rasteret 0.3.12

Index-first GeoTIFF access layer for ML and analysis, powered by queryable Parquet indexes.

🛰️ Rasteret

The AI practitioner's multiplier for cloud-native satellite data.
A high-performance rasterio/GDAL alternative for scaleable ML workflows.

Rasteret helps you manage and read massive satellite imagery collections with zero friction.
It provides a 20x faster "drop-in" backend for TorchGeo, interops with all Arrow-based tools like DuckDB, Polars, while maintaining xarray, and NumPy support.

---

Why Rasteret?

Geospatial data science is often 80% "plumbing." You spend hours writing STAC reading loops, manual ThreadPoolExecutor, and fragile CRS-alignment logic just to get a batch of pixels for your model.

Rasteret turns those 80% into a single line of code.

It separates the Control Plane (managing your scenes, labels, and splits in a local Parquet index) from the Data Plane (streaming pixels directly from cloud COGs).

The "Friction" vs. "Flow" Comparison

The Old Way (25+ lines of fragile plumbing):

1. Search STAC catalog
2. Loop over items
3. Handle pagination
4. Filter by cloud cover
5. Wait 500ms per file to parse remote TIFF headers (GDAL cold start)
6. Manage ThreadPoolExecutor manually
7. Manually stack results and align CRS

The Rasteret Way (3 lines of robust code):

import rasteret

# 1. Load or Build your collection
collection = rasteret.load("my_s2_experiment")

# 2. Query like a Table: "Give me the training scenes with <10% clouds"
filtered = collection.subset(split="train", cloud_cover_lt=10)
# OR send Collection to DuckDB or Polars for enriching with your own data and bring it back to Rasteret
filtered = duckdb.sql("""
SELECT
    c.*,
    p.plot_id,
    p.is_target
FROM collection c
JOIN plots p ON ST_Intersects(c.geometry, p.geometry)
WHERE p.is_target = true AND c.cloud_cover_lt < 10
""")
filtered = rasteret.as_collection(filtered)

# 3. Batch Read: "Fetch aligned pixels for all geometries in the filtered collection"
data = filtered.get_numpy(geometries=filtered.geometry, bands=["B04", "B08"])

---

Key Features

• 🚀 20x Faster Cold Starts: By caching tile-layout metadata locally, Rasteret jumps straight to the pixels, skipping expensive remote header parsing, which happens in every new environment.
• 📦 Seamless "Drop-in" Backends: Boost TorchGeo or xarray performance by simply swapping the reader. No need to rewrite your analysis code.
• 🧬 Relational Imagery: Store your labels, train/val/test splits, and custom metadata directly in the imagery index. No more separate CSVs.
• 🛠️ Zero-Config Throughput: Automatic cloud storage presigning with Obstore, and custom async I/O handles the networking so you don't have to.

Performance

Rasteret's claims are backed by rigorous, reproducible benchmarks. We measure across three dimensions: cold-start latency, cloud-native scale, and comparison against legacy "data-inside-parquet" patterns.

1. Cold-start comparison with TorchGeo

Same AOIs, same scenes, same sampler, same DataLoader. Rasteret eliminates the "cold start tax" by caching IFD headers in the local Parquet index.

Scenario	rasterio/GDAL (Standard)	Rasteret (Index-First)	Speedup
Single AOI, 15 scenes	9.08 s	1.14 s	8x
Multi-AOI, 30 scenes	42.05 s	2.25 s	19x
Cross-CRS boundary	12.47 s	0.59 s	21x

2. The Cloud vs. Edge Comparison

How does Rasteret stack up against Google Earth Engine (GEE) or a highly parallelized Rasterio setup for time-series extraction?

Library	First Run (Cold)	Subsequent Runs (Hot)
Rasterio + ThreadPool	32 s	24 s
Google Earth Engine	10–30 s	3–5 s
Rasteret	3 s	3 s

3. HuggingFace MajorTOM vs. Rasteret

Recent "images-inside-Parquet" approaches (like MajorTOM) try to store image bytes in Parquet files. Rasteret keeps imagery in cloud COGs while using Parquet as a high-performance index—delivering better throughput without the data movement overhead.

Patches	HF datasets (streaming)	Rasteret index+COGs	Speedup
120	46.83 s	12.09 s	3.88x
1000	771.59 s	118.69 s	6.50x

All numbers measured on AWS us-west-2 4CPU machine (same region as data) vs. cold-start GDAL.

---

Technical Deep Dives

For the full architectural rationale, methodology, and reproducibility scripts, see:

• Full Benchmarks Guide: Methodology and results.
• Design Decisions: Why we chose Parquet + COGs
• Schema Contract: The internal anatomy of a Collection.

STAC API / GeoParquet  -->  Parquet Collection  -->  Tile-level byte reads
       (once)                  (queryable)             (no GDAL hot path)

Quick Start

1. Build a Collection

import rasteret

# Build from any STAC API or Parquet Metadata table
collection = rasteret.build(
    "earthsearch/sentinel-2-l2a",
    name="s2_training",
    bbox=(77.5, 12.9, 77.7, 13.1),
    date_range=("2024-01-01", "2024-06-30")
)

2. Turbocharge your ML (TorchGeo)

Rasteret provides a high-performance backend that honors the GeoDataset contract.

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

# Same API as TorchGeo, much faster pixel pipe
dataset = collection.to_torchgeo_dataset(bands=["B04", "B08"], chip_size=256)

sampler = RandomGeoSampler(dataset, size=256, length=100)
loader  = DataLoader(dataset, sampler=sampler, batch_size=4)

3. Fast Xarray creation

ds = collection.get_xarray(geometries=my_aoi, bands=["B04", "B08"])
ndvi = (ds.B08 - ds.B04) / (ds.B08 + ds.B04)

Key Entry Points

Rasteret is built for flexibility. Choose the output format that fits your existing workflow:

Method	Output	Purpose
to_torchgeo_dataset()	RasteretGeoDataset	Drop-in high-performance backend for TorchGeo training.
get_xarray()	xarray.Dataset	Quick create Xarray for analysis.
get_numpy()	numpy.ndarray	Raw pixel arrays ([N, C, H, W]) directly.
get_gdf()	GeoDataFrame	Metadata and pixel arrays as a standard geopandas dataframe.
sample_points()	DataFrame	Exact pixel values at points geometries with intuitive configurable fallback for nodata pixels

---

Full documentation at terrafloww.github.io/rasteret:

• Concepts: Why Rasteret?
• Migrating from Rasterio: Side-by-side patterns.
• TorchGeo Integration: Use Rasteret collections with TorchGeo.
• Tutorials: Hands-on examples.

License

Code: Apache-2.0