Sampler¶

HiRISE strips are thin rotated parallelograms (a few km wide, tens of km long), so naive bounding-box sampling wastes 60–90 % of patches on empty pixels. HiRISEGeoSampler pre-computes a grid of patch centres that actually intersect each strip's convex-hull polygon and samples uniformly from that set each epoch.

Geometry idea¶

flowchart LR
    A["Strip polygon
(convex hull)"] --> B["Erode by patch_size/2"]
    B --> C["Valid-center region"]
    C --> D["Pack patch centres
(grid or optimal)"]
    D --> E["Sample uniformly
each epoch"]

The "optimal" mode (in src/dataset/sampling/geometry.py) packs centres to maximize non-overlapping coverage; "simple" mode uses a regular grid.

Geographic splits¶

The sampler supports train / val / test splits via geographic partitioning:

• split_method="geographic" — split by split_axis ("longitude" or "latitude").
• split_fractions=(0.8, 0.1, 0.1) — fractions of strips (not patches) per split.
• K-fold cross-validation is also supported via n_folds / fold_idx.

This prevents leakage from a strip's patches appearing in more than one split.

Usage¶

from dataset import HiRISEGeoSampler
from torchgeo.samplers import Units

sampler = HiRISEGeoSampler(
    dataset,
    size=0.018,                  # in CRS units
    length=None,                 # use the full prebuilt set
    units=Units.CRS,
    split_fractions=(0.8, 0.1, 0.1),
    split_method="geographic",
    split_axis="longitude",
    split="train",
)

See dataset.sampling.sampler and dataset.sampling.geometry.