In recent years, tragedies caused by forest fires have been frequently reported. Forest fires not only result in significant economic losses but also cause environmental damage. The utilization of computer vision techniques and unmanned aerial vehicles (UAVs) for forest fire monitoring has become a primary approach to accurately locate and extinguish fires during their early stages. However, traditional computer-based methods for UAV forest fire image segmentation require a large amount of pixel-level labeled data to train the networks, which can be time-consuming and costly to acquire. To address this challenge, we propose a novel weakly supervised approach for semantic segmentation of fire images in this study. Our method utilizes self-supervised attention foreground-aware pooling (SAP) and context-aware loss (CAL) to generate high-quality pseudo-labels, serving as substitutes for manual annotation. SAP collaborates with bounding box and class activation mapping (CAM) to generate a background attention map, which aids in the generation of accurate pseudo-labels. CAL further improves the quality of the pseudo-labels by incorporating contextual information related to the target objects, effectively reducing environmental noise. We conducted experiments on two publicly available UAV forest fire datasets: the Corsican dataset and the Flame dataset. Our proposed method achieved impressive results, with IoU values of 81.23% and 76.43% for the Corsican dataset and the Flame dataset, respectively. These results significantly outperform the latest weakly supervised semantic segmentation (WSSS) networks on forest fire datasets.

Unmanned aerial vehicles (UAVs) are an efficient tool for monitoring forest fire due to its advantages, e.g., cost-saving, lightweight, flexible, etc. Semantic segmentation can provide a model aircraft to rapidly and accurately determine the location of a forest fire. However, training a semantic segmentation model requires a large number of labeled images, which is labor-intensive and time-consuming to generate. To address the lack of labeled images, we propose, in this paper, a semi-supervised learning-based segmentation network, SemiFSNet. By taking into account the unique characteristics of UAV-acquired imagery of forest fire, the proposed method first uses occlusion-aware data augmentation for labeled data to increase the robustness of the trained model. In SemiFSNet, a dynamic encoder network replaces the ordinary convolution with dynamic convolution, thus enabling the learned feature to better represent the fire feature with varying size and shape. To mitigate the impact of complex scene background, we also propose a feature refinement module by integrating an attention mechanism to highlight the salient feature information, thus improving the performance of the segmentation network. Additionally, consistency regularization is introduced to exploit the rich information that unlabeled data contain, thus aiding the semi-supervised learning. To validate the effectiveness of the proposed method, extensive experiments were conducted on the Flame dataset and Corsican dataset. The experimental results show that the proposed model outperforms state-of-the-art methods and is competitive to its fully supervised learning counterpart.