FZSNet: Burning Area Segmentation Network for Real-time Detection in Remote Sensing Images

The dataset used in this study is designed for burning area segmentation in remote sensing images, aiming to support accurate and real-time detection of fire regions. The dataset is collected from multi-source remote sensing imagery, including high-resolution RGB images and partial multispectral data, covering diverse scenarios such as forests, grasslands, and urban fringes, which ensures strong scene diversity and complexity.The image resolution ranges from 512×512512 \times 512512×512 to 1024×10241024 \times 10241024×1024. All images are uniformly preprocessed and cropped to meet the input requirements of the proposed network. Pixel-level annotations are provided for burning regions, which are manually labeled by experts based on flame color characteristics, smoke distribution, and thermal radiation cues, ensuring high annotation accuracy and consistency. The annotation categories are defined as foreground (burning area) and background.To enhance the generalization capability of the model, the dataset incorporates various challenging factors, including illumination variations, smoke occlusion, background interference (e.g., red roofs and bare soil), and scale variations. Additionally, considering the small proportion of burning regions in many cases, a significant number of small-scale fire targets are retained to improve the model’s sensitivity to fine-grained objects.The dataset is divided into training, validation, and testing subsets with a typical ratio (e.g., 7:2:1), ensuring a balanced distribution of different scenarios. Furthermore, data augmentation techniques such as random rotation, flipping, scaling, and color jittering are applied during training to improve model robustness.本研究使用的数据集专为遥感图像中的燃烧区域分割（burning area segmentation）任务设计，旨在支持火灾区域的精准实时检测。该数据集采集自多源遥感影像，涵盖高分辨率RGB图像与部分多光谱数据，覆盖森林、草原、城市边缘等多样化场景，具备较强的场景多样性与复杂性。图像分辨率范围为512×512至1024×1024，所有图像均经过统一预处理与裁剪，以适配所提出网络的输入需求。数据集提供燃烧区域的像素级标注（pixel-level annotations），由专家基于火焰颜色特征、烟雾分布与热辐射线索人工标注，确保了极高的标注精度与一致性。标注类别仅分为前景（燃烧区域）与背景两类。为提升模型的泛化能力，该数据集纳入了多种挑战性因素，包括光照变化、烟雾遮挡、背景干扰（如红色屋顶与裸土）以及尺度变化。此外，考虑到多数场景中燃烧区域占比极小，数据集保留了大量小规模火灾目标，以提升模型对细粒度目标的感知能力。数据集按照典型比例（如7:2:1）划分为训练集、验证集与测试集，确保不同场景的分布均衡。此外，训练阶段采用随机旋转、翻转、缩放与色彩抖动等数据增强技术（data augmentation techniques），以提升模型的鲁棒性。