TF-1770: A Visible–Infrared Forest Fire Image Dataset for Multimodal Perception and Image Fusion Research

TF-1770 Dataset Description Additional note: In addition to the TF-1770 main dataset, an independently collected deployment test set containing 1,116 aligned visible–infrared image pairs is provided for edge-deployment and quantization-effect evaluation. This additional test set was not used for model training and can be accessed at: https://zenodo.org/records/20450596 Overview TF-1770 is a multimodal visible–infrared image dataset designed for research on forest fire monitoring, multimodal perception, and visible–infrared image fusion in complex wildfire environments. The dataset contains 1,770 spatially aligned pairs of visible and thermal infrared images, collected from both unmanned aerial vehicle (UAV) platforms and ground-based firefighter perspectives. Forest fire monitoring often occurs under challenging environmental conditions, including: dense smoke low illumination complex vegetation backgrounds varying observation distances In such scenarios, visible and infrared imaging provide complementary information.Visible images capture detailed scene structures and textures, while infrared images are highly sensitive to thermal targets such as flames and high-temperature regions. The TF-1770 dataset was constructed to facilitate research on effectively integrating these complementary modalities for improved wildfire perception, monitoring, and intelligent sensing systems. Dataset Composition The TF-1770 dataset contains 1,770 aligned visible–infrared image pairs collected from two observation perspectives: UAV viewpoint 944 image pairs Captured from UAV platforms at different flight altitudes Provide large-scale views of wildfire spread and thermal distribution Firefighter viewpoint 826 image pairs Captured from ground-level sensing devices simulating firefighter perspectives Include close-range observations of flames, smoke, and complex forest environments Each visible image has a corresponding infrared image with the same filename, ensuring straightforward pairing for multimodal processing. Data Acquisition Data were collected in representative forest environments using a multi-platform sensing strategy. UAV acquisition Aerial observations were collected from UAV platforms at different flight altitudes, providing large-scale monitoring of wildfire scenes and thermal distribution patterns. Ground-level acquisition Ground-based observations were captured using sensing devices simulating the perspectives of firefighters during field operations. This hierarchical acquisition strategy allows the dataset to capture variations in: observation distance target scale environmental complexity smoke conditions illumination variations These characteristics make TF-1770 suitable for developing and evaluating multimodal perception algorithms under realistic wildfire monitoring scenarios. Data Processing All visible–infrared image pairs were spatially aligned before release using the following processing pipeline: Synchronous acquisition of visible and thermal infrared images Pixel-level image registration using the Enhanced Correlation Coefficient (ECC) algorithm Field-of-view alignment by cropping the visible images to match the thermal sensor view Resolution normalization by resizing aligned image pairs to 640 × 480 pixels This processing ensures that the image pairs are spatially consistent and suitable for multimodal learning tasks such as image fusion and cross-modal perception. Dataset Structure The Zenodo record contains the following files and folders: TF-1770 Dataset│├── TF1770_visible_images_v1.zip│ All visible-spectrum images (1770 images)│├── TF1770_infrared_images_v1.zip│ All thermal infrared images (1770 images)│├── TF1770_annotations_v1.zip│ Annotation files corresponding to the image pairs│├── TF1770_preview_samples_v1│ Preview samples (14 aligned visible–infrared image pairs)│├── README└── LICENSE Preview Samples The folder TF1770_preview_samples_v1 contains 14 aligned visible–infrared image pairs that allow users to quickly inspect the dataset structure and data quality before downloading the full dataset. Each visible image corresponds to an infrared image with the same filename. Dataset Applications The TF-1770 dataset supports research in several areas related to wildfire monitoring and multimodal perception, including: visible–infrared image fusion multimodal wildfire detection cross-modal feature learning edge intelligent perception systems ecological monitoring and disaster response technologies Citation If you use the TF-1770 dataset in your research, please cite the following work. License The dataset is released under the license specified in the LICENSE file included in this repository.