In this study, we present an advanced forest fire detection method tailored for China UAV platforms. The approach builds upon the YOLOv11n architecture, introducing three key improvements: a lightweight Adown downsampling module, a mixed aggregation network (MANet) for enhanced feature extraction, and a Separated and Enhancement Attention Module (SEAM) in the detection head. These modifications significantly boost detection accuracy for early-stage forest fires, particularly in challenging scenarios with small flame targets and indistinct smoke textures. Our comprehensive experiments demonstrate that the proposed method achieves superior performance over existing algorithms, making it well-suited for deployment on China UAV systems in real-time monitoring applications.
1. Introduction
Forest fires represent one of the most destructive natural disasters, causing severe ecological damage and threatening human lives. Rapid detection in the early stages is critical for effective firefighting and damage mitigation. Traditional methods such as manual patrols suffer from low efficiency and high costs, especially in large forested areas. This has motivated the adoption of deep learning-based computer vision techniques for fire and smoke detection from aerial platforms, including China UAVs.
Recent advances in object detection, particularly the YOLO family, have shown great promise in fire monitoring tasks. However, existing models often struggle with small objects, complex backgrounds, and occlusions prevalent in UAV imagery. To address these challenges, we propose an improved YOLOv11n algorithm specifically designed for China UAV-based forest fire detection. Our contributions are threefold:
- We introduce a lightweight Adown downsampling module that reduces computational cost while preserving key spatial information.
- We design a Mixed Aggregation Network (MANet) module that integrates multiple convolutional structures to enhance feature representation.
- We incorporate a SEAM attention mechanism in the detection head to improve accuracy under occlusion and background clutter.
Through extensive experiments on a curated dataset of 2,313 early-stage forest fire images (augmented to 6,940 images), we demonstrate that our method outperforms baseline YOLOv11n and other state-of-the-art detectors, achieving a mAP@0.5 of 80.9% with only 3.2 million parameters and 6.8 GFLOPs. These results underscore the suitability of our approach for deployment on resource-constrained China UAV platforms.
2. Proposed Method
The overall architecture is illustrated in the figure below. Our improvement focuses on three key components: replacing the standard downsampling with Adown, embedding MANet in the backbone, and replacing the detection head with SEAM-enhanced layers.
2.1 Lightweight Adown Downsampling
To reduce computational overhead while maintaining feature richness, we adopt the Adown module. Traditional downsampling using stride-2 convolution compresses feature maps too aggressively, losing fine-grained details of small fire and smoke regions. Adown combines average pooling and max pooling with 1×1 and 3×3 convolutions to achieve efficient spatial reduction with minimal information loss.
The computation process is as follows:
$$
\begin{aligned}
X_1, X_2 &= \text{Split}\big(\text{Avgpool2d}(X)\big), \quad X \in \mathbb{R}^{B \times C \times H \times W} \\
Y_1 &= \text{Conv}_{1\times 1}\big(\text{MaxPool2d}(X_1)\big) \\
Y_2 &= \text{Conv}_{3\times 3}(X_2) \\
Y &= \text{Concat}(Y_1, Y_2), \quad Y \in \mathbb{R}^{B \times 2C \times \frac{H}{2} \times \frac{W}{2}}
\end{aligned}
$$
The Adown module reduces the number of parameters by approximately 19% in our experiments (from 2.6×10⁶ to 2.1×10⁶) and decreases FLOPs by 17% (from 6.4×10⁹ to 5.3×10⁹) compared to standard convolution, while even slightly improving mAP.
2.2 Mixed Aggregation Network (MANet)
The MANet module enhances feature representation by combining three different convolutional paths: a 1×1 convolution for channel calibration, a depthwise separable convolution for efficient spatial feature extraction, and a series of residual ConvNeck blocks for hierarchical feature fusion. The design follows an “input expansion – multi-path extraction – fusion compression” paradigm.
The mathematical formulation is given by:
$$
\begin{aligned}
X_{\text{mid}} &= \text{Conv}_1(X_{\text{in}}) \\
X_1 &= \text{Conv}_2(X_{\text{mid}}) \\
X_2 &= \text{DSConv}\big(\text{Conv}_3(X_{\text{mid}})\big) \\
X_3, X_4 &= \text{Split}(X_{\text{mid}}) \\
X_5 &= \text{ConvNeck}_1(X_4) + X_4 \\
X_6 &= \text{ConvNeck}_2(X_5) + X_5 \\
&\vdots \\
X_{4+n} &= \text{ConvNeck}_n(X_{3+n}) + X_{3+n}
\end{aligned}
$$
Finally, all outputs are concatenated and passed through a 1×1 convolution to produce the fused feature map:
$$
X_{\text{out}} = \text{Conv}_0\big( X_1 \parallel X_2 \parallel \cdots \parallel X_{4+n} \big)
$$
The MANet module introduces a moderate increase in parameters (from 2.6×10⁶ to 3.8×10⁶) but yields a 3.5 percentage point improvement in mAP@0.5 over the baseline, as shown in the ablation study.
2.3 SEAM Detection Head
To better handle occlusion and small targets in forest fire scenes, we replace the standard detection head with a Separated and Enhancement Attention Module (SEAM). The SEAM mechanism first splits the feature map into multiple branches, applies channel-wise and spatial attention independently, then reweights and fuses the features to suppress background interference while enhancing target-relevant information.
The SEAM operation can be summarized as:
$$
F_{\text{out}} = \text{SEAM}(F_{\text{in}}) = \text{Concat}\big( \text{Att}_1(F_{\text{in}}), \text{Att}_2(F_{\text{in}}), \dots, \text{Att}_k(F_{\text{in}}) \big) \cdot W
$$
Where Att₁, Att₂, …, Attₖ are different attention mechanisms (e.g., channel attention, spatial attention) applied to the same input, and W is a learned weighting matrix. This enables the model to focus on discriminative regions of tiny flames and semi-transparent smoke.
3. Experiments and Results
All experiments were conducted on a system with an Intel Xeon Platinum 8470Q CPU, NVIDIA GeForce RTX 4090 GPU (24 GB VRAM), Python 3.10.8, PyTorch 2.1.0, and CUDA 11.8. Training settings are summarized in Table 1.
| Parameter | Value | Parameter | Value |
|---|---|---|---|
| epoch | 100 | momentum | 0.937 |
| batch | 32 | Weight_decay | 0.0005 |
| patience | 50 | Warm_up_epochs | 3 |
| Image_size | 640×640 | Warm_up_momentum | 0.8 |
| optimizer | SGD | Close_mosaic | 10 |
| workers | 8 | Lr_0 | 0.1 |
Our dataset consists of 2,313 original forest fire images collected from public flame datasets, satellite imagery, and web searches, focusing on early-stage fires. Data augmentation (rotation, scaling, blurring, brightness adjustment, and noise addition) expanded the dataset to 6,940 images, split into 8:1:1 for training, validation, and testing. All images were annotated with two classes: fire and smoke.
3.1 Ablation Study
We performed an ablation study to evaluate the contribution of each proposed module. Results are shown in Table 2.
| Experiment | Adown | MANet | SEAM | P (%) | R (%) | mAP@0.5 (%) | Params (×10⁶) | GFLOPs (×10⁹) |
|---|---|---|---|---|---|---|---|---|
| 1 (baseline) | ✗ | ✗ | ✗ | 81.5 | 69.5 | 75.3 | 2.6 | 6.4 |
| 2 | ✓ | ✗ | ✗ | 81.9 | 71.5 | 76.2 | 2.1 | 5.3 |
| 3 | ✗ | ✓ | ✗ | 82.9 | 72.3 | 78.8 | 3.8 | 8.4 |
| 4 | ✗ | ✗ | ✓ | 82.5 | 72.1 | 78.3 | 2.5 | 5.8 |
| 5 | ✓ | ✓ | ✗ | 83.4 | 74.0 | 79.5 | 3.4 | 7.6 |
| 6 (ours) | ✓ | ✓ | ✓ | 84.2 | 75.3 | 80.9 | 3.2 | 6.8 |
The baseline YOLOv11n achieves 75.3% mAP. Adding only Adown reduces parameters while slightly improving mAP (+0.9%). MANet alone gives a significant boost of +3.5% in mAP. SEAM alone improves mAP by +3.0%. The full combination yields +5.6% mAP over baseline, with only 23% more parameters and 6.3% more FLOPs, indicating an excellent trade-off.
3.2 Comparison with State-of-the-Art
We compare our method with several popular detectors: YOLOv5n, YOLOv8n, hyper-yolo, YOLOv10n, and YOLOv11n under the same conditions. Results are presented in Table 3.
| Method | P (%) | R (%) | mAP@0.5 (%) | Params (×10⁶) | GFLOPs (×10⁹) |
|---|---|---|---|---|---|
| YOLOv5n | 79.2 | 65.1 | 72.4 | 2.2 | 5.9 |
| YOLOv8n | 80.1 | 67.3 | 74.9 | 2.7 | 6.9 |
| hyper-yolo | 83.5 | 74.7 | 78.2 | 3.6 | 9.7 |
| YOLOv10n | 82.6 | 67.2 | 73.1 | 1.8 | 6.7 |
| YOLOv11n | 81.5 | 69.5 | 75.3 | 2.6 | 6.4 |
| Ours | 84.2 | 75.3 | 80.9 | 3.2 | 6.8 |
Our method achieves the highest precision, recall, and mAP among all compared methods. Although hyper-yolo has comparable performance, our model has significantly lower computational cost (6.8 GFLOPs vs. 9.7 GFLOPs) and fewer parameters, making it more suitable for deployment on China UAVs with limited onboard resources.
3.3 Qualitative Analysis
We visualize detection results to illustrate the improvements. In scenarios with small flames and faint smoke, the baseline YOLOv11n often misses detections or produces low-confidence bounding boxes. Our enhanced model consistently identifies both fire and smoke regions with higher confidence and fewer false positives. This is particularly critical for China UAV applications where early fire detection can prevent large-scale disasters.
4. Conclusion
We have proposed an improved YOLOv11n-based method for forest fire detection from China UAV platforms. By integrating a lightweight Adown downsampling module, a mixed aggregation network (MANet), and a SEAM attention head, our approach significantly enhances detection accuracy for early-stage fires while maintaining low computational complexity. Experimental results show a 2.7% improvement in precision, 5.8% in recall, and 5.6% in mAP@0.5 over the baseline YOLOv11n, with only 3.2 million parameters and 6.8 GFLOPs. Comparisons with other state-of-the-art detectors confirm the superiority of our method. Future work will focus on deploying the model on actual China UAV hardware for real-time surveillance, as well as further optimizing the network for edge inference.