A Dataset for Fine-grained Aircraft Classification in Low-Resolution Optical Remote Sensing Images

Our dataset is built on the public remote sensing image military aircraft recognition dataset called MAR20. The MAR20 dataset was published by Northwestern Polytechnical University in 2023 in the National Remote Sensing Bulletin. The MAR20 dataset contains 3842 high-resolution optical remote sensing images of airport scenes collected from Google Earth, including 22341 aircrafts of 20 military aircraft categories (SU-35, C-130, C-17, C-5, F-16, TU-160, E-3, B-52, P-3C, B-1B, E-8, TU-22, F-15, KC-135, F-22, FA-18, TU-95, KC-10, SU-34, SU-24). The training and testing sets of the MAR20 dataset contain 1331 and 2511 airport scene images, respectively.  The construction process of our dataset is as follows. Firstly, based on the bounding box annotation provided by the MAR20 dataset, we crop aircrafts from airport scene images to generate single-object slices as high-resolution remote sensing images of aircrafts. The division of the training and testing sets for these remote sensing images of aircrafts is consistent with the airport scene images where they are from. After all aircraft images are obtained and divided, 10% of the aircraft images of each category are randomly selected from the training set and moved to the validation set. Then, the high-resolution remote sensing images of aircrafts are downsampled via bicubic interpolation to obtain corresponding low-resolution remote sensing images. Following the common pixel-based definition of low-resolution small objects and previous studies on small objects in remote sensing images, we generate low-resolution remote sensing images by downsampling aircrafts to three kinds of absolute sizes including 32 pixels, 20 pixels, and 12 pixels, where the absolute size is the square root of the product of the object length and width pixels.