The Satellite dataset encompasses Point of Interest (POI) data, each associated with one or more Areas of Interest (AOI).
Every AOI incorporates a unique location identifier, a land type, and 16 low-resolution, 13-channel images (13 spectral bands, up to 10 m/pixel), each taken during a satellite visit to the location.
The Satellite dataset encompasses four land types as labels, namely Amnesty POI (4.8%), ASMSpotter (8.9%), Landcover (61.3%), and UNHCR (25.0%), making the task a 4-class classification problem of 3927 locations.
See Satellite.ipynb for more details.
-
Download the
satellite.zipfile from here -
Verify the file integrity using the following command:
(Make sure that the result matches the value provided in
satellite.zip.md5sum.)
md5sum satellite.zip #
- Unzip the
satellite.zip, If the file integrity check is successful, the zip file can be extracted using any standard unzipping tool. After extraction, the dataset's file structure will be as follows:
.
├── satellite_party0_test.csv
├── satellite_party0_test.pkl
├── satellite_party0_train.csv
├── satellite_party0_train.pkl
├── satellite_party1_test.csv
├── ...
└── satellite_party15_train.pklThe dataset is organized into 16 parties, each containing test and training files in CSV formats.
The naming convention follows the pattern satellite_partyX_test.csv and satellite_partyX_train.csv, where X ranges from 0 to 15.
Each CSV file consists of a header and the following columns (total 324,534 columns):
id, y, x0, x1, ..., x324531
For example the satellite_party0_test.csv
You can use the CSV data directly, or convert it to an image, column x0 to x324531 represents the 13-channel 158x158 pixel image.
the pixel values of each image are scaled to integer values within the range of [0, 255]
324532 = 13 x 158 x 158
See Satellite.ipynb for more details.
- algorithm:
SplitResNet(from torchvision.models import resnet18) - #classes:
4(four land types) - #parties:
16 - #epochs:
50 - metric:
accuracy - learning_rate:
1e-5 - batch_size:
32 - loss_function:
Cross-Entropy - early_stopping:
None - channel:
13 - kernel_size :
9 - out_activation:
None - agg_hidden:
[1000, out_dim] - out_dim:
4(same as #classes) - optimizer:
Adam - weight_decay:
1e-5 - lr_scheduler:
StepLR(step_size=10, gamma=0.5)
class SplitResNet(nn.Module):
def __init__(self, n_parties, channels, kernel_size=9, agg_hidden=None, out_activation=None):
super().__init__()
self.n_parties = n_parties
self.out_activation = out_activation
self.local_resnet_list = nn.ModuleList()
local_output_dims = []
for i in range(self.n_parties):
resnet = resnet18(weights=None)
local_output_dims.append(resnet.fc.in_features)
resnet.fc = nn.Identity()
resnet.conv1 = nn.Conv2d(channels, 64, kernel_size, stride=2, padding=3, bias=False)
self.local_resnet_list.append(resnet)
print("local output dims", local_output_dims)
self.cut_dim = sum(local_output_dims)
if agg_hidden is None:
self.agg_hidden = [100, 1]
else:
self.agg_hidden = agg_hidden
self.agg_mlp = MLP(self.cut_dim, self.agg_hidden)
if out_activation is None:
self.out_activation = nn.Identity()
else:
self.out_activation = out_activation
def forward(self, Xs):
# print("Shape of Xs: ", [Xi.shape for Xi in Xs])
local_outputs = [resnet(Xi) for resnet, Xi in zip(self.local_resnet_list, Xs)]
agg_input = torch.cat(local_outputs, dim=1)
agg_output = self.agg_mlp(agg_input)
return self.out_activation(agg_output)The VertiBench Satellite Dataset is licensed under the Creative Commons Attribution 4.0 International License. You are free to share and adapt the material for any purpose, even commercially, as long as you provide appropriate credit, link to the license, and indicate if changes were made.
Please refer to the full license text for more detailed information.