales of forest flames and smoke, coupled with complex environmental distractors, necessitate the collaborative processing of the three parallel branches. The outputs from these branches are integrated via element-wise summation to fuse features extracted across different dimensions. This fused representation then undergoes a non-linear transformation through the SiLU activation function, significantly bolstering the model’s capacity to extract features from subtle fire and smoke signals. Within this architecture, each Batch Normalization (BN) operation incorporates four key parameters: the mean (μ), variance (σ 2), scaling factor (γ), and shift (β). These parameters are continuously optimized through backpropagation and running statistics during training, ensuring that the BN operation precisely adapts to the feature distribution and enhances the network’s representational power. The final output feature map of the RepVGGBlock during training is defined in Equation 18:

To minimize computational overhead and satisfy the deployment requirements of resource-constrained edge devices, the multi-branch topology is equivalently transformed into a single-path 3 × 3 convolutional structure during the inference phase. This structural reparameterization involves the equivalent fusion of convolutional kernels and BN parameters from each branch. Specifically, let the kernel weights of a given branch is W ∈ R K × K × C in × C out; the fused equivalent weight W′ and equivalent bias b’ are calculated using Equations 19, 20: where ε is a negligible constant introduced to prevent division by zero. To ensure dimensional consistency during fusion, the 1 × 1 convolutional kernels are zero-padded to a 3 × 3 size. Similarly, the identity mapping branch is treated as a specialized 1 × 1 convolution (an identity matrix) and subsequently padded to 3 × 3. Following this alignment, the equivalent weights and biases from all three branches are summed element-wise to produce the parameters for a single 3 × 3 convolutional layer, as shown in Equations 21, 22:

Consequently, during inference, the RepVGGBlock performs only a single 3 × 3 convolution to complete feature processing. This mechanism substantially reduces both the computational workload and Memory Access Cost (MAC). Furthermore, this ensures that the inference precision remains identical to that of the training phase, achieving high-speed execution without any loss in detection accuracy. The simplified inference operation is expressed in Equation 23:

As the architectural cornerstone of the neck feature fusion network, the RepBlock is configured not as a solitary unit, but as a sequential stack of n RepVGGBlocks, forming a deep feature processing structure. This deep stacking, integrated with the structural reparameterization mechanism, facilitates a rich gradient flow during the training phase to enhance feature representation. This characteristic is particularly advantageous for learning the diffuse textures of smoke and the irregular boundaries of flames. During inference, the structure maintains the high efficiency of a single-path convolutional chain, effectively reducing the computational overhead of the inference graph. Given an input Z 0, the final output Z out is expressed in Equation 24: where f i denotes the mapping function of the i-th RepVGGBlock.

Within the overall architecture, four RepBlocks are deployed at the multi-scale fusion nodes of RepFPN, with each RepBlock comprising a sequential stack of 12 RepVGGBlocks. These RepBlocks work in synergy with the upsampling and downsampling operations to perform cross-scale feature fusion and nonlinear feature refinement. Each pyramid level corresponds to a specific scale: lower-level feature maps encapsulate rich spatial details, such as flame edges and minute smoke particles, while higher-level maps provide robust semantic information regarding integrated fire and smoke regions. This configuration facilitates the effective integration of multi-scale information, thereby enhancing the semantic representation of tiny smoke plumes and irregular flames and providing high-quality feature inputs for subsequent detection tasks.

In forest fire and smoke detection, early-stage fire targets are often minuscule, while smoke exhibits high diffusivity and sparsity. During deep feature extraction, continuous downsampling operations inevitably lead to a severe loss of spatial information, which weakens the model’s localization precision for faint ignition points and long-distance smoke. To mitigate this loss of spatial detail, a Transpose upsampling module is integrated into the RepFPN architecture. Centered on the two-dimensional transposed convolution (ConvTranspose2d), this module functions as an upsampling operator with learnable parameters. It remaps highly compressed deep semantic features back into the high-resolution spatial domain, providing the necessary geometric structure and positional cues for precise target localization and classification. The specific structure is shown in Figure 6.

Because it’s necessary to restore deep, low-resolution feature maps to high resolution to reconstruct spatial details such as flame edges and smoke contours, the Transpose module achieves this by padding the elements of the input feature map with zeros and applying learnable convolutional kernels. Therefore, the Transpose module uses two-dimensional transposed convolution as its core operator. Let the input feature map is X ∈ R C in × H in × W in and the output feature map after the transposed convolution operation is Y ∈ R C out × H out × W out, where C represents the number of channels, and H and W represent the height and width of the feature map. The mapping relationship performed by this module is expressed in Equation 25: where W represents the learnable convolutional kernel and b denotes the bias term.

In forest fire scenarios, the initial fire source targets are small, the smoke is thin, and the spread is large. Therefore, the feature map size needs to be precisely doubled to retain the location information of these small targets while avoiding the introduction of unnecessary padding noise. Consequently, the module is configured with a kernel size (k) of 2, a stride (s) of 2, and zero padding (p = 0). No output padding is utilized, and the bias term was retained in the implemented transposed-convolution layer. The spatial dimensions of the output feature map are calculated using Equations 26, 27:

Under this configuration, the output dimensions satisfy H out = 2H in and W out = 2W in. This effectively doubles the spatial dimensions while simultaneously transforming the channel count from C in to C out.

Compared to traditional nearest-neighbor or bilinear interpolation, the Transpose module (Hai et al., 2025) demonstrates a superior ability to preserve critical information such as flame boundaries and smoke silhouettes. Since transition regions between fire centers and forest backgrounds, as well as the edges of thin smoke, contain complex textural features, the module’s internal learnable weight matrix can adaptively capture target edges and contours based on the distribution characteristics of forest fires. This mechanism effectively bridges the semantic gap during feature fusion, enabling the model to maintain high spatial response accuracy and boundary definition when processing distant, small-scale fire points.

WIoU loss function

In visual recognition tasks for forest fires and smoke, target boundaries typically exhibit characteristics such as fuzziness, high diffusivity, and significant morphological variability. These traits make it difficult for bounding boxes to align precisely with the actual object contours, leading to the inevitable inclusion of low-quality samples within training datasets. Consequently, traditional IoU loss functions struggle to balance convergence speed with localization precision. Although loss functions like CIoU improve regression performance by introducing penalty terms for center distance and aspect ratio, their fixed weight allocation mechanisms fail to account for disparities in anchor- box quality, rendering the model susceptible to interference from low-quality anchors in complex environments. To address these issues, this study introduces Wise-IoU (WIoU) (Tong et al., 2023), an advanced bounding box regression loss function based on a dynamic non-monotonic focusing mechanism. By employing outlier degree evaluation and non-monotonic focusing, WIoU actively suppresses harmful gradients generated by annotation noise and low-quality samples. Simultaneously, the distance attention coefficient, R WIoU enhances sensitivity to center-point offsets, significantly improving localization for semi-transparent targets like smoke. By concentrating the model’s learning focus on effective samples of moderate quality, this approach substantially bolsters robustness and generalization in complex scenarios.

The Wise-IoU (WIoU) loss function achieves optimization by incorporating a non-monotonic focusing coefficient. The standard IoU loss, ℒ IoU, is defined in Equation 28: where A and B represent the predicted and ground truth bounding boxes, respectively, and |·| denotes the area. In forest fire scenarios, significant center offsets often exist between the flame core and the smoke diffusion area. Since IoU alone is insufficient to effectively penalize center-point deviations, a distance attention weight, R WIoU, is introduced to enhance sensitivity to these offsets. R WIoU measures the correlation of center distances between the predicted and ground truth boxes, as defined in Equation 29: where (x, y) and (x gt, y gt) are the center coordinates of the predicted and ground truth boxes, while W g and H g represent the width and height of the minimum enclosing rectangle. The superscript * signifies a pivotal design feature of WIoU: the gradient separation operation. Due to the semi-transparent boundaries of smoke and the irregular morphologies of flames, manual annotations are inherently prone to noise. Including the parameters of the enclosing rectangle in the gradient update would amplify this interference; therefore, a separation operation is employed to stabilize the training process. Specifically, during backpropagation, the term marked with * is treated as a constant, effectively detaching it from the gradient calculation.

Furthermore, training datasets often contain numerous low-quality anchors, such as fire points occluded by dense smoke. These samples generate harmful gradients that can degrade performance. To address this, WIoU introduces the outlier degree β to assess anchor quality, as defined in Equation 30:

In this expression, ℒ IoU ∗ is the current predicted box’s IoU loss, and ℒ IoU ¯ is the exponential moving average loss during training. Based on this outlier degree, WIoU constructs a non-monotonic focusing coefficient r and obtains the final WIoU loss, as shown in Equations 31, 32: where α and δ are hyperparameters that control the gain curve. In this study, WIoU v3 was adopted, with α = 1.9 and δ = 3.0.

In forest fire and smoke detection, a small β indicates a high-quality anchor with precise localization; in such cases, a smaller gain coefficient is assigned to prevent overfitting. Conversely, a large β identifies the anchor as an outlier (e.g., a fire point with extremely blurred boundaries due to smoke occlusion). Through its non-monotonic mechanism, WIoU actively suppresses the gradient contribution of these outliers, preventing low-quality anchors from causing significant interference during parameter updates. This strategy enables the model to concentrate on moderate-quality anchors, substantially improving both the localization of flame core regions and the training efficiency in complex outdoor environments.

Experimental environment and parameter settings

This experiment used the PyTorch framework, version 2.0.1, and the Python programming language, version 3.8. All models were trained on a Windows 10 operating system. The specific hardware configuration was as follows: Intel Xeon Platinum 8,163 processor, 2.50 GHz, 24 cores; NVIDIA GeForce RTX 4090D GPU (47 GB VRAM); CUDA 11.7. The baseline model used in this study was YOLOv8n. Both YOLOv8n and FRW-YOLOv8 were trained from scratch without pretrained weights. The AdamW optimizer was adopted with an initial learning rate of 0.002 and a weight decay of 0.0005. A cosine learning-rate schedule was used, with the final learning-rate factor set to 0.01, corresponding to a final learning rate of 2 × 10 −5. The data augmentation settings were kept identical across all models. HSV augmentation was applied with hue, saturation, and value gains of 0.02, 0.4, and 0.4, respectively. Mosaic, MixUp, and copy-paste augmentations were applied with probabilities of 0.4, 0.3, and 0.3, respectively. The detailed experimental settings are summarized in Table 1.

All baseline, ablation, comparative, and external-validation experiments reported in this study were initially conducted using a fixed random seed of 3 to ensure a consistent controlled experimental protocol. To assess the sensitivity of the principal comparison to stochastic training variation, YOLOv8n and FRW-YOLOv8 were additionally trained using random seeds 0 and 1. Consequently, the principal comparison between the two models was conducted using three random seeds, namely 0, 1, and 3. For all repeated runs, only the random seed was changed, while the fixed dataset partition and all other training and evaluation settings remained unchanged.

For all internal and external evaluations, candidate predictions were retained using a minimum confidence threshold of 0.001, followed by class-aware non-maximum suppression with an IoU threshold of 0.70. These thresholds were applied uniformly to all evaluated models on the M4SFWD test set and to both YOLOv8n and FRW-YOLOv8 on the flame and Wildfire Smoke datasets, without model-specific or dataset-specific adjustment.

Model evaluation

This study evaluates the trade-off between detection accuracy and model efficiency. Therefore, mAP50 and mAP50-95 were selected as the principal accuracy metrics, while parameters, GFLOPs, and FPS were used to evaluate model complexity, computational cost, and inference throughput, respectively. For the repeated random-seed experiments, mAP50, mAP50-95, and FPS are reported as the mean ± sample standard deviation over three runs using random seeds 0, 1, and 3. Mean Average Precision (mAP)

mAP is a core metric in object detection tasks, used to comprehensively evaluate a model’s detection accuracy across different categories. Specifically, mAP50 denotes the mean average precision at an Intersection over Union (IoU) threshold of 0.5. It considers a prediction as a correct detection only if the overlap between the predicted bounding box and the ground truth bounding box is at least 50%. Consequently, mAP50 is relatively insensitive to the precise localization of bounding boxes, focusing more on the overall classification and rough positioning. The specific calculation formula for mAP50 is given in Equation 33:

Compared with the mean average precision at an Intersection-over-Union (IoU) threshold of 0.5 (mAP50), mAP50-95 serves as a core metric in object detection for comprehensively assessing both classification capability and bounding box localization precision. This metric encompasses a spectrum of IoU thresholds ranging from loose to stringent. Specifically, it represents the mean of mAP values calculated across Q discrete IoU thresholds (amounting to 10) from 0.50 to 0.95 with an incremental step of 0.05. Consequently, mAP50-95 provides a holistic reflection of model performance under varying spatial localization constraints. A higher score implies superior proficiency in both target identification and precise bounding box alignment, thereby ensuring a more rigorous validation of detection accuracy and adaptability across diverse scenarios. The mathematical formulation of mAP50-95 is defined in Equation 34: Parameter count

The parameter count refers to the total number of trainable parameters in a model. It serves as a core indicator for measuring model complexity and lightweight degree. A smaller parameter count generally reduces model-storage and memory requirements, although inference speed also depends on operator type, memory access, parallelism, and hardware implementation. Conversely, a larger parameter count enhances the model’s fitting capacity and its ability to handle complex scenarios. However, increasing the number of parameters inevitably leads to higher computational demands and stricter hardware requirements. Floating Point Operations (FLOPs)

FLOPs are a critical quantitative metric for measuring a model’s computational complexity. They specifically refer to the total number of floating-point operations executed during a single forward inference process, directly reflecting the model’s demand for hardware arithmetic power. Since flame and smoke detection must often be deployed on computationally constrained edge devices, such as Unmanned Aerial Vehicles (UAVs), an excessively high FLOP count would lead to poor practical applicability, even if the model is designed for high accuracy. In this experiment, this metric is expressed in GFLOPs (Giga Floating-point Operations), representing billions of floating-point operations. The calculation of GFLOPs is given in Equation 35: Real-time performance

Real-time performance is measured in Frames Per Second (FPS), which represents the number of images processed by a model per second under the specified inference configuration. A higher FPS indicates a higher image-processing throughput under the same experimental conditions. The inference-speed experiments were conducted on the NVIDIA GeForce RTX 4090D GPU using the native PyTorch inference pipeline. All evaluated models used the same hardware environment, software implementation, and timing procedure. The input resolution was uniformly fixed at 640 × 640 pixels, the inference batch size was set to 32, and FP32 precision was used without mixed-precision inference, quantization, TensorRT acceleration, or model-specific hardware optimization. Before formal timing, all models underwent the same warm-up procedure to reduce the influence of CUDA initialization, memory allocation, and kernel scheduling. The timed interval included image preprocessing, model inference, and post-processing. The same number of test images and repeated runs were used for all models, and the FPS values were calculated using the same timing and aggregation procedure. Therefore, the reported FPS values represent comparative batch-inference throughput under the specified desktop-GPU environment rather than batch-one latency or embedded-device deployment performance. The FPS is calculated using Equation 36:

Ablation experiment

To determine whether the introduced components provide measurable individual and combined contributions, ablation experiments were conducted on the M4SFWD dataset. The modules were evaluated not only in terms of Fire AP50, Smoke AP50, mAP50, and mAP50-95, but also in terms of parameter count, GFLOPs, and FPS. Therefore, the necessity of each component was assessed according to its intended function and its contribution to the overall accuracy–complexity trade-off, rather than whether it improved every individual metric. In Table 2, A, B, and C denote FasterNeXt, RepFPN, and WIoU, respectively.

Experiment 1 used the original YOLOv8n model as the baseline. When FasterNeXt was independently introduced in Experiment 2, Fire AP50, Smoke AP50, mAP50, and mAP50-95 increased by 0.1, 1.5, 0.8, and 0.5 percentage points, respectively. Meanwhile, the parameter count decreased from 3.01 M to 2.56 M and GFLOPs decreased from 8.1 to 6.8, corresponding to reductions of 14.93 and 16.05%, respectively. These results indicate that FasterNeXt primarily contributes to lightweight backbone feature extraction by reducing redundant computation while preserving or slightly improving the representation of fire and smoke targets.

Experiment 3 evaluated RepFPN independently. Compared with the baseline, Smoke AP50, mAP50, and mAP50-95 increased by 1.3, 0.5, and 0.7 percentage points, respectively, while the parameter count decreased from 3.01 M to 2.83 M and GFLOPs decreased from 8.1 to 7.5. Fire AP50 decreased slightly by 0.2 percentage points. Therefore, RepFPN does not uniformly improve every individual metric, but its improvements in Smoke AP50 and mAP50-95 are consistent with its intended role in enhancing multi-scale feature interaction and the representation of diffuse smoke regions.

Experiment 4 introduced WIoU as the bounding-box regression loss. Without increasing the parameter count or GFLOPs, WIoU increased Fire AP50, Smoke AP50, mAP50, and mAP50-95 by 1.5, 0.5, 1.0, and 1.6 percentage points, respectively. The relatively larger improvement in mAP50-95 indicates that WIoU mainly contributes to improving bounding-box localization under stricter IoU thresholds, which is relevant to fire and smoke targets with irregular shapes and blurred boundaries.

Experiment 5 combined FasterNeXt and RepFPN. Compared with the FasterNeXt-only configuration, Fire AP50 and mAP50 increased by 0.7 and 0.2 percentage points, respectively, while the parameter count further decreased from 2.56 M to 2.38 M and GFLOPs decreased from 6.8 to 6.2. Although Smoke AP50 decreased slightly from 85.3 to 85.0% and mAP50-95 remained unchanged at 49.4%, the combined configuration achieved a further reduction in model complexity while maintaining improved detection performance relative to the original baseline. This result indicates that FasterNeXt and RepFPN provide complementary functions in lightweight feature extraction and cross-scale feature fusion, although their effects are not strictly additive across all metrics.

Finally, WIoU was added to the FasterNeXt–RepFPN architecture in Experiment 6 to construct the complete FRW-YOLOv8 model. Compared with Experiment 5, Fire AP50, Smoke AP50, mAP50, and mAP50-95 increased by 0.2, 0.9, 0.6, and 1.2 percentage points, respectively, without increasing the parameter count or GFLOPs. This result indicates that WIoU further improves overall detection performance and bounding-box localization on the lightweight structural configuration, particularly under the stricter mAP50-95 metric.

Compared with the original YOLOv8n baseline, the complete model improved Fire AP50, Smoke AP50, mAP50, and mAP50-95 by 1.0, 2.1, 1.6, and 1.7 percentage points, respectively. At the same time, the parameter count decreased from 3.01 M to 2.38 M and GFLOPs decreased from 8.1 to 6.2, corresponding to reductions of 20.67 and 23.46%. The complete model achieved the highest Smoke AP50, mAP50, and mAP50-95 among the ablation configurations while maintaining the lowest parameter count and GFLOPs.

It should be noted that the complete model did not achieve the best result for every individual metric. For example, the WIoU-only configuration achieved a Fire AP50 of 91.7%, which was slightly higher than the 91.2% obtained by the complete model. In addition, the FPS decreased from 277.8 for the baseline to 256.4 for the complete model. These results show that the effects of the modules are not strictly monotonic or additive across all metrics. The three components were retained because they provide distinct and complementary contributions to lightweight feature extraction, multi-scale feature fusion, and bounding-box regression, thereby producing a more favorable overall accuracy–complexity trade-off rather than maximizing every individual metric.

Comparison with other models

To objectively and comprehensively evaluate the detection performance of the proposed model, representative object detection frameworks were selected for comparative experiments. Comparative experiments were conducted using four representative general-purpose detectors, including YOLOv8, YOLOv9 (Wang C.-Y. et al., 2024), YOLOv10 (Wang A. et al., 2024), and YOLOv11 (He et al., 2025), together with two recent task-specific models, FireYOLO-Lite and a-FSDM. The exact model variants used in the comparative experiments are specified in Table 3. FireYOLO-Lite is a lightweight forest fire detection model that integrates efficient feature extraction, multi-scale attention, and small-target bounding-box regression optimization. a-FSDM is a dedicated fire-and-smoke detection model that introduces an Attentive Transparency Detection Head to enhance the representation of transparent flame and smoke targets. These models cover both general-purpose object detection architectures and task-specific fire and smoke detection strategies, providing a broader basis for evaluating the detection accuracy, computational complexity, and inference throughput of FRW-YOLOv8. All comparative models were trained and evaluated using the same training, validation, and test subsets of the M4SFWD dataset. The input image size was uniformly set to 640 × 640 pixels, and all experiments were conducted using the same hardware environment and evaluation implementation. Fire AP50, Smoke AP50, mAP50, mAP50-95, parameter count, GFLOPs, and FPS were used to evaluate detection accuracy and model efficiency.

As shown in Table 3, FRW-YOLOv8 achieved the highest Smoke AP50, mAP50, and mAP50-95 among all evaluated models. Meanwhile, it required the fewest parameters and the lowest computational cost, demonstrating that the reduction in model complexity was achieved without sacrificing detection accuracy. Compared with the task-specific FireYOLO-Lite and a-FSDM models, FRW-YOLOv8 delivered better overall fire and smoke detection performance with a more lightweight architecture. Although its inference throughput was lower than that of YOLOv10n, FRW-YOLOv8 maintained competitive inference throughput. Overall, FRW-YOLOv8 achieved a favorable trade-off among detection accuracy, model complexity, computational cost, and inference speed.

Table 3 reports the results of all comparative models obtained under the common seed-3 experimental protocol. Although this controlled setting ensures a fair comparison among the evaluated models, a single training run cannot fully reflect the influence of stochastic initialization. Therefore, to further assess the stability of the principal comparison, YOLOv8n and FRW-YOLOv8 were independently trained using the same three random seeds, namely 0, 1, and 3, while all other experimental settings remained unchanged.

Across the three runs, YOLOv8n achieved an mAP50 of 87.01 ± 0.20%, an mAP50-95 of 48.87 ± 0.15%, and an inference speed of 277.50 ± 3.19 FPS, whereas FRW-YOLOv8 achieved 88.60 ± 0.20%, 50.60 ± 0.20%, and 256.37 ± 2.65 FPS, respectively. FRW-YOLOv8 maintained higher mAP50 and mAP50-95 values across all three random seeds, with mean improvements of 1.59 and 1.73 percentage points, respectively. The limited run-to-run variation indicates that the observed accuracy improvements remained stable across the three tested random seeds rather than being driven by a single favorable initialization. Although the inference speed of FRW-YOLOv8 was lower than that of YOLOv8n, it maintained a mean throughput above 250 FPS while requiring fewer parameters and lower computational cost.

Performance across different target scales

Small fire and smoke targets usually contain limited visual information and are more susceptible to background interference, making them difficult to detect in complex forest monitoring scenes. To investigate the scale sensitivity of the proposed model, the annotated targets in the M4SFWD test set were divided into small (S), medium (M), and large (L) groups according to their bounding-box areas.

All models were evaluated at an input resolution of 640 × 640 pixels. Following the COCO-style area criteria in the evaluation space, targets with a bounding-box area of A < 32 2 pixels 2 were classified as small, those with 32 2 ≤ A < 96 2 pixels 2 as medium, and those with A ≥ 96 2 pixels 2 as large, where A denotes the bounding-box area. The class-averaged mAP50 and mAP50-95 values were calculated separately for each target-scale group.

Table 4 compares the scale-specific detection performance of YOLOv8n and FRW-YOLOv8 on the M4SFWD test set. The reported mAP values were averaged over the fire and smoke categories.

As shown in Table 4, FRW-YOLOv8 achieved mAP50 values of 75.7, 88.4, and 92.8% for small, medium, and large targets, respectively, representing improvements of 0.9, 1.5, and 1.7 percentage points over YOLOv8n. Under the stricter IoU thresholds, FRW-YOLOv8 achieved mAP50-95 values of 34.4, 51.1, and 58.8%, improving the baseline by 1.2, 0.6, and 1.6 percentage points, respectively. The proposed model consistently outperformed YOLOv8n across all three target-scale groups, indicating improved multi-scale feature representation and localization performance. Although positive improvements were obtained for small targets, their absolute detection performance remained lower than that of medium and large targets, suggesting that small fire and smoke detection remains challenging.

External validation on independent real fire and smoke datasets

Although the M4SFWD dataset is a representative public UAV-perspective benchmark dataset for forest fire detection and provides diverse synthetic scenes with different terrains, weather conditions, illumination levels, and multi-scale fire and smoke targets, certain differences may still exist between simulation-based data and real fire monitoring scenarios in terms of imaging noise, target morphology, smoke diffusion, background interference, occlusion, and motion blur. Therefore, on the basis of model training and internal evaluation using the M4SFWD dataset, introducing real-scene data for external validation can provide a complementary assessment of the model’s adaptability and robustness in practical application environments, thereby improving the completeness of the experimental evaluation and the credibility of the conclusions. For this purpose, two independent external datasets, namely FLAME dataset and the Wildfire Smoke dataset, were selected to directly test the YOLOv8n baseline model and the proposed FRW-YOLOv8 model trained on the M4SFWD dataset.

For the FLAME dataset, only the fire class was evaluated. Table 5 reports Fire AP50, Fire AP50-95, parameters, and GFLOPs. For the Wildfire Smoke dataset, only the smoke class was evaluated. Table 6 reports Smoke AP50, Smoke AP50-95, parameters, and GFLOPs.

As shown in Table 5, FRW-YOLOv8 achieved a Fire AP50 of 88.3% and a Fire AP50-95 of 56.8% on the real fire dataset, which were 1.2 and 1.3 percentage points higher than those of YOLOv8n, respectively. As shown in Table 6, FRW-YOLOv8 achieved a Smoke AP50 of 93.5% and a Smoke AP50-95 of 57.9% on the real smoke dataset, improving upon YOLOv8n by 1.7 and 0.2 percentage points, respectively. Meanwhile, the number of parameters decreased from 3.01 M to 2.38 M, and the computational cost decreased from 8.1 GFLOPs to 6.2 GFLOPs. These results indicate that FRW-YOLOv8 maintains improved category-specific detection performance on independently collected real fire and smoke images while reducing model complexity.

To further examine whether the external-validation results were sensitive to stochastic training variation, the checkpoints obtained from the three seed-specific training runs were directly evaluated on the two external datasets without retraining or fine-tuning. On the FLAME dataset, YOLOv8n achieved mean Fire AP50 and Fire AP50–95 values of 87.07 ± 0.15% and 55.47 ± 0.15%, whereas FRW-YOLOv8 achieved 88.30 ± 0.20% and 56.80 ± 0.20%, respectively. On the Wildfire Smoke dataset, the corresponding Smoke AP50 and Smoke AP50–95 values were 91.80 ± 0.20% and 57.69 ± 0.15% for YOLOv8n, compared with 93.50 ± 0.20% and 57.91 ± 0.15% for FRW-YOLOv8. These results indicate that the external performance trend was generally maintained across the three training runs. The mean difference in Smoke AP50–95 was only 0.22 percentage points. Although the difference remained positive across the three seed-specific evaluations, its magnitude was modest. Therefore, it is interpreted as a small consistently observed numerical difference rather than as statistically conclusive evidence of superiority.

Visualization and analysis of results

Figure 7 presents a qualitative comparison between YOLOv8n and FRW-YOLOv8 using the same six representative images from the M4SFWD test set. The selected examples cover a complex multi-target scene, small or distant fire and smoke targets in snowy terrain, nighttime conditions, low-contrast diffuse smoke, dense or overlapping targets, and a foggy or visually confusing background. For direct comparison, the detection results of YOLOv8n and FRW-YOLOv8 are presented using the same image order and a unified visualization format.

In relatively clear scenes, both models successfully detect the main fire and smoke regions. Compared with YOLOv8n, FRW-YOLOv8 detects additional weak or small targets in several challenging scenes and provides more complete localization for some diffuse or overlapping fire–smoke regions. The improvement is particularly visible under low illumination, complex terrain, and weak target–background contrast. However, extremely small or weak targets may still be missed, while fog, clouds, bright sky, and other smoke-like background patterns may cause false alarms or prediction uncertainty. Therefore, the visualization is interpreted as qualitative supporting evidence, whereas the overall performance comparison is primarily determined by the quantitative AP results.

The results of this study indicate that FRW-YOLOv8 achieves an improved accuracy–complexity trade-off for fire and smoke detection on the M4SFWD dataset. Although the accuracy gains over the baseline YOLOv8n are moderate, they are accompanied by reductions in model size and computational complexity. The ablation results further demonstrate that FasterNeXt, RepFPN, and WIoU contribute to the model in distinct but complementary ways. FasterNeXt reduces redundant computation in the backbone, RepFPN strengthens the interaction between shallow spatial details and deep semantic features, and WIoU improves bounding-box regression for small and irregular fire-related targets. The combination of these components therefore aims not to maximize every individual metric, but to provide a more balanced solution for forest fire monitoring, where detection accuracy, localization reliability, and computational efficiency must be considered simultaneously.

From the perspective of forest fire monitoring and forest ecosystem protection, detecting early fire-related visual cues is important because small flames and incipient smoke may provide useful warning information before fire spread becomes severe. In forest environments, delayed detection may increase burned areas, vegetation loss, habitat degradation, carbon release, and fire suppression difficulty. Therefore, an efficient vision-based detection model may provide potential technical support for early warning systems, especially in monitoring scenarios where large forest areas need to be observed under limited computational resources. In this context, FRW-YOLOv8 is not intended to replace existing fire monitoring systems, but rather to serve as a lightweight visual detection component that may be integrated with UAV monitoring, fixed cameras, or other forest surveillance platforms in future applications.

To further examine the synthetic-to-real transferability of the proposed model, two independent real-world external datasets were introduced. The FLAME dataset was used for fire-specific external validation, and the Wildfire Smoke dataset was used for smoke-specific external validation. The trained YOLOv8n baseline and FRW-YOLOv8 models were directly evaluated on these two external datasets. According to the annotation characteristics of the two datasets, fire-specific and smoke-specific evaluations were conducted separately. The results show that FRW-YOLOv8 achieved improved detection performance over the baseline model on both real fire and real smoke images, indicating its potential transferability to real-world forest fire monitoring.

The three datasets used in this study collectively cover multiple target scales, illumination levels, weather conditions, real-world imaging characteristics, diffuse smoke appearances, and complex backgrounds. Therefore, the reported internal and external AP results represent overall performance across diverse mixed conditions rather than performance under a single idealized fire scenario. However, the datasets do not provide unified condition labels for independently quantifying faint smoke, occlusion, fog, or different levels of background complexity. The current results should therefore be interpreted as an overall mixed-condition evaluation rather than a condition-by-condition quantitative benchmark.

In addition, a dedicated negative-only benchmark composed exclusively of fire-free images containing clouds, sunlight, dust, steam, or fog was not included in the current experimental design. Although visually confusing backgrounds occur in the evaluated datasets, the aggregate detection results cannot replace a dedicated false-alarm assessment. Future work should construct condition-labeled challenge subsets and independent fire-free image or continuous-video benchmarks to evaluate false-alarm image rates, false positives per image, and false-alarm frequency over time.

It should also be noted that the internal evaluation on M4SFWD was conducted using a fixed image-level partition rather than a scene- or sequence-exclusive grouping strategy. Although the same partition was consistently applied to all baseline, ablation, and comparative experiments to ensure controlled and reproducible comparisons, some degree of visual correlation among samples originating from related simulated environments cannot be completely ruled out. Therefore, the internal results are primarily interpreted as comparative evidence obtained under an identical evaluation protocol, rather than as a dedicated assessment of generalization to entirely unseen simulated scenarios. The external evaluations on the flame and Wildfire Smoke datasets provide complementary evidence of transferability to real-world fire and smoke images. Future work will further investigate scene- or sequence-exclusive evaluation when a larger and more evenly distributed collection of simulated scenarios becomes available.

Lightweighting is not universally necessary for all forest fire detection systems. When sufficient server-side computing resources and stable communication conditions are available, larger detection models may also be applicable. However, lightweight design is particularly relevant to the UAV-based and edge-based monitoring scenarios considered in this study because these platforms are commonly constrained by payload, onboard computation, memory capacity, battery life, and communication bandwidth. Therefore, reducing model size and computational workload can improve the potential feasibility of local fire and smoke detection and reduce dependence on remote computing resources.

Nevertheless, reduced parameters and FLOPs do not automatically guarantee efficient embedded deployment. At the model level, excessive structural simplification may potentially weaken the representation of extremely small flames, diffuse smoke, low-contrast boundaries, and visually confusing backgrounds. At the hardware level, actual deployment performance may also be affected by memory access, operator compatibility, inference frameworks, hardware acceleration, batch size, power consumption, and continuous video-stream processing. In this study, hardware-level deployment experiments on UAV-compatible edge platforms were not conducted. Therefore, the present results demonstrate computational efficiency at the model-complexity level, whereas practical real-time deployment still requires validation on embedded devices under real flight or field-monitoring conditions.

Building on these findings, future work should move from controlled dataset evaluation toward field-oriented forest fire monitoring. First, larger real-world fire–smoke datasets should be collected from UAV and ground-based monitoring platforms under consistent annotation standards, covering different forest types, fuel conditions, topographic settings, seasons, weather patterns, and illumination environments. Such datasets would make it possible to evaluate the model more comprehensively under realistic forest disturbance scenarios. Second, future studies should extend the current image-based evaluation to continuous video monitoring, where temporal information, changing viewpoints, smoke diffusion, and fire spread dynamics can be incorporated to improve detection stability and reduce false alarms. Third, the deployment feasibility of FRW-YOLOv8 should be assessed on UAV-compatible edge devices by measuring inference latency, memory usage, power consumption, and performance under real flight or field-monitoring conditions.

This study proposed FRW-YOLOv8, a lightweight YOLOv8-based model for fire and smoke detection in forest fire monitoring. The model integrates FasterNeXt into the backbone to reduce redundant computation, introduces RepFPN to enhance multi-scale feature fusion, and adopts WIoU to improve bounding-box regression for small and irregular fire-related targets. On the M4SFWD dataset, FRW-YOLOv8 achieved a Fire AP50 of 91.2%, a Smoke AP50 of 85.9%, an overall mAP50 of 88.6%, and an mAP50-95 of 50.6%. Compared with the baseline YOLOv8n, these values increased by 1.0, 2.1, 1.6, and 1.7 percentage points, respectively. At the same time, the number of parameters was reduced from 3.01 M to 2.38 M, and the computational cost decreased from 8.1 GFLOPs to 6.2 GFLOPs. These results indicate that FRW-YOLOv8 achieves an improved accuracy–complexity trade-off on the M4SFWD dataset.

External validation on two independent real-world datasets further indicated that the proposed model demonstrates transferability to real fire and smoke images. Overall, FRW-YOLOv8 provides a lightweight visual detection approach for identifying early fire and smoke cues under resource-constrained monitoring conditions. From the perspective of forest ecosystem protection, such capability to identify early fire-related visual cues may support more timely fire management decisions and help reduce vegetation damage and ecological disturbance.