alarmod/forest_fire

The initial pictures from UAV (FLAME dataset, file 9) with the size 3840 × 2160 were splitted in nine non-overlapping parts, each such part having a size of 1280 × 720 pixels. Information about locations with fires was extracted from masks, prepared by developers of primary dataset for segmentation (FLAME, file 10), then, with the help of marking algorithm of connected components the data about points with fire were extracted: information about their sizes and coordinates are getting saved in text files. Further, the fragments of snaps with information about the exact location of place with fire are submitted to ANN. In order to form both the test images and training ones from one dataset, the breakdown of processed data into batches with 90 images each is made. The first packet goes for training, the second one is purposefully discarded, the third one is forwarded for testing, and the fourth one is discarded. This is done on purpose, in order to ensure that the tests and training set will be from video stream’s frames different in time. Eventually, training and test sets appeared to have 4500 images each (with textual description of fire zones). Respectively, 2601 and 2604 of them – images of forest without places with fire.

Thanks to Shamsoshoara, A.; Afghah, F.; Razi, A.; Zheng, L.; Fulé, P. The Flame Dataset: Aerial Imagery Pile Burn Detection Using Drones (UAVS). 2021. Available online: https://ieee-dataport.org/open-access/flame-dataset-aerial-imagery-pile-burn-detection-using-drones-uavs, doi: 10.21227/qad6-r683