Dataset of Fire Image Classification by Neural Networks for EdgeAI Applications

Available here, are the image datasets used to train and evaluate a fully connected feed-forward neural network (FireFFNN) for detecting wildfire in images. The work is part of a thesis on lightweight, compressible fire classifiers for edge devices such as UAVs. All variants were built from two public datasets: FlameVision (Jafar et al., 2023, Mendeley Data, doi:10.17632/fgvscdjsmt.4): 8,600 aerial wildfire images (5,000 fire / 3,600 no fire), originally 512×512 px, with many near-identical consecutive video frames. The Wildfire Dataset (El-Madafri et al., 2023, Forests 14(9):1697, doi:10.3390/f14091697): 2,699 aerial and ground images (1,045 fire / 1,654 no fire), including look-alike scenes such as fog, cloud and sunlight. Images are stored as PNG or JPG. The folders flamevision/, flamevision_originals_114/, flamevision_augmented500/ and flamevision_augmented1000/ are stored at 128×128 px. Other folders are stored at 512×512 px. Each class has its own folder, so the data loads directly with torchvision.datasets.ImageFolder. Resizing to 128×128 and scaling to [0, 1] happen at training time in the accompanying code. Dataset variants Folder Thesis name Classes Images Purpose flamevision/ FlameVision fire / nofire 8,600 (provided train/val/test) Source dataset wildfire/ Wildfire fire / nofire 2,699 (provided train/val/test) Source dataset flamevision_wildfire_combined/ Combined fire / nofire 11,200 (6,046 / 5,154) Both sources pooled; the 100 no-fire images that FlameVision contains twice are kept once flamevision_wildfire_combined_AfterLabelling/ Combined-relabelled fire_or_smoke / nofire 10,365 (5,291 / 5,074) Cleaned and relabelled binary set; Fire covers flame and/or smoke. Used to train the binary model M1 multiclass/ Multiclass-raw fire_smoke / smoke / nofire 10,236 (4,701 / 589 / 4,946) Fire images sorted by hand into visible flame vs. smoke only. Strongly imbalanced. Used for the controlled model M2-C multiclass_600/ Multiclass-600 fire_smoke / smoke / nofire 1,789 (600 / 589 / 600) Balanced subset of Multiclass-raw multiclass_augmented1000/ Multiclass-Aug1000 fire_smoke / smoke / nofire 3,000 (1,000 per class) 589 originals plus 411 offline-augmented copies per class multiclass_augmented1500/ Multiclass-Aug1500 fire_smoke / smoke / nofire 4,500 (1,500 per class) 589 originals plus 911 offline-augmented copies per class. Used to train the final multiclass model M2 crossdatasetexp/ Cross-source fire / nofire train/val from FlameVision; test from Wildfire (159 / 242) Cross-dataset generalisation test Exploratory variants from early experiments are also included: flamevision_originals_114/, flamevision_augmented500/, flamevision_augmented500folderSplit/, flamevision_augmented1000/ and fire_combined120/. None of them is used for the final results. Offline augmentation Each augmented copy was made from one original image using one to three of these transformations: horizontal flip, small rotation, zoom, brightness, contrast, and a slight saturation change. Vertical flips and hue shifts were left out on purpose. The file name records the source image and the transformations, e.g. aug_fire (1406)_hflip-rotate_0094.jpg. When splitting the augmented sets, keep each copy in the same split as its source image, or the test results will be inflated by leakage. Splits Folders with train/val/test subfolders use the split the source dataset provides. All other folders are split at training time with a stratified random split (seed 42): 70/20/10 for the binary model and 70/10/20 for the balanced three-class sets. Out-of-distribution testing The models were also tested on a hand-picked sample of the FLAME 3 dataset (Hopkins et al., 2024, doi:10.21227/w0mz-aq48). Those images are not part of this record and should be downloaded from the original source. Note: If you use this data, please also cite FlameVision and The Wildfire Dataset.