Multi-Dataset Training

Rasteret Collections produce standard TorchGeo GeoDataset objects, so TorchGeo's dataset composition operators (&, |) work directly.

---

Combining datasets with & (intersection)

Use & when you need aligned data from two sources, e.g. Sentinel-2 imagery paired with AEF embeddings, or imagery paired with a label mask. The sampler only draws chips where both datasets have spatial and temporal coverage.

from torch.utils.data import DataLoader
from torchgeo.samplers import RandomGeoSampler
from torchgeo.datasets.utils import stack_samples

s2 = s2_collection.to_torchgeo_dataset(bands=["B04", "B03", "B02"], chip_size=256)
aef = aef_collection.to_torchgeo_dataset(bands=["emb_0", "emb_1"], chip_size=256)

combined = s2 & aef  # IntersectionDataset: channels concatenated

sampler = RandomGeoSampler(combined, size=256, length=100)
loader = DataLoader(combined, sampler=sampler, batch_size=4,
                    collate_fn=stack_samples)

for batch in loader:
    # batch["image"] shape: [B, C_s2 + C_aef, H, W]
    ...

---

Combining datasets with | (union)

Use | when you want to train on whichever source has coverage, e.g. fall back to Landsat when Sentinel-2 is cloudy, or mosaic adjacent coverage from different providers.

ds1 = col1.to_torchgeo_dataset(bands=["B04", "B03", "B02"], chip_size=256)
ds2 = col2.to_torchgeo_dataset(bands=["B04", "B03", "B02"], chip_size=256)

combined = ds1 | ds2  # UnionDataset

---

How alignment works

• CRS and resolution are aligned to the first dataset at init time
• & computes spatial intersection via GeoPandas overlay; only areas where both datasets exist are valid for sampling
• | concatenates indices; the sampler can draw from either dataset's coverage area
• Band tensors are concatenated (for &) or merged (for |) automatically

---

xarray path

For analysis workflows, load each Collection separately and combine with standard xarray:

import xarray as xr

ds_s2 = s2_collection.get_xarray(geometries=aoi, bands=["B04", "B08"])
ds_aef = aef_collection.get_xarray(geometries=aoi, bands=["emb_0"])
combined = xr.merge([ds_s2, ds_aef])

Both must be in the same CRS. Use target_crs= if they differ.