Technical Summary: Dynamic Lightweight YOLO for UAV-Based Forest Fire Detection

1. Problem Statement

Unmanned Aerial Vehicle (UAV) based forest fire detection faces significant hurdles due to the limitations of onboard computational resources and the inherent complexity of fire scenarios. Traditional detection methods, such as manual patrols, ground sensors, and satellite remote sensing, suffer from issues regarding coverage, response time, or resolution. While deep learning offers a promising alternative, existing object detection models struggle with:

• Small Target Detection: Smoke and flames often appear as small, blurred, or irregular targets, leading to high rates of missed detections.
• Complex Backgrounds: Variations in lighting, weather, and terrain (e.g., vegetation, buildings) cause high false detection rates.
• Resource Constraints: UAVs have limited processing power and memory, necessitating lightweight models that do not compromise real-time performance.
• Dynamic Nature: The rapid movement and deformation of flames and smoke challenge static feature extraction methods.

2. Methodology

The authors propose an improved lightweight object detection algorithm based on the YOLOv8n architecture. The model integrates four specific modifications to address the aforementioned challenges while maintaining a compact footprint:

A. Dynamic Convolution Backbone

To handle uneven brightness and dynamic target shapes, the standard convolution layers in the backbone are replaced with a Dynamic Convolution (DynamicConv) module.

• Mechanism: Instead of using fixed kernels, this module dynamically generates weights and biases based on the input features. It employs an attention mechanism to select and combine multiple convolution kernels ($K$) weighted by $\pi_k(x)$.
• Benefit: This allows the network to adaptively adjust its parameters for different input samples, enhancing feature extraction for irregular smoke and flame boundaries without significantly increasing Floating Point Operations (FLOPs).

B. Multi-dimensional Dynamic Detection Head (DyHead)

The original detection head is replaced with the DyHead module to improve the perception of object size, spatial information, and task characteristics.

• Mechanism: The head utilizes three parallel attention mechanisms:
  1. Scale-Aware Attention ($\pi_L$): Models positional information to better handle irregular flame boundaries.
  2. Spatial-Aware Attention ($\pi_S$): Optimizes spatial relationships to distinguish fire from background elements like leaves or sunlight reflections.
  3. Task-Aware Attention ($\pi_C$): Adjusts channel weights to focus on high-brightness flame regions and suppress background noise.
• Benefit: This coordinated processing significantly improves localization accuracy and reduces false positives in complex forest environments.

C. Efficient Multi-scale Attention (EMA) in the Neck

An EMA module based on cross-spatial learning is introduced into the feature fusion neck.

• Mechanism: Unlike traditional attention mechanisms (e.g., SENet, CBAM) that compress channel information via global average pooling, EMA encodes global contextual information across both channel and spatial dimensions. It dynamically recalibrates channel weights for each parallel branch.
• Benefit: This strengthens the representation of multi-scale features, specifically enhancing the detection of small flames and early-stage smoke while suppressing irrelevant background information.

D. Lightweight Feature Fusion (Slim-neck)

To further reduce model size and computational cost, the Slim-neck module is applied.

• Mechanism: This structure utilizes VoV-GSCSP and GSConv (Group Shuffle Convolution) layers. It combines depthwise separable convolutions with structural reparameterization and channel shuffling.
• Benefit: This design reduces the number of parameters and FLOPs while maintaining feature diversity and fusion efficiency, making the model suitable for resource-constrained UAV deployment.

3. Key Contributions

The paper claims the following primary contributions:

1. Dynamic Resource Allocation: The integration of DynamicConv optimizes computational resource usage, improving performance under low FLOPs conditions.
2. Enhanced Multi-scale Perception: The DyHead module enables flexible adjustment of detection strategies based on target size and spatial context, crucial for varying flame scales.
3. Robust Feature Representation: The EMA mechanism improves the model's ability to capture subtle fire features in complex, dynamic environments.
4. Lightweight Architecture: The Slim-neck design successfully balances detection accuracy with a compact model size, facilitating real-time inference on UAVs.

4. Experimental Results

The proposed model was evaluated on the C4-AI public dataset (9,848 images of fire and smoke), split into training, validation, and test sets. Experiments were conducted on an NVIDIA RTX4060 GPU.

• Performance Metrics:
  • Precision: 75.7% (an increase of 5.1% over the baseline YOLOv8n).
  • mAP50: 69.5% (an improvement of 2.7% over YOLOv8n).
  • mAP50-95: 39.3%.
  • Model Size: 10.6 MB.
  • Parameters: 5.4 Million.
  • FLOPs: 13.0 G.
• Comparative Analysis:
  • The proposed model outperformed other YOLO variants (YOLOv5n, YOLOv7t, YOLOv9t, YOLOv10n, YOLOv11n) in terms of the balance between accuracy and model size.
  • Compared to YOLOv7t (6.0M params), the proposed model achieved a 7.9% higher mAP50-95 with fewer parameters (5.4M).
  • Ablation studies confirmed that the combination of all four modules yielded the optimal performance, with the EMA module showing the most significant individual impact on accuracy.
• Qualitative Results: Visual comparisons demonstrated that the improved model produced tighter bounding boxes around flames, reduced false detections in non-fire areas, and successfully identified multiple small fire points without merging or missing targets.

5. Significance and Claims

The authors assert that this research provides a cost-effective and reliable technical solution for real-time forest fire monitoring using UAVs. The significance of the work lies in:

• Synergistic Optimization: Achieving simultaneous improvements in detection accuracy and computational efficiency through multi-dimensional optimization strategies.
• Edge Deployment Feasibility: Demonstrating that high-precision fire detection is possible on devices with limited storage (10.6 MB) and computing power, overcoming the bottlenecks of traditional heavy models.
• Practical Impact: The model significantly reduces false detection rates (by 37% in complex backgrounds) and increases recall for small targets (by 21%), which is critical for early fire warning systems.

Limitations and Future Work:
The authors modestly note that while the model performs well on the C4-AI dataset, challenges remain under extreme lighting conditions (e.g., strong backlighting, nighttime infrared). They suggest that future work could involve fusing infrared spectral data and developing adaptive exposure compensation. Additionally, the current model focuses on static flame detection, and tracking rapidly spreading fire lines remains an area for further validation. Future research directions include developing fusion mechanisms for extreme environments and building fire spread prediction models.