Using the GeoZarr Converter

The GeoZarr converter provides tools to transform EOPF datasets into GeoZarr-spec 0.4 compliant format. This guide explains how to use the converter effectively.

Command Line Interface

The converter can be accessed via the eopf-geozarr command-line tool. Below are some common use cases:

Basic Conversion

Convert an EOPF dataset to GeoZarr format:

eopf-geozarr convert input.zarr output.zarr

S3 Output

Convert and save the output directly to an S3 bucket:

eopf-geozarr convert input.zarr s3://my-bucket/output.zarr

Parallel Processing

Enable parallel processing for large datasets using a Dask cluster:

eopf-geozarr convert input.zarr output.zarr --dask-cluster

Validation

Validate the GeoZarr compliance of a dataset:

eopf-geozarr validate output.zarr

Python API

The converter also provides a Python API for programmatic usage:

Example: Basic Conversion

# test: skip
import xarray as xr
from eopf_geozarr import create_geozarr_dataset

# Load your EOPF DataTree
dt = xr.open_datatree("path/to/eopf/dataset.zarr", engine="zarr")

# Convert to GeoZarr format
dt_geozarr = create_geozarr_dataset(
    dt_input=dt,
    groups=["/measurements/r10m", "/measurements/r20m", "/measurements/r60m"],
    output_path="path/to/output/geozarr.zarr",
    spatial_chunk=4096,
    min_dimension=256,
    max_retries=3
)

Example: S3 Output

import os
from eopf_geozarr import create_geozarr_dataset

# Configure S3 credentials
os.environ['AWS_ACCESS_KEY_ID'] = 'your_access_key'
os.environ['AWS_SECRET_ACCESS_KEY'] = 'your_secret_key'
os.environ['AWS_DEFAULT_REGION'] = 'us-east-1'

# Convert and save to S3
dt_geozarr = create_geozarr_dataset(
    dt_input=dt,
    groups=["/measurements/r10m", "/measurements/r20m", "/measurements/r60m"],
    output_path="s3://my-bucket/output.zarr",
    spatial_chunk=4096,
    min_dimension=256,
    max_retries=3
)

Advanced Features

Chunk Alignment

The converter ensures proper chunk alignment to optimize storage and prevent data corruption. It uses the calculate_aligned_chunk_size function to determine optimal chunk sizes.

Multiscale Support

The converter writes the native resolution arrays at the group root and adds factor-of-two overviews as sibling subgroups named r{2**level} (r2, r4, r8, ...). Each overview is a complete dataset with its own coordinates and spatial: / proj: attributes; the parent group's multiscales metadata records each level via asset, derived_from, and transform.

Native CRS Preservation

The converter maintains the native coordinate reference system (CRS) of the dataset, avoiding reprojection to Web Mercator.

Sentinel-2 Optimized Conversion

The Sentinel-2 optimized converter (convert-s2-optimized / convert_s2_optimized) builds an efficient multiscale pyramid by reusing the original multi-resolution data (r10m, r20m, r60m) without duplication, and adding coarser overview levels (r120m, r360m, r720m) for visualization at lower resolutions.

The general convert command emits the same flat r{N} sibling layout for any input, but only the S2-optimized command takes advantage of S2's native multi-resolution structure.

Layout

Both converters produce a flat pyramid where each level is a sibling group:

output.zarr/
└── measurements/
    └── reflectance/
        ├── r10m/           # Native 10m data (reused as-is)
        ├── r20m/           # Native 20m data (reused as-is)
        ├── r60m/           # Native 60m data (reused as-is)
        ├── r120m/          # Computed from r60m (2x downsampling)
        ├── r360m/          # Computed from r120m (3x downsampling)
        └── r720m/          # Computed from r360m (2x downsampling)

Why these specific resolution levels?

The resolution levels are chosen to balance data preservation with storage optimization:

• Native ESA resolutions (10m, 20m, 60m): These are the original resolutions delivered by ESA for Sentinel-2 data and are reused as-is to preserve the source data without any loss
• Computed overview levels (120m, 360m, 720m): These additional levels were specifically chosen because their downsampling factors allow the data to be chunked and sharded in complete pieces, ensuring:
• 120m (2x from 60m): Standard doubling for the first computed overview
• 360m (3x from 120m): Selected for optimal chunking alignment
• 720m (2x from 360m): Final level for global-scale visualization

This approach maintains the integrity of ESA's original multi-resolution data while adding computationally efficient overview levels for performance at coarser scales.

Benefits: - No data duplication — native resolutions are reused directly - Efficient storage - Simple, flat hierarchy - Natural fit for Sentinel-2's multi-resolution data model

Key Capabilities

• Smart Resolution Consolidation: Combines Sentinel-2's native multi-resolution structure (10m, 20m, 60m) into a unified multiscale pyramid
• Non-Duplicative Downsampling: Reuses original resolution data instead of recreating it, adding only the coarser levels (120m, 360m, 720m)
• Variable-Aware Processing: Applies appropriate resampling methods for different data types (reflectance, classification, quality masks, probabilities)
• Efficient Testing: Improved test infrastructure for faster local development

Usage Example

from eopf_geozarr.s2_optimization.s2_converter import convert_s2_optimized
import xarray as xr

# Load Sentinel-2 DataTree
dt_input = xr.open_datatree("path/to/s2/product.zarr", engine="zarr")

# Convert to optimized multiscale structure
dt_optimized = convert_s2_optimized(
    dt_input=dt_input,
    output_path="path/to/output/optimized.zarr",
    enable_sharding=True,
    spatial_chunk=256,
    compression_level=3,
    validate_output=True
)

The result is a space-efficient multiscale pyramid: /measurements/reflectance/{r10m, r20m, r60m, r120m, r360m, r720m} where the native resolutions are preserved as-is and only the coarser levels are computed.

Error Handling

The converter includes robust error handling and retry logic for network operations, ensuring reliable processing even in challenging environments.

For more details, refer to the API Reference.