GeoBench Substation

GeoBench Substation#

The substation dataset is a multi-spectral S2 dataset for instance segmentation power substation globally. Originally multi-temporal, as part of GEO-Bench-2 we created cloud free composites and subsampled to deduce computational complexity.

Reference

  • Jindgar, K., & Lindsay, G. W. (2024). Improving satellite imagery segmentation using multiple Sentinel-2 revisits. arXiv preprint arXiv:2409.17363.

[ ]:

from pathlib import Path

from geobench_v2.datamodules import GeoBenchSubstationDataModule
from geobench_v2.datasets import GeoBenchSubstation
from geobench_v2.datasets.visualization_util import (
    compute_batch_histograms,
    plot_batch_histograms,
)

%load_ext autoreload
%autoreload 2

The autoreload extension is already loaded. To reload it, use:
  %reload_ext autoreload

[18]:

PROJECT_ROOT = Path("../../")

[19]:

band_order = GeoBenchSubstation.band_default_order

datamodule = GeoBenchSubstationDataModule(
    img_size=256,
    batch_size=4,
    num_workers=4,
    root=PROJECT_ROOT,
    band_order=band_order,
    download=True,  # we do custom normalization in the tutorial
)
datamodule.setup("fit")

Initializing normalizer from class: ZScoreNormalizer
Initializing normalizer from class: ZScoreNormalizer

[20]:

sample_dist_fig = datamodule.visualize_geospatial_distribution()
../_images/dataset_notebooks_substation_5_0.png

Raw Batch Statistics#

[21]:

# Get a batch of data from the dataloader
train_dataloader = datamodule.train_dataloader()
raw_batch = next(iter(train_dataloader))

raw_batch_stats = compute_batch_histograms(raw_batch, n_bins=100)


raw_figs = plot_batch_histograms(
    raw_batch_stats, band_order, title_suffix=" (Raw Data)"
)
raw_figs

[21]:

[<Figure size 1200x500 with 1 Axes>]
../_images/dataset_notebooks_substation_7_1.png

Visualize Batch Data#

[22]:

fig, batch = datamodule.visualize_batch()
../_images/dataset_notebooks_substation_9_0.png
Contents