Graph Neural Network-Based Structured Scene Graph Generation for Efficient Wildfire Detection

Yanning Ye, Shimin Luo, Mengmeng Jing, Yongqi Ding, Kunbin He, Lin Zuo

Abstract: Wildfires are large, rapidly spreading fires in the natural environment that pose a significant threat to economic property and human safety, and it is important to detect and extinguish them quickly. Usually smoke is one of the signs of wildfires, it helps us to detect the wildfires. However, smoke is visually similar to reflections of sunlight in clouds or rivers, which typically results in high false positive rates and low accuracy. To mitigate this problem, we propose a novel method based on the structured scene graph information for wildfire detection. Specifically, we first learn the structured information based on the graph neural network to model neighborhood relations among samples. Then, we employ the structured information fusion based on the gated graph neural network to dynamically capture semantic features and flexibly fuse multi-dimensional information, making the understanding of wildfire object more comprehensive and accurate. Extensive experiments on public and self-constructed real datasets confirm the effectiveness of the proposed method. In comparison with existing methods, the proposed method has a higher recall rate.