Download the dataset from https://zenodo.org/records/10939100
Dataset paper published in arXiv:2404.05180
This dataset comprises a total of 13,703 samples of multispectral satellite images captured across continents from 2021 to 2023 in various solar panel farms. The detailed metadata for this dataset is provided in the following table:
| Dataset | Positive Sample | Negative Sample | Sum |
|---|---|---|---|
| 2023 US | 706 | 969 | 1675 |
| 2022 US | 723 | 968 | 1691 |
| 2023 Global | 1774 | 1689 | 3463 |
| 2022 Global | 1803 | 1655 | 3458 |
| 2021 Global | 1787 | 1629 | 3416 |
| Total | 6793 | 6910 | 13703 |
Input sample dimension (H, W, C): (256, 256, 13) Mask label dimension (H, W, C): (256, 256, 1) Dataset resolution: 10 meters / pixel Dataset spectral names: [B1, B2, B3, B4, B5, B6, B7, B8, B8A, B9, B10, B11, B12] Dataset spectral detail information: Sentinel-2 Bands Info Label classes: {0: "background", 1: "solar_panel"}
Examples from our GloSoFarID dataset. Each row represents one sample with all 13 bands along with the ground truth mask where the gray color indicates the solar farm area.
We trained the FCN, Half-UNet and UNet models on this dataset for 50 epochs with a batch size of 16. The following presents the outcomes:
| Model | IoU | FScore |
|---|---|---|
| FCN | 71.81 | 82.87 |
| Half-UNet | 71.48 | 82.17 |
| UNet | 79.31 | 87.80 |
For TensorFlow users, you can load the dataset using import load_solar_dataset from dataset_converter, and execute tf_data = load_solar_dataset() to obtain a dataset object with the tf.data.Dataset class.
If you are working with frameworks other than TensorFlow, you may wish to convert the dataset into images for use as input data. By running the command python dataset_converter.py, the original dataset can be transformed into multispectral .tif images and .png mask annotations.
The dataset_converter supports the following optional arguments:
- --path, type=str, default=".", help='path to the dataset folder'
- --file_name, type=str, default=None, help='name of a specific tfrecord.gz file to convert, if not specified, convert all dataset files in the path'
- --include_jpg, type=bool, default=False, help='include jpg previews when converting'
In order to execute dataset_converter.py, ensure that the following libraries are installed in your environment:
pip install tensorflow pillow numpy tifffile
After converting the dataset, the .tif files will take about 24 GB space. Please make sure you have enough space on your disk before runing convertion