DeepFire dataset: A manually annotated sentinel 2 deep learning dataset for burned area semantic segmentation

DeepFire Dataset: A Manually Annotated Sentinel-2 Deep Learning Dataset for Burned Area Mapping The DeepFire dataset is a manually annotated, Sentinel-2-based deep learning dataset developed for burned area semantic segmentation using deep learning and computer vision approaches. The dataset is designed to support the development, training, validation, and benchmarking of semantic segmentation models for the automatic detection and delineation of wildfire-affected areas from multispectral satellite imagery. The dataset was developed to address the need for high-quality, pixel-level annotated training data for burned area mapping. Unlike datasets based solely on automatically generated fire products or threshold-based classifications, DeepFire provides manually prepared burned-area reference masks intended to support supervised deep learning and reproducible evaluation. Data Source The dataset is based on Sentinel-2 multispectral satellite imagery, which provides high-resolution optical observations suitable for detecting changes associated with wildfire disturbances. The imagery was processed and prepared specifically for semantic segmentation applications. The input data consist of multispectral image patches, while the corresponding target dzzata contain binary pixel-level masks representing burned and unburned areas. The dataset is organized into image–mask pairs so that each input image can be directly associated with its corresponding ground-truth segmentation mask. Dataset Structure The released dataset contains: 3,509 image patches Image patch size: 128 × 128 pixels 6 input channels/bands per image Corresponding binary burned-area segmentation masks Data provided in NumPy (.npy) format The principal files include: X_train.npy — Sentinel-2 image patches used as model inputs. y_train.npy — Corresponding manually annotated burned-area masks. The input array is organized as: X_train: (3509, 128, 128, 6) and the corresponding segmentation masks are provided in: y_train: (3509, 128, 128, 1) The masks represent the spatial distribution of burned and unburned pixels and are intended for binary semantic segmentation. Annotation Burned areas were manually delineated to provide pixel-level reference information for supervised learning. The annotations were prepared with the objective of accurately representing the spatial extent of wildfire-affected areas visible in the satellite imagery. The manually annotated masks can therefore be used as ground-truth/reference data for training and evaluating segmentation algorithms. Intended Applications DeepFire is intended primarily for research and development in: Burned area semantic segmentation Wildfire mapping Remote sensing image segmentation Deep learning for Earth observation Forest fire and disaster mapping Satellite-based environmental monitoring Multispectral image analysis Computer vision for remote sensing Benchmarking semantic segmentation architectures The dataset can be used with a wide range of deep learning architectures, including U-Net, DeepLabV3+, SegFormer, Mask2Former, Swin-based segmentation models, TransUNet, and other convolutional or transformer-based semantic segmentation approaches. Dataset Format The data are provided in NumPy binary (.npy) format, allowing direct integration with Python-based machine learning and deep learning frameworks such as NumPy, TensorFlow/Keras, and PyTorch. The input data and corresponding masks are spatially aligned, allowing researchers to load the arrays and directly construct training, validation, and testing subsets according to their experimental design.