Cutana - Astronomical Cutout Pipeline

Cutana is a high-performance Python pipeline for creating astronomical image cutouts from large FITS datasets. It provides both an interactive Jupyter-based UI and a programmatic API for efficient processing of astronomical survey data like ESA Euclid observations.

Documentation: https://esa.github.io/Cutana/

Note: Cutana is currently optimised for Euclid Q1/IDR1 data. Some defaults and assumptions are Euclid-specific:

Flux conversion expects the MAGZERO header keyword (configurable via config.flux_conversion_keywords.AB_zeropoint) to convert to Jy

Filter detection patterns are tuned for Euclid bands (VIS, NIR-Y, NIR-H, NIR-J)

FITS structure assumes one file per channel/filter

For other surveys, you may need to adjust these settings or disable flux conversion (config.apply_flux_conversion = False).

Support for Datalabs Users

For users experiencing problems with Cutana in the ESA Datalabs environment, please open a service desk ticket at: https://support.cosmos.esa.int/situ-service-desk/servicedesk/customer/portal/5

Quick Start

Installation

pip install cutana

Or for development:

# Create conda environment
conda env create -f environment.yml
conda activate cutana

Interactive UI (Recommended)

For most users, the interactive interface provides the easiest way to process astronomical data:

import cutana_ui

cutana_ui.start()  # optionally can specify e.g. ui_scale=0.6 for smaller UI

This launches a step-by-step interface where you can:

Select your source catalogue (CSV format)
Configure processing parameters (image extensions, output format, resolution)
Monitor progress with live previews and status updates

Programmatic API

For automated workflows or integration into larger systems:

import pandas as pd
from cutana import get_default_config, Orchestrator

# Configure processing
config = get_default_config()
config.source_catalogue = "sources.csv"  # See format below
config.output_dir = "cutouts_output/"
config.output_format = "zarr"  # or "fits"
config.target_resolution = 256
config.selected_extensions = [{"name": "VIS", "ext": "PrimaryHDU"}]  # Extensions to process
# 1 output channel for VIS, details explained below
config.channel_weights = {
    "VIS": [1.0],
}
config.console_log_level = "INFO"  # Show INFO logs in console

# Process cutouts
orchestrator = Orchestrator(config)
results = orchestrator.run()

Input Data Format

Your source catalogue must be a CSV/FITS/PARQUET file containing these columns:

SourceID,RA,Dec,diameter_pixel,fits_file_paths
TILE_102018666_12345,45.123,12.456,128,"['/path/to/tile_vis.fits', '/path/to/tile_nir.fits']"
TILE_102018666_12346,45.124,12.457,256,"['/path/to/tile_vis.fits','/path/to/tile_nir.fits']"

Required Columns:

SourceID: Unique identifier for each astronomical object. SourceIDs must be unique across all rows — cutout extraction is keyed by SourceID, so rows sharing one collapse into a single cutout. Cutana detects duplicates in catalogues with fewer than 100,000 sources and reformats IDs as SourceID_RA_Dec; rows that duplicate both the ID and the position cannot be told apart and are rejected with CatalogueValidationError (deduplicate first, e.g. df.drop_duplicates(subset=['SourceID', 'RA', 'Dec'])). For larger catalogues the check is skipped for performance reasons — Cutana never scans a whole catalogue, so ensuring uniqueness is your responsibility. Note that streaming processes small internal batches, so the check still applies within each batch.
RA: Right Ascension in degrees (0-360°, ICRS coordinate system)
Dec: Declination in degrees (-90 to +90°, ICRS coordinate system)
diameter_pixel: Cutout size in pixels (creates square cutouts). Alternatively diameter_arcsec.
fits_file_paths: JSON-formatted list of FITS file paths containing the source, use consistent order

Note

Cutana checks that each source actually falls inside the FITS products its row names. A source whose centre lies outside a file it names is rejected with CatalogueValidationError — without the check it would silently yield a full-size cutout made entirely of edge padding. A centre inside the image with the cutout box crossing an edge is only a warning, since a partial cutout is usually what you want at a survey boundary. Large catalogues are sampled rather than checked row by row (1000 rows, at most 100 distinct files opened), and what was skipped is logged.

Output Formats

ZARR Format (recommended): All cutouts stored in a efficient archives, ideal for large datasets and analysis workflows. Cutana uses the Zarr format for high-performance storage and the images_to_zarr library for conversion. (See the Output section below for sample code to access)

FITS Format: Individual FITS files per source, best for compatibility with existing astronomical software. Mandatory format for do_only_cutout_extraction (Raw cutouts), which skips all processing aside from the flux conversion, which can be disabled.

Flux Conversion

To align cutout units, pixels are converted to Janskys by default using the MAGZERO header keyword (configurable via config.flux_conversion_keywords.AB_zeropoint). Disable this flux conversion via config.apply_flux_conversion = False.

Pixel Units in FITS Headers

FITS cutouts record their pixel unit in these keywords:

Keyword	HDU	Meaning
`BUNIT`	image	Standard, machine-readable unit. Written only when cutana knows the unit, which means `Jy`; absent in every other case
`FLUXAPPX`	image	`T` when the resize did not conserve flux, so the values only approximate `BUNIT`. Written only alongside `BUNIT`
`UNIT`	primary	Deprecated 0.3.2 keyword, kept for existing readers. Prefer `BUNIT`
`CONSVFLX`	primary	Whether flux-conserved resizing was used

UNIT takes one of these values, prefixed approx when the resize did not conserve flux:

`UNIT`	When	`BUNIT`
`Jy`	Flux conversion on, pixels still on their physical scale	`Jy`
`OriginalUnit`	Flux conversion off, so pixels keep the parent tile's unit — which cutana cannot name without reading it back from the parent header	absent
`UserConversionUnit`	`config.user_flux_conversion_function` replaced the AB-zeropoint maths, so the pixels were converted, just not by cutana and not necessarily to Jy. Only you know the resulting unit	absent
`normalised`	Normalisation discarded the physical scale — the values are dimensionless	absent

Important: normalisation stretches pixels into the data_type range and discards the physical scale, so a normalised cutout is dimensionless regardless of apply_flux_conversion. To keep pixels in Jy you need both:

config.normalisation_method = "none"  # or config.do_only_cutout_extraction = True
config.data_type = "float32"  # uint8 rescales to 0-255, losing the scale

Note that BUNIT is written when the resize did not conserve flux: the unit is still Jy, and FLUXAPPX = T carries the caveat that the values only approximate it.

WCS (World Coordinate System) Handling

FITS Output

FITS cutouts preserve full WCS information with accurate astrometric calibration:

Automatic pixel scale correction: When cutouts are resized (e.g., from extracted size to target_resolution), the WCS pixel scale is automatically adjusted to maintain accurate coordinate transformations
Parent WCS reproduction: the cutout reproduces the parent tile mapping exactly — CRVAL, CTYPE and the CD/PC orientation are inherited from the parent unchanged, and only CRPIX is shifted to the extraction origin. When the parent tile carries no usable WCS, a minimal TAN header is synthesised instead, with CRPIX at the cutout centre and CRVAL at the source position
Format compatibility: Supports CD matrix, CDELT, and PC+CDELT WCS formats from original FITS files
Sky area preservation: Total sky coverage remains constant while pixel scale adjusts for resize operations

Pixel units are documented separately under Pixel Units in FITS Headers; they are not WCS keywords.

Zarr Output

Important: Zarr archives do not contain WCS information. The WCS data is not recorded in the image metadata stored within the Zarr files. The central point and image size (in pixels or arcseconds, depending on what is provided) is recorded.

Flux-Conserved Output

For scientific analysis requiring photometric accuracy, Cutana supports flux-conserving resizing:

config.flux_conserved_resizing = True
config.data_type = "float32"
config.normalisation_method = "none"

Important: When using flux-conserved resizing, you should:

Set data_type = "float32" to preserve numerical precision
Use normalisation_method = "none" to keep original flux values
This mode preserves total flux during the resizing operation, essential for photometric measurements

Note: Flux conservation uses drizzle, which may affect performance

Multi-Channel Processing

Cutana automatically handles sources with multiple FITS files and allows channel mixing through configurable weights.

Channel Weights Configuration

The channel_weights parameter controls how multiple FITS files are combined into output channels. Each key represents a FITS extension name, and the corresponding value is a list of weights for image channels respectively.

Use channel names as channel_weights keys; dictionary order does not matter. Weights are not normalised. The UI discovers channel order from a bounded catalogue sample. See channel mapping and sampling for matching rules and sampling limits.

# Map each input band to output channel weights
config.channel_weights = {
    "VIS": [1.0, 0.0, 0.5],  # RGB weights for VIS band
    "NIR-H": [0.0, 1.0, 0.3],  # RGB weights for NIR H-band
    "NIR-J": [0.0, 0.0, 0.8],  # RGB weights for NIR J-band
}

Example: If your source catalogue contains:

SourceID,RA,Dec,diameter_pixel,fits_file_paths
TILE_123_456,45.1,12.4,128,"['/path/to/vis.fits', '/path/to/nir_h.fits', '/path/to/nir_j.fits']"

The following names identify the input bands regardless of dictionary order:

/path/to/vis.fits → VIS extension
/path/to/nir-h.fits → NIR-H extension
/path/to/nir-j.fits → NIR-J extension

Image Normalisation

Cutana supports multiple stretch algorithms with unified parameter configuration:

config = get_default_config()

# Set normalisation method
config.normalisation_method = "asinh"  # "linear", "log", "asinh", "zscale"

# Configure normalisation parameters (method-specific defaults applied automatically)
config.normalisation.percentile = 99.8  # Data clipping percentile
config.normalisation.a = 0.7  # Transition parameter (asinh/log)
config.normalisation.n_samples = 1000  # ZScale samples
config.normalisation.contrast = 0.25  # ZScale contrast

Image stretching is powered by fitsbolt for consistent processing.

Output

In the case of zarr files the output will be organised in batches. Per batch one folder is created each with an images.zarr and an images_metadata.parquet. The folders are named using the format batch_cutout_process_{index}_{unique_id} where each process gets a unique identifier.

Setting do_only_cutout_extraction to True along with output_format fits allows cutouts to be directly created without normalisation/resizing or channel combination. This is currently incompatible with zarr output. The flux conversion will still be applied and the data_type determined by the input.

Metadata

With the output zarr files, a metadata parquet file is created containing the following information: source_id, ra, dec, diameter_arcsec, diameter_pixel, tile, processing_timestamp and the extraction geometry.

This can be read with:

import pandas as pd

metadata=pd.read_parquet("output_path/batch_cutout_process_*/images_metadata.parquet", engine='pyarrow')

Row i of this parquet describes image i of the .zarr file in the same folder, which is how you map Source IDs to cutouts. Note: No source IDs will be stored in the .zarr files!

Images

To open the .zarr files and look at example images, the following code can be used:

# Open the images
import zarr

# File: a dictionary that contains "images" of shape n_images,H,W,C
file = zarr.open("output_path/batch_cutout_process_*/images.zarr", mode='r')
print(list(file.keys()))

# Example plots
from matplotlib import pyplot as plt
import numpy as np

fig, axes = plt.subplots(4, 4, figsize=(8, 8))
for ax in axes.flatten():
    index=np.random.randint(0, file['images'].shape[0])
    image = file['images'][index]
    ax.imshow(image, cmap='gray',origin="lower")
    ax.axis('off')
plt.tight_layout()
plt.show()

If fits was selected as the output, then the fits files contain the info in the header of the PRIMARY extension.

To select a specific image from the .zarr files, the above output parquet file must be used as described in the Metadata section.

Cutout Extraction Behavior

Padding Factor

The padding_factor (in UI zoom-out) parameter controls the extraction area relative to the source size from your catalogue:

padding_factor = 1.0 (default): Extracts exactly the source size (diameter_pixel or diameter_arcsec)
padding_factor < 1.0 (zoom-in): Extracts a smaller area (source_size × padding_factor pixels)
- Minimum value: 0.25 (extracts 1/4 of source_size)
- Example: padding_factor = 0.5 with 10px source extracts 5×5 pixels
padding_factor > 1.0 (zoom-out): Extracts a larger area (source_size × padding_factor pixels)
- Maximum value: 10.0 (extracts 10× source_size)
- Example: padding_factor = 2.0 with 10px source extracts 20×20 pixels

All extracted cutouts are then resized to the final target_resolution (e.g., 256×256) in the processing pipeline.

Stretch/Normalisation Configuration

Cutana supports multiple image stretching methods with unified parameter naming for optimal visualisation and analysis. The normalisation_method parameter controls the stretch algorithm, with a single set of parameters that automatically apply appropriate defaults based on the chosen method:

Unified Parameter Details: All normalisation parameters are now stored in the config.normalisation DotMap object:

config.normalisation.percentile: Percentile for data clipping, applied to all stretch methods (default: 99.8, range: 0-100)
config.normalisation.a: Unified transition parameter with method-specific defaults:
- ASINH: 0.7 (controls linear-to-logarithmic transition, range: 0.001-3.0)
- Log: 1000.0 (scale factor for transition point, range: 0.01-10000.0)
config.normalisation.n_samples: Number of samples for ZScale algorithm (default: 1000, range: 100-10000)
config.normalisation.asinh_n_samples: Pixels per channel sampled to estimate the asinh percentile bounds (default: None, range: 100-1000000). None uses every pixel, which is exact and is what earlier releases did. Set it only to trade accuracy for speed — sampling biases the bright tail and changes output values
config.normalisation.contrast: Contrast adjustment for ZScale (default: 0.25, range: 0.01-1.0)

Note: The unified a parameter automatically applies the appropriate default value based on the selected normalisation method, eliminating the need for method-specific parameter names.

Linear Stretch

from cutana import get_default_config

config = get_default_config()
config.normalisation_method = "linear"
config.normalisation.percentile = 99.8  # Percentile clipping (default)

ASINH Stretch (Recommended)

from cutana import get_default_config

config = get_default_config()
config.normalisation_method = "asinh"
config.normalisation.percentile = 99.8  # Percentile clipping (default)
config.normalisation.a = 0.7  # Transition parameter (default for asinh)
# config.normalisation.asinh_n_samples = 10000  # optional: subsample for speed, not exact

Log Stretch

from cutana import get_default_config

config = get_default_config()
config.normalisation_method = "log"
config.normalisation.percentile = 99.8  # Percentile clipping (default)
config.normalisation.a = 1000.0  # Scale factor (default for log)

ZScale Stretch

from cutana import get_default_config

config = get_default_config()
config.normalisation_method = "zscale"
config.normalisation.percentile = 99.8  # Percentile clipping (default)
config.normalisation.n_samples = 1000  # Number of samples (default)
config.normalisation.contrast = 0.25  # Contrast parameter (default)

Performance Considerations

Memory and CPU Optimization

Cutana is designed to optimally utilize available system memory and CPU cores while maintaining stability and preventing system overload. The pipeline includes intelligent load balancing that automatically adjusts worker processes and memory allocation based on real-time system monitoring.

Memory Constraints

In most deployment scenarios, memory will be the limiting factor rather than CPU cores. Cutana's load balancer continuously monitors memory usage and dynamically adjusts the number of worker processes to prevent memory exhaustion.

Critical Usage Guidelines

IMPORTANT: Users SHALL NOT run other memory-intensive activities in parallel with Cutana processing.

Running competing memory-intensive processes will interfere with Cutana's load balancing algorithms and may lead to:

System crashes due to memory exhaustion
Degraded performance and increased processing time
Inconsistent or failed processing results
Potential data corruption in extreme cases

Optimal Usage

For best performance:

Dedicated Resources: Run Cutana on dedicated compute resources when possible
Monitor System: Use system monitoring tools to track memory usage during processing
Batch Size Tuning: Adjust N_batch_cutout_process based on available memory
Worker Configuration: Let the load balancer automatically determine optimal worker count, or manually set max_workers conservatively

The load balancer will automatically scale down processing if memory pressure increases, but prevention through proper resource management is always preferable.

Common Errors and Pitfalls

CatalogueValidationError: N catalogue rows are exact duplicates Cutout extraction is keyed by SourceID, so rows sharing an ID and a position cannot be told apart and would collapse into one cutout. Deduplicate first: df.drop_duplicates(subset=["SourceID", "RA", "Dec"]). Sources that merely share an ID at different positions are fine — they are separated automatically. Unique IDs remain the caller's responsibility; Cutana never scans a whole catalogue, so this check is skipped above 100,000 rows in a single-shot load.

Streaming ended after N of M expected batches get_batch_count() is derived from the catalogue row count, so the run expects one cutout per row. Fewer arrived. Call get_delivery_report() for how many are missing and which worker they were assigned to — the usual causes are sources whose cutout window falls outside its tile (no cutout is produced, and this is legitimate) and workers that died before delivering.

Worker <id> failed with exit code ... The message quotes the worker's own error; the full traceback is in the session log and in logs/subprocesses/<id>_stderr.log. Exit code -9 means the kernel OOM-killed the worker — lower N_batch_cutout_process or max_workers.

Streaming without cleanup() in a finally Workers block indefinitely waiting to hand over their next chunk, so leaving the loop early — including via an exception from next_batch() — strands them holding their shared memory pools. Always wrap init_streaming() and the batch loop in try: ... finally: orchestrator.cleanup().

Advanced features

To avoid bright objects at the border of an cutout making the target too faint, the user can set

int x # larger than 1, smaller than config.target_resolution
config.normalisation.crop_enable=True
config.normalisation.crop_height = x
config.normalisation.crop_width = x

Then after resizing, during normalisation the image is cropped to crop_height x crop_width around the center and the maximum value is determined inside this cropped region.

Configuration Parameters

The following table describes all configuration parameters available in Cutana:

Parameter	Type	Default	Range/Allowed Values	Description
General Settings
`name`	str	"cutana_run"	-	Run identifier
`log_level`	str	"INFO"	DEBUG, INFO, WARNING, ERROR, CRITICAL, TRACE	Logging level for files
`console_log_level`	str	"WARNING"	DEBUG, INFO, WARNING, ERROR, CRITICAL, TRACE	Console/notebook logging level
Input/Output Configuration
`source_catalogue`	str	None	File path	Path to source catalogue CSV file (required)
`output_dir`	str	"cutana_output"	Directory path	Output directory for results
`output_format`	str	"zarr"	zarr, fits	Output format
`data_type`	str	"float32"	float32, uint8	Output data type
`flux_conserved_resizing`	bool	False	-	Enable flux-conserving resizing (use with float32 + none normalisation, uses drizzle (slower))
Preprocessing Configuration
`skip_catalogue_validation`	bool	False	-	If True, catalogue validation is skipped during the preprocessing step
`skip_fits_check`	bool	False	-	If True, the validation checks that open FITS files are skipped: that each row's tiles exist and are readable, and that the source falls inside them. Much faster on a short run, at the cost of those two failures going undetected
Processing Configuration
`max_workers`	int	effective CPU count	1-1024	Maximum number of worker processes
`N_batch_cutout_process`	int	1000	10-10000	Batch size within each process
`max_workflow_time_seconds`	int	1354571	600-5000000	Maximum total workflow time (~2 weeks default)
Cutout Processing Parameters
`do_only_cutout_extraction`	bool	False	-	If True, must set "fits", no norm/resize/combination img
`target_resolution`	int	256	16-2048	Target cutout size in pixels (square cutouts)
`padding_factor`	float	1.0	0.25-10.0	Padding factor for cutout extraction (1.0 = no padding)
`normalisation_method`	str	"linear"	linear, log, asinh, zscale, none	Normalisation method, method must not be none for unit8 output
`interpolation`	str	"bilinear"	bilinear, nearest, bicubic, lanczos	Interpolation method
FITS File Handling
`fits_extensions`	list	["PRIMARY"]	List of str/int	Default FITS extensions to process
`selected_extensions`	list	[]	List of str/int/dict	Bands selected by user (set by UI); also narrows which files of a set are loaded — see Band selection
`available_extensions`	list	[]	List	Available extensions (discovered during analysis)
Flux Conversion Settings
`apply_flux_conversion`	bool	True	-	Whether to apply flux conversion (for Euclid data)
`flux_conversion_keywords.AB_zeropoint`	str	"MAGZERO"	-	Header keyword for AB magnitude zeropoint
`user_flux_conversion_function`	callable	None	-	Custom flux conversion function (deprecated)
Image Normalization Parameters
`normalisation.percentile`	float	99.8	0-100	Percentile for data clipping
`normalisation.a`	float	0.7 (asinh), 1000.0 (log)	0.001-10000.0	Unified transition parameter
`normalisation.n_samples`	int	1000	100-10000	Number of samples for ZScale algorithm
`normalisation.asinh_n_samples`	int	None	100-1000000	Pixels/channel sampled for asinh percentile bounds; None = exact
`normalisation.contrast`	float	0.25	0.01-1.0	Contrast adjustment for ZScale
`normalisation.crop_enable`	bool	False	-	Enable cropping during normalization
`normalisation.crop_width`	int	-	0-5000	Crop width in pixels
`normalisation.crop_height`	int	-	0-5000	Crop height in pixels
Advanced Processing Settings
`channel_weights`	dict	{"PRIMARY": [1.0]}	Dict of str: list[float]	Channel weights for multi-channel processing
`external_fitsbolt_cfg`	DotMap	None	FITSBolt config or None	External FITSBolt config for ML pipeline integration (overrides normalisation settings). Initialize this parameter using fitsbolt.create_config(), then modify the specific parameters by updating the returned dictionary.
File Management
`tracking_file`	str	"workflow_tracking.json"	-	Job tracking file
`config_file`	str	None	File path	Path to saved configuration file
Load Balancer Configuration
`loadbalancer.memory_safety_margin`	float	0.15	0.01-0.5	Safety margin for memory allocation (15%)
`loadbalancer.memory_poll_interval`	int	3	1-60	Poll memory every N seconds
`loadbalancer.memory_peak_window`	int	30	10-300	Track peak memory over N second windows
`loadbalancer.main_process_memory_reserve_gb`	float	4.0	0.5-10.0	Reserved memory for main process
`loadbalancer.initial_workers`	int	1	1-8	Start with N workers until memory usage is known
`loadbalancer.max_sources_per_process`	int	150000	1+	Maximum sources per job/process
`loadbalancer.log_interval`	int	30	5-300	Log memory estimates every N seconds
`loadbalancer.event_log_file`	str	None	File path	Optional file path for LoadBalancer event logging
`loadbalancer.skip_memory_calibration_wait`	bool	False	-	Skip waiting for first worker memory measurements on launch of cutout creation and proceed immediately with a static memory estimate
`process_threads`	int	None	1-128, None	Thread limit per process (None = auto: cores // 4)
UI Configuration
`ui.preview_samples`	int	10	1-50	Number of preview samples to generate
`ui.preview_size`	int	256	16-512	Size of preview cutouts
`ui.auto_regenerate_preview`	bool	True	-	Auto-regenerate preview on config change

Band selection

selected_extensions decides which files of a FITS set are loaded, on every backend. Names are matched against the band in each filename (VIS, NIR-H, …), so selecting three bands from a four-file Euclid set loads exactly those three, in catalogue order, and channel_weights needs one entry per selected band. Use "PRIMARY" to switch band selection off and load every file in the set.

A selection naming no band Cutana recognises narrows nothing, so every file in each set loads. On non-Euclid data that is the normal case: the UI labels a file the recogniser cannot classify UNKNOWN, and there is no band behind that label to select on. It is logged once per run, because the other way to reach it is a typo — "NIRH" narrows nothing just as quietly.

Warning

A selection that names real bands but matches no file in a set asks for data that set does not contain. create_cutouts_direct() raises ValueError. The orchestrator and streaming backends skip that set and log Skipping FITS set, so one tile that lacks the selected bands does not cost the tiles around it — grep the log for it before reading a short run as a complete one. If the selection matches no set, every backend fails: that is a misspelled or inapplicable selected_extensions, not a gap in the data, and the error names both the bands you asked for and the bands that are there.

Backend API Reference

Orchestrator Class

The Orchestrator class is the main entry point for programmatic access to Cutana's cutout processing capabilities.

Constructor

from cutana import Orchestrator
from cutana import get_default_config

orchestrator = Orchestrator(config, status_panel=None)

Parameters:

config (DotMap): Configuration object created with get_default_config()
status_panel (optional): UI status panel reference for direct updates

Main Processing Methods

`start_processing(catalogue_data)`

Start the main cutout processing workflow.

import pandas as pd
from cutana import Orchestrator, get_default_config

# Load your source catalogue
catalogue_df = pd.read_csv("sources.csv")

# Configure processing
config = get_default_config()
config.output_dir = "cutouts_output/"
config.output_format = "zarr"
config.target_resolution = 256
config.selected_extensions = [
    {"name": "VIS", "ext": "PrimaryHDU"},
    {"name": "NIR-H", "ext": "PrimaryHDU"},
    {"name": "NIR-J", "ext": "PrimaryHDU"},
]
config.channel_weights = {
    "VIS": [1.0, 0.0, 0.5],
    "NIR-H": [0.0, 1.0, 0.3],
    "NIR-J": [0.0, 0.0, 0.8],
}

# Process cutouts
orchestrator = Orchestrator(config)
result = orchestrator.start_processing(catalogue_df)

Parameters:

catalogue_data (pandas.DataFrame): DataFrame containing source catalogue with required columns

Returns:

dict: Result dictionary containing:
- status (str): "completed", "failed", or "stopped"
- total_sources (int): Number of sources processed
- completed_batches (int): Number of completed processing batches
- mapping_parquet (str): Path to source-to-zarr mapping *.parquet file
- error (str): Error message if status is "failed"

`run()`

Simplified entry point that loads catalogue from config and runs processing.

config = get_default_config()
config.source_catalogue = "sources.csv"
config.output_dir = "cutouts_output/"

orchestrator = Orchestrator(config)
result = orchestrator.run()

Returns:

dict: Same format as start_processing()

Direct Cutout Generation (Fast, Small Batches)

For small batches (< ~1000 sources), create_cutouts_direct() runs entirely in-process — ~3x faster than the full orchestrator by avoiding subprocess overhead.

import pandas as pd
from cutana import create_cutouts_direct, get_default_config

config = get_default_config()
config.target_resolution = 256
config.selected_extensions = [{"name": "VIS", "ext": "PrimaryHDU"}]
config.channel_weights = {"VIS": [1.0]}

catalogue_df = pd.read_csv("sources.csv")
results = create_cutouts_direct(catalogue_df, config)

for result in results:
    cutouts = result["cutouts"]  # ndarray (N, H, W, C)
    metadata = result["metadata"]  # list of per-source dicts

selected_extensions narrows which files of a FITS set are loaded here as it does on every other backend — see Band selection for what it matches and what happens when it matches nothing.

When the sources span many FITS tiles, pass max_workers to process tiles concurrently on a thread pool (each tile is loaded and closed in isolation). None (default) auto-selects min(n_tiles, effective_cpus, cap) using the k8s/cgroup-aware CPU count; a single tile always runs serially. The cap (8) is a conservative ceiling: the work is I/O bound, so beyond a handful of threads extra concurrency mostly oversubscribes a shared/networked filesystem rather than adding throughput. Pass an explicit max_workers to override it:

results = create_cutouts_direct(catalogue_df, config, max_workers=8)

Extract once, render many times

For an interactive view — a preview whose stretch or colour mix the user changes — extract the raw arrays once and combine and normalise them repeatedly. All three steps are public:

from cutana import apply_normalisation, combine_channels, create_cutouts_direct

config.do_only_cutout_extraction = True  # raw arrays: no resize, mix or stretch
results = create_cutouts_direct(catalogue_df, config)

raw = results[0]["cutouts"]  # (N, H, W, N_extensions)
names = results[0]["channel_names"]  # the extension order the weights bind to

config.channel_weights = {"VIS": [0.0, 0.0, 1.0], "NIR-H": [0.66, 0.0, 0.0]}
mixed = combine_channels(raw, config.channel_weights, names)  # -> (N, H, W, 3)
display = apply_normalisation(mixed, config)  # re-run per stretch change

Warning

Pass result["channel_names"] to combine_channels. Names are required to resolve weights; missing, duplicate, or ambiguous mappings raise an error.

When to use which API:

Use case	API
Quick-look / previews (< 1000 sources)	`create_cutouts_direct()`
Full catalogue processing	`Orchestrator`
Streaming / ML pipeline integration	`StreamingOrchestrator`

Streaming Mode (Advanced)

Process large catalogues in batches for integration into data pipelines using StreamingOrchestrator.

The StreamingOrchestrator class offers a dedicated API for streaming workflows and supports configurable background worker counts.

from cutana import get_default_config, StreamingOrchestrator

config = get_default_config()
config.source_catalogue = "sources.csv"
config.output_dir = "streaming_output/"
config.target_resolution = 256
config.selected_extensions = [
    {"name": "VIS", "ext": "PrimaryHDU"},
    {"name": "NIR-H", "ext": "PrimaryHDU"},
]
config.channel_weights = {"VIS": [1.0, 0.0], "NIR-H": [0.0, 1.0]}

# Create streaming orchestrator
orchestrator = StreamingOrchestrator(config)

try:
    # Initialize streaming. Kept inside the try: init_streaming() pre-spawns
    # workers one by one, so a failure partway through leaves the already-spawned
    # ones blocked on their pools with only cleanup() able to release them.
    orchestrator.init_streaming(
        batch_size=10000,
        write_to_disk=False,  # Return cutouts in memory (zero disk I/O)
    )

    # Process batches
    for i in range(orchestrator.get_batch_count()):
        result = orchestrator.next_batch()

        # result['cutouts']: list of numpy arrays, one per source (H, W, C)
        # result['metadata']: list of source metadata dicts
        # result['batch_number']: 1-indexed batch number

        # Your ML inference or analysis here...
        process_cutouts(result["cutouts"])

        # The next batches are already being prepared in background!
finally:
    # Required: workers block indefinitely waiting to hand over their next chunk,
    # so leaving the loop early - including via an error raised by next_batch() -
    # without cleanup() leaves them running and holding their shared memory pools.
    orchestrator.cleanup()

Key Parameters for init_streaming():

batch_size (int): Sources per batch. With write_to_disk=False batches are cut to exactly this size from a rolling buffer. With write_to_disk=True each batch is one zarr archive written by one worker and cannot be re-split, so batch_size is a minimum there: Cutana enlarges it to fit more sources into each worker process, and get_batch_count() reports correspondingly fewer batches
write_to_disk (bool): If False, returns cutouts via shared memory (recommended for ML pipelines)
max_workers (int): Maximum number of parallel workers
min_workers (int): Number of workers to pre-spawn at init time so they are already processing when next_batch() is first called
max_shm_memory_consumption (int): Total SHM memory budget in bytes

Progress and Status Methods

`get_progress()`

Get current progress and status information.

progress = orchestrator.get_progress()
print(f"Completed: {progress['completed_sources']}/{progress['total_sources']}")

Returns:

dict: Progress information including completed/total sources, runtime, errors

`get_progress_for_ui(completed_sources=None)`

Get progress information optimized for UI display.

from cutana.progress_report import ProgressReport

progress_report = orchestrator.get_progress_for_ui()
print(f"Progress: {progress_report.progress_percent:.1f}%")
print(f"Memory: {progress_report.memory_used_gb:.1f}/{progress_report.memory_total_gb:.1f} GB")

Parameters:

completed_sources (int, optional): Override completed sources count

Returns:

ProgressReport: Dataclass with UI-relevant progress information including system resources

`get_delivery_report()`

Check how many cutouts finished workers delivered against the sources they were assigned. A source whose cutout window falls outside its tile produces no cutout, so a run can legitimately end with fewer cutouts than the catalogue has rows — this reports what was lost and which worker lost it.

report = orchestrator.get_delivery_report()
if report["missing"]:
    print(f"{report['missing']} cutouts not produced: {report['shortfalls']}")

Returns:

dict: missing (total undelivered cutouts) and shortfalls (list of (process_id, assigned, delivered) tuples)

Control Methods

`stop_processing()`

Stop all active subprocesses gracefully.

result = orchestrator.stop_processing()
print(f"Stopped {len(result['stopped_processes'])} processes")

Returns:

dict: Stop operation results with list of stopped process IDs

`can_resume()`

Check if a workflow can be resumed from saved state.

if orchestrator.can_resume():
    print("Previous workflow can be resumed")

Returns:

bool: True if resumption is possible

Configuration Functions

`get_default_config()`

Get the default configuration object.

from cutana import get_default_config

config = get_default_config()
config.output_dir = "my_cutouts/"
config.target_resolution = 512

Returns:

DotMap: Configuration object with all default parameters

`save_config_toml(config, filepath)`

Save configuration to TOML file.

from cutana import save_config_toml

config_path = save_config_toml(config, "cutana_config.toml")

Parameters:

config (DotMap): Configuration to save
filepath (str): Path to save TOML file

Returns:

str: Path to saved file

`load_config_toml(filepath)`

Load configuration from TOML file.

from cutana import load_config_toml

config = load_config_toml("cutana_config.toml")

Parameters:

filepath (str): Path to TOML configuration file

Returns:

DotMap: Loaded configuration merged with defaults

Catalogue Functions

`load_and_validate_catalogue(filepath)`

Load and validate source catalogue from CSV file using catalogue_preprocessor.

from cutana.catalogue_preprocessor import load_and_validate_catalogue

try:
    catalogue_df = load_and_validate_catalogue("sources.csv")
    print(f"Loaded {len(catalogue_df)} sources")
except CatalogueValidationError as e:
    print(f"Validation error: {e}")

Parameters:

filepath (str): Path to CSV catalogue file

Returns:

pandas.DataFrame: Validated catalogue DataFrame

Raises:

CatalogueValidationError: If catalogue format is invalid

Name		Name	Last commit message	Last commit date
.github		.github
assets		assets
benchmarking		benchmarking
cutana		cutana
cutana_ui		cutana_ui
docs		docs
paper_scripts		paper_scripts
scripts		scripts
tests		tests
.gitignore		.gitignore
.python-version		.python-version
.vulture_whitelist.py		.vulture_whitelist.py
CHANGELOG.md		CHANGELOG.md
CITATION.cff		CITATION.cff
CONTRIBUTING.md		CONTRIBUTING.md
LICENCE.txt		LICENCE.txt
MANIFEST.in		MANIFEST.in
README.md		README.md
colours.txt		colours.txt
conftest.py		conftest.py
environment.yml		environment.yml
playwright.config.py		playwright.config.py
pyproject.toml		pyproject.toml
run_playwright_tests.py		run_playwright_tests.py
zensical.toml		zensical.toml

Latest commit

History

Folders and files

Repository files navigation

Cutana - Astronomical Cutout Pipeline

Support for Datalabs Users

Quick Start

Installation

Interactive UI (Recommended)

Programmatic API

Input Data Format

Output Formats

Flux Conversion

Pixel Units in FITS Headers

WCS (World Coordinate System) Handling

FITS Output

Zarr Output

Flux-Conserved Output

Multi-Channel Processing

Channel Weights Configuration

Image Normalisation

Output

Metadata

Images

Cutout Extraction Behavior

Padding Factor

Stretch/Normalisation Configuration

Linear Stretch

ASINH Stretch (Recommended)

Log Stretch

ZScale Stretch

Performance Considerations

Memory and CPU Optimization

Memory Constraints

Critical Usage Guidelines

Optimal Usage

Common Errors and Pitfalls

Advanced features

Configuration Parameters

Band selection

Backend API Reference

Orchestrator Class

Constructor

Main Processing Methods

start_processing(catalogue_data)

run()

Direct Cutout Generation (Fast, Small Batches)

Extract once, render many times

Streaming Mode (Advanced)

Progress and Status Methods

get_progress()

get_progress_for_ui(completed_sources=None)

get_delivery_report()

Control Methods

stop_processing()

can_resume()

Configuration Functions

get_default_config()

save_config_toml(config, filepath)

load_config_toml(filepath)

Catalogue Functions

load_and_validate_catalogue(filepath)

About

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Contributors

Languages

`start_processing(catalogue_data)`

`run()`

`get_progress()`

`get_progress_for_ui(completed_sources=None)`

`get_delivery_report()`

`stop_processing()`

`can_resume()`

`get_default_config()`

`save_config_toml(config, filepath)`

`load_config_toml(filepath)`

`load_and_validate_catalogue(filepath)`