.png){kind=icon}
.png){kind=icon}
Return
Forest fire object detection based on multi-task model and extreme weather simulation algorithm
.png){kind=icon}
PRE
AI
DOI：10.1016/j.engappai.2025.113099.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
The publicly available dataset contains only a single forest fire scenario, leading to a model that excessively depends on a restricted set of features, thereby lacking the robustness to address the intricacies of practical wildfire detection environments. For this reason, an image enhancement method for simulating complex weather is proposed, which converts the original images into foggy, high light, nighttime and rainy images to increase the diversity of the data. In this study, a multi-task learning framework for wildfire monitoring is developed, prioritizing wildfire object recognition as the primary objective, while integrating three complementary tasks, including wildfire object identification, forest fire semantic segmentation and forest fire image classification. The algorithmic model enhances the network's capability for forest fire image feature extraction through a multi-task shared backbone network. This shared feature extraction network allows multiple tasks to learn from each other, improving their performance and detection accuracy, while also providing implicit data enhancement effects. In addition, we developed an efficient lightweight backbone that includes a backward residual structure to increase the sensory field of the model. In the neck network, a three-feature fusion module (TFFM) and a multi-feature aggregation network module (MFANM) are introduced. Evaluation results demonstrate that, relative to the YOLOv7(You Only Look Once version 7) baseline, the developed model achieves a 5.5 % enhancement in mean average precision (mAP) and a 6.5 % increase in average precision for small-scale object detection (APS) within wildfire-specific scenarios, other metrics are improved by 3 %–8 %, and the amount of parameters is reduced by 67 %. In the early detection of forest fires, the model considers both flame and smoke information to characterize the fire conditions and effectively combines their semantic information for early fire warning.
Journal
IF:
8
Papers:
5.3K
Citations:
3.5W