GAFNet: a novel semantic segmentation network across aerial and ground imagery for wildfire detection

Wildfires pose significant threats to ecosystems and human safety, making timely and accurate fire monitoring essential for assessing fire expansion trends and controlling fires. Although many studies have been devoted to detecting wildfires using aerial or ground imagery, separately processing aerial and ground imagery can complicate wildfire monitoring. Due to the different viewpoints and scales of the two data sources, it is essential to address the challenges posed by small-area fire in aerial imagery and occlusions in ground imagery, and the difficulties introduced by multi-view and multi-scale scenarios in across-data-source processing. Therefore, an efficient wildfire detection method across the two sources is urgently needed. To address these issues, a novel network named ground and aerial fire network (GAFNet) was designed for wildfire detection in aerial and ground imagery. In GAFNet, a dual-attention efficient layer aggregation network (DAELAN) module was introduced to learn wildfire features across aerial and ground imagery, which reduced the impact of multi-view scenarios and occlusions on fire detection. Meanwhile, a multi-level feature fusion (MLFF) module and feature aggregation (FA) structure were designed to improve the integration of semantic and detailed features and enhance the capability to detect small fire areas from different-scale images. GAFNet was evaluated comprehensively on publicly available aerial, ground, and mixed air – ground datasets. The results demonstrate that the proposed network achieves the best performance on all datasets. Especially on the most difficult hybrid ground-to-air wildfire detection missions, GAFNet increased IoU by 2.4%+ and F1 scores by 1.2%+ compared to other semantic‑segmentation architectures and state‑of‑the‑art models. This study has provided an effective solution for obtaining accurate wildfire information.