Wildfire Dataset

A new dataset was prepared because the open-source datasets prepared for forest fires are single-class, require very high processing power owing to the excessive number of images, and have very long processing times. At the beginning of a fire, dense smoke rises, and after this rising smoke, fire flames begin to grow. Because our aim was early fire detection, the dataset was created from aerial smoke images and images of smoke-like objects in nature. First, 737 images were taken from the open-source Wildfire Smoke Dataset created by the AI for the Makind technology community[1]. Then, 124 smoke-like images were downloaded from various open source websites. By applying data augmentation to 72 downloaded images, 268 images were obtained and combined with the images of the Wildfire Smoke Dataset. A dataset consisting of two classes, divided into three classes, training, testing, and validation, was prepared. A large dataset was created with 737 images containing smoke images and 268 images downloaded from open source websites. At the end of this process, 80% of the new data set was reserved for training, 10% for validation, and 10% for testing. 

1. “Open Wildfire Smoke Datasets.” Accessed: Sep. 11, 2023. [Online]. Available: https://github.com/aiformankind/wildfire-smoke-dataset