asterra

Support-aware machine learning for Earth observation.

Motivation

Earth observation (EO) machine learning workflows routinely mix data sources and label types with mismatched spatial supports:

• sensors with different pixel sizes (e.g., Sentinel-2 vs PlanetScope)
• labels defined on parcels/fields, tiles, scenes, or time windows (not per-pixel)
• coarse-to-fine (and fine-to-coarse) supervision
• patch overlap leakage and neighborhood dependence in evaluation

Ignoring these mismatches often leads to:

• biased features/labels due to incorrect aggregation
• silent leakage (overlapping patches, neighboring pixels, same-tile same-date)
• metrics that do not correspond to the true label support

What Asterra does

Asterra is a NumPy-first, scikit-learn-compatible package for building support-aware pipelines.

The core abstraction is a sparse, overlap-based SupportMatrix that maps one support to another. It powers:

• mixed-resolution aggregation and projection (grid ↔ grid, samples → groups)
• support-aware feature and label projection
• leakage-safe splitting utilities (buffers, tile/time grouping)
• support-aware metrics

Installation

# Once published to PyPI:
# python -m pip install asterra

# Install from source (GitHub):
python -m pip install "asterra @ git+https://github.com/ArnaBannonymus/asterra.git@main"

Optional geospatial extras (not required for pixel-space workflows):

python -m pip install "asterra[geo] @ git+https://github.com/ArnaBannonymus/asterra.git@main"

Quickstart

import numpy as np
from asterra.data import EOData
from asterra.io import sensors
from asterra.support import SupportMatrix

# Synthetic Sentinel-2-like grid (H, W, B)
arr_s2 = np.random.RandomState(0).randn(32, 32, 4).astype("float32")
e_s2 = EOData.from_array(
    arr_s2,
    band_schema=sensors.sentinel2_rgbn(),
    support={"kind": "grid", "resolution": (10.0, 10.0), "origin": (0.0, 0.0)},
)

# Synthetic PlanetScope-like grid on a different resolution
arr_ps = np.random.RandomState(1).randn(64, 64, 4).astype("float32")
e_ps = EOData.from_array(
    arr_ps,
    band_schema=sensors.planetscope_4band(),
    support={"kind": "grid", "resolution": (5.0, 5.0), "origin": (0.0, 0.0)},
)

# Map PlanetScope pixels (source) onto Sentinel-2 pixels (target)
M = SupportMatrix.from_grid_to_grid(source=e_ps.support, target=e_s2.support)
X_ps_on_s2 = M.project_features(e_ps.as_samples())
print(X_ps_on_s2.shape)  # (32*32, 4)

Visual proofs (local datasets)

Asterra is NumPy-first and does not ship heavy geospatial file I/O. For GeoTIFF/SAFE/NetCDF products you typically read data with tools like rasterio/xarray and then construct EOData with a BandSchema and SupportSpec.

The figures below are generated from local datasets (not included in this repo) to sanity-check the support-aware operators on non-synthetic inputs:

Planet (3m PF-SR) NDVI → Sentinel-2 (10m) NDVI window (SupportMatrix overlap projection + sparse structure view)

Sentinel-1 VV/VH (dB) window + label map (leakage-aware spatial CV demo data)

complex SAR patch example (HH/HV magnitudes)

To regenerate these visuals on your machine (with your own file paths):

python scripts/generate_readme_visuals.py \
  --planet-pf-sr /path/to/planet_pf_sr.tif \
  --s2-lr-ndvi /path/to/s2_lr_ndvi.tif \
  --s1-vv /path/to/s1_vv.tif \
  --s1-vh /path/to/s1_vh.tif \
  --label-map /path/to/label_map.tif \
  --cvdl-city-dir /path/to/S1SLC_CVDL/City

Supported inputs

• .npy arrays with shapes (H, W, B), (T, H, W, B), (N, B)
• generic EO arrays with user-provided metadata:
  • band_names
  • georeferencing (resolution/origin or affine transform/crs) when available
  • pixel-space coordinates when georeferencing is not available
  • explicit group identifiers for parcel/tile/time supports

Built-in sensor presets

Sensor helpers are convenience presets; the core library is sensor-agnostic.

• Sentinel-2 (common optical bands)
• Sentinel-1 (VV/VH-style SAR schema)
• PlanetScope (4-band and 8-band styles)
• NISAR-style configurable SAR schemas

Architecture

The project is organized to keep EO-specific functionality separate from potentially generic sparse support logic:

• asterra.data: EO data model (array + band schema + support metadata)
• asterra.support: sparse support operators (SupportMatrix, projection)
• asterra.preprocessing: reshape/masking and band-aware transformers
• asterra.model_selection: leakage-aware splitters/utilities
• asterra.metrics: support-aware metrics
• asterra.io: .npy loader + sensor presets

See DESIGN_BOUNDARIES.md and UPSTREAMING.md for boundary notes and candidate generic components.

Release status

0.1.x is an early, focused release line. The API is intentionally narrow and may evolve based on user feedback and scientific validation.

Roadmap (high level)

• richer support specifications (polygons/parcels via optional geo extras)
• additional support-aware scorers and splitters
• integration examples with local EO stacks (while keeping the core sensor-agnostic)