Data argumentation facility used during modifiable areal unit problem research project. Read more in our article on Medium - Generative adversarial networks in satellite image datasets augmentation.
Sample data can be downloaded from our S3 bucket or by utilizing the scgan/data_download.py.
To produce you own data you can use scgan/gdal_operations.py. Please note that three files will be needed: TIFF representing land cover, TIFF with satellite image and generate grid in form of an ArcGis shape file.
Dataset has to meet following criteria in terms of the directory structure:
dataset (name of the dataset)
├── train (samples used during training)
│ ├── data_descriptor.csv (names / ids of the files)
│ ├── LC (land cover data folder)
│ │ ├── LC_10.tif
│ │ ├── LC_1.tif
....................
│ │ └── LC_n.tif
│ └── S
│ ├── S_10.tif
│ ├── S_1.tif
....................
│ └── S_n.tif
├── plot (samples used during intermediate result plotting after each epoch)
│ ├── data_descriptor.csv (names / ids of the files)
│ ├── LC (land cover data folder)
│ │ ├── LC_10.tif
│ │ ├── LC_1.tif
....................
│ │ └── LC_n.tif
│ └── S
│ ├── S_10.tif
│ ├── S_1.tif
....................
│ └── S_n.tif
└── test (samples used during predict phase)
├── data_descriptor.csv (names / ids of the files)
├── LC (land cover data folder)
│ ├── LC_10.tif
│ ├── LC_1.tif
....................
│ └── LC_n.tif
└── S
├── S_10.tif
....................
└── S_1.tif
The default training configuration can be run from scgan/train.py. Default dataset is called bdot. Please note that chosen hyperparameters were set for the training to perform best on the sample dataset related to central Poland and Sentinel-2 images.
To generate artificial satellite images from predefined mask use scgan/predict. If you did not train a model you can download one of ours from S3. Masks have to placed in relevant dataset test subdirectory.
