UAVDisaster36K benchmark

This repository containes the data for the following publication: Demetris Shianios, Christos Kyrkou, "Advancing Efficiency and Accuracy in Aerial Image Classification for Disaster Incidents with SqueezeViT and UAVDisaster36K benchmark." Neurocomputing, vol. 681, p. 133282, Jun. 2026, doi: 10.1016/j.neucom.2026.133282. It contains the zip files with all imeages in each category. The csv file contains the different fold setups used in the paper. The 4th fold which demonstrated the most correlration with the average is also provide in train/val/test folders in the Disaster36K_Fold4.zip file. Detail Description Subject Area The dataset focuses on UAV-captured images encompassing disasters like earthquakes, floods, fires, and normal cases. Data Collection The dataset was obtained by collecting aerial images from drones, UAVs (Unmanned Aerial Vehicles), and other low-altitude aircraft, utilizing targeted disaster keywords. Data Partition The dataset was partitioned into training, validation, and test sets using a video-based approach to mitigate data leakage. Data Cleaning Removal of irrelevant, branded, terrestrial, and low-quality (e.g., blurry) images. Validity The dataset was reviewed and curated by a team of two researchers. Data Format Images: .jpg, Videos: .mp4 Altidue and bearing Viewing angles range from near-nadir to approximately oblique at altitude between 10 and 300 meters. Number of Data Earthquake Images: 6595, Videos:45 Flood Images: 8905, Videos: 37 Fire Images: 10160, Videos: 45 Normal Images: 10340, Videos: 32 Resolutions per Class Earthquake Width: Min 694, Max 1920, Mean 1310.20 Height: Min 480, Max 1080, Mean 763.71 Flood Width: Min 960, Max 1920, Mean 1329.70 Height: Min 620, Max 1080, Mean 749.72 Fire Width: Min 640, Max 1920, Mean 1200.82 Height: Min 360, Max 1080, Mean 680.109 Normal Width: Min 426, Max 1920, Mean 1103.36 Height: Min 240, Max 1080, Mean 597.083