Wildfires pose a great threat to the ecological environment and human safety. Therefore, rapid and accurate detection of wildfires holds significant importance. However, existing wildfire detection methods neglect the full integration of spatial–temporal relationships across different scales, and thus suffer from issues of low robustness and accuracy in varying wildfire scenes. To address this, we propose a deep learning model for near-real-time wildfire detection, where the core idea is to integrate multi-scale spatial–temporal features (MSSTF) to efficiently capture the dynamics of wildfires. Specifically, we design a multi-kernel attention-based convolution (MKAC) module for extracting spatial features representing the differences between fire and non-fire pixels within multi-scale receptive fields. Moreover, a long short-term Transformer (LSTT) module is used to capture the temporal differences from the image sequences with different window lengths. The two modules are combined into multiple streams to integrate the multi-scale spatial–temporal features, and the multi-stream features are then fused to generate the fire classification map. Extensive experiments on various fire scenes show that the proposed method is superior to JAXA Wildfire products and representative deep learning models, achieving the best accuracy scores (i.e., average fire accuracy (FA): 88.25%, average false alarm rate (FAR): 20.82%). The results also show that the method is sensitive to early-stage fire events and can be applied in the task of near-real-time wildfire detection with 10-minute Himawari-8/9 satellite data. The data and codes used in the study are detailed in: https://github.com/eagle-void/MSSTF.

中文翻译：

---

基于向日葵 8/9 地球静止卫星数据集成多尺度时空特征的近实时野火探测方法


野火对生态环境和人类安全构成巨大威胁。因此，快速准确地检测野火具有重要意义。然而，现有的野火探测方法忽视了不同尺度上时空关系的充分整合，因此在不同的野火场景中存在鲁棒性和准确性低的问题。为了解决这个问题，我们提出了一种用于近实时野火检测的深度学习模型，其核心思想是集成多尺度时空特征 （MSSTF） 以有效捕获野火的动态。具体来说，我们设计了一个多核基于注意力的卷积 （MKAC） 模块，用于提取表示多尺度感受野中火像素和非火像素之间差异的空间特征。此外，长短期 Transformer （LSTT） 模块用于捕获不同窗口长度的图像序列的时间差异。将两个模块组合成多个流，整合多尺度时空特征，然后将多流特征融合生成火灾分类图。在各种火灾现场的广泛实验表明，所提方法优于 JAXA Wildfire 产品和代表性深度学习模型，取得了最佳准确率得分 （即平均火灾准确率 （FA）：88.25%，平均误报率 （FAR）：20.82%）。结果还表明，该方法对早期火灾事件很敏感，可以应用于使用 10 分钟向日葵 8/9 卫星数据的近实时野火探测任务。研究中使用的数据和代码详见：https://github.com/eagle-void/MSSTF。