Dataset of Fire Image Classification by Neural Networks for EdgeAI Applications

This record contains the image datasets used to train and evaluate FireFFNN, a fully connected feed-forward neural network for detecting wildfire in images. The work is part of a thesis on lightweight, compressible fire classifiers for edge devices such as UAVs. All datasets 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) at 512×512 px, divided into training, validation and test parts. 100 no-fire images occur twice, and many images are 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 of varying size (1,045 fire / 1,654 no fire), divided into training, validation and test parts. It includes look-alike scenes such as fog, cloud and sunlight. The thesis calls it Wildfire. Files and format Each dataset is a separate ZIP archive named after its folder, and each archive unpacks to that folder, e.g. multiclass_augmented1500.zip → multiclass_augmented1500/smoke/…. Each class has its own subfolder, so a dataset (or each of its train/val/test parts) loads directly with torchvision.datasets.ImageFolder. All images are 3-channel RGB, stored as PNG or JPG. In the augmented datasets, all images, including the originals, are JPG. The images in flamevision/, flamevision_originals_114/, flamevision_augmented500/ and flamevision_augmented1000/ are 128×128 px. All other images are 512×512 px; the Wildfire images were resized to this size from their original resolutions. For training, the accompanying code resizes every image to 128×128 px (antialiased bilinear interpolation, aspect ratio not preserved) and scales the values to [0, 1]. Class folders: fire and nofire are the labels of the source datasets. In the relabelled binary dataset, fire_or_smoke holds every image with flame and/or smoke (class Fire in the thesis). In the three-class datasets, fire_smoke holds images with visible flame, possibly together with smoke (Fire), smoke holds images with smoke but no visible flame (Smoke), and nofire corresponds to No Fire. Datasets Archive Name in thesis Class folders Images Description flamevision.zip FlameVision fire / nofire 8,600 (5,000 / 3,600) Source dataset at 128×128 px, with its provided parts: train 3,600 / 3,200, val 700 / 200, test 700 / 200 wildfire.zip Wildfire fire / nofire 2,699 (1,045 / 1,654) Source dataset with its provided parts: train 730 / 1,157, val 156 / 246, test 159 / 251 flamevision_wildfire_combined.zip FlameVisionWildfireCombined fire / nofire 11,200 (6,046 / 5,154) Both sources pooled with their original labels; the 100 no-fire images that FlameVision contains twice are kept once. fire/ also contains one duplicated file, fire (3613) - Copy.png flamevision_wildfire_combined_AfterLabelling.zip FlameVisionWildfireCombined_afterlabelling fire_or_smoke / nofire 10,365 (5,291 / 5,074) Cleaned and relabelled binary dataset. Used to train the binary model M1 multiclass.zip multiclass fire_smoke / smoke / nofire 10,236 (4,701 / 589 / 4,946) Fire images sorted by hand into visible flame vs. smoke only. Strongly imbalanced (5.75 % smoke). Used to train the controlled multiclass model M2-C multiclass_600.zip multiclass_600 fire_smoke / smoke / nofire 1,789 (600 / 589 / 600) Balanced subset of multiclass, without augmented copies multiclass_augmented1000.zip multiclass_augmented1000 fire_smoke / smoke / nofire 3,000 (1,000 per class) 589 originals plus 411 offline-augmented copies per class; each original has at most one copy multiclass_augmented1500.zip multiclass_augmented1500 fire_smoke / smoke / nofire 4,500 (1,500 per class) 589 originals plus 911 offline-augmented copies per class; each original has one or two copies. Used to train the final multiclass model M2 crossdatasetexp.zip Cross-source fire / nofire 8,101 Cross-dataset generalisation test. train and val are the FlameVision training and validation parts (3,600 / 3,200 and 700 / 200); test holds 159 / 242 images from the Wildfire test part Four exploratory datasets from early experiments are also included. They are not used for the results reported in the thesis: flamevision_originals_114.zip: 114 fire and 114 no-fire FlameVision images (128×128 px). flamevision_augmented500.zip: 500 images per class, made up of 94 of these originals and 406 augmented copies (128×128 px). flamevision_augmented1000.zip: 1,000 images per class, made up of the same 94 originals and 906 augmented copies (128×128 px). fire_combined120.zip: 120 images per class; fire/ holds the 114 FlameVision fire images above plus 6 Wildfire images, and nofire/ holds 120 FlameVision images (512×512 px). Offline augmentation Each augmented copy was made from one original image with one to three of these transformations: horizontal flip, rotation (±12°), zoom in (80–100 % of the frame), brightness (factor 0.75–1.25), contrast (factor 0.80–1.30) and saturation (factor 0.85–1.15). Vertical flips and hue shifts were left out on purpose, because flames and smoke rise upwards and colour is the main cue for flame. Originals carry the prefix orig_. A copy is named aug_<source image>_<transformations>_<index>.jpg, e.g. aug_11713547914_dd11630b77_o_hflip-saturation-brightness_0000.jpg is a copy of orig_11713547914_dd11630b77_o.jpg. When you split an augmented dataset, keep each copy in the same split as its source image; otherwise near-identical images end up in both the training and the test data and inflate the test results. The random split used in the thesis did not group copies with their source images, and the thesis reports how many test images are affected. Splits The train/val/test subfolders of flamevision/, wildfire/ and crossdatasetexp/ follow the parts provided by the source datasets. In the thesis, the parts were used as provided only for the cross-dataset test (crossdatasetexp/). All other datasets, including FlameVision and Wildfire, were pooled and split at training time with a stratified random split (seed 42): 70/20/10 (train/val/test) for the binary datasets and for multiclass (M1, M2-C), and 70/10/20 for multiclass_600, multiclass_augmented1000 and multiclass_augmented1500 (M2). Out-of-distribution testing The models were also tested on two hand-picked samples of RGB images from the FLAME 3 dataset (Hopkins et al., 2024, IEEE DataPort, doi:10.21227/w0mz-aq48), 102 distinct images in total. These 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 Wildfire and FLAME 3 dataset.