Sentinel generative conditional adversarial network implementation
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.
