BohdanZasymovych/aircraft-classification

Russian Military Aircraft in Satellite Imagery

This is the data for the aircraft-classification project. That project detects and classifies Russian military aircraft in satellite images of airbases. It contains:

• real images derived from a public Kaggle dataset
• artificial images generated by placing aircraft cut-outs onto empty airbase images
• the assets used to generate the artificial images

The layout of this repository matches the data/ directory of the GitHub project. To download it, run python3 download_data.py from the project root.

Structure

artifitial-data/
├── airbase-images.zip        # same as airbase-images/, zipped
├── airbase-images/
│   ├── images/               # 30 empty airbase images (no aircraft)
│   └── labels/               # real size and aircraft spawn points for each airbase
├── plane-images.zip          # aircraft cut-outs for 12 types
├── plane-images/             # unzipped subset: 6 types used in the 6-class datasets
│   ├── class_names_to_id_mapping.json
│   └── <plane-type>/
│       ├── no-background/    # PNG cut-outs with transparent background
│       ├── with-background/  # original crops
│       └── size-labels/      # real wingspan of the aircraft, in meters
└── combined-images-dataset/  # unsplit source of datasets/russian-planes-6cls-combined
    ├── images/
    └── labels/
datasets/
├── dataset-archives/
│   ├── original-dataset.zip                # original Kaggle dataset (Pascal VOC .xml labels)
│   ├── all-planes-yolov8-dataset.zip       # Kaggle dataset converted to YOLO, all 20 classes
│   └── russian-planes-yolov8-dataset.zip   # Kaggle dataset converted to YOLO, Russian aircraft only
├── russian-planes-6cls/
├── russian-planes-6cls-artifitial/
├── russian-planes-6cls-combined/
└── russian-planes-11cls-artifitial/        # configs and test results only, no images

Each dataset folder contains:

• images/{train,val,test}/ and labels/{train,val,test}/, in YOLO format
• data.yaml, the YOLO data file
• config.yaml, the training and augmentation parameters used for the corresponding model
• class_names_mapping.json, which maps each class id to its name
• test-results-<model>/, the evaluation plots of the model trained on the dataset

Datasets

Dataset	Images	Train	Val	Test
russian-planes-6cls	real	933	234	83
russian-planes-6cls-artifitial	artificial train/val, real test	3,200	800	83
russian-planes-6cls-combined	real + artificial train/val, real test	3,541	626	83

All three datasets use the same 83 real test images, so models trained on them can be compared directly.

Classes (6-class datasets)

Id	Class
0	Su-27/Su-35
1	Tu-160
2	Tu-22
3	Tu-95
4	Su-30/Su-34
5	Su-24

Su-27 and Su-35, and Su-30 and Su-34, are each merged into one class. Their small visual differences make them hard to tell apart in satellite images.

The 11-class configuration adds A-50, Il-76, MiG-29, MiG-31 and Su-25. plane-images.zip also contains L-39.

Label formats

Dataset labels use the YOLO format: one object per line, with coordinates normalized to [0, 1]:

<class_id> <x_center> <y_center> <width> <height>

Airbase labels (artifitial-data/airbase-images/labels/<n>.txt): the first line is the real width and height of the image in meters. Each following line is a point where an aircraft can be placed, written as x,y,type. The type is s (small aircraft only) or b (any aircraft):

700,700
455,144,s
415,174,b

Plane size labels (plane-images/<type>/size-labels/<name>.txt) contain the real wingspan of the aircraft in meters. It's used to scale the cut-out to the resolution of the airbase image.

How the artificial images were made

Aircraft cut-outs and empty airbase images were taken from public Google Earth imagery. For each generated image:

• Aircraft are placed on randomly chosen spawn points of an airbase, with a small random offset.
• A shadow is added for each aircraft, with one random shadow direction shared by the whole image.
• Saturation, brightness, contrast, blur and sharpness are randomized for the aircraft and their shadows.
• Gaussian noise and blur are applied to the final image.

The generator code is in scripts/artifitial_image_generator.py in the GitHub project.

Source and license

The real images come from the Kaggle Military Aircraft Recognition dataset, which contains 3,842 images of 20 aircraft types. Check that dataset's license before reusing any data derived from it. This applies to dataset-archives/, russian-planes-6cls, and the real images in russian-planes-6cls-combined and in the test sets.

The airbase images and aircraft cut-outs were made from Google Earth imagery and are subject to Google Earth's terms of use.