Galaxy10 DECaLS

Galaxy10 DECaLS is a set of 17,736 colour images of galaxies in ten morphology classes, published on Zenodo on 4 July 2024. The images come from the DESI Legacy Imaging Surveys in the g, r and z bands, 256 × 256 pixels in the release. The class labels come from the votes of Galaxy Zoo volunteers.

This page follows the data from the survey camera and the volunteers to the 17,736 × 12,288 pixel matrix: what the provider documents, what is held here, and how 17,736 released images become matrix rows. This site holds neither the release nor the 64 × 64 copy the model was trained on; the images on this page are the demo’s saved display copies of its test galaxies.

The images come from the DESI Legacy Imaging Surveys. Its DECaLS part was observed with the Blanco telescope at Cerro Tololo Inter-American Observatory (proposal 2014B-0404; PIs David Schlegel and Arjun Dey), as the survey acknowledgement in the astroNN documentation states. Galaxy Zoo volunteers classified the galaxies by eye: Galaxy Zoo DR2 re-imaged with DECaLS, and the DECaLS campaigns ab and c, about 441,000 unique galaxies in all. Henry Leung and Jo Bovy selected 17,736 of them into ten broad classes “using volunteer votes with more rigorous filtering” (astroNN documentation). The thresholds of that filtering are not published there.

From the sky to the held files

1. DocumentedDESI Legacy Imaging Surveys

   g, r and z imaging; the DECaLS part observed with the Blanco telescope, Cerro Tololo

2. DocumentedGalaxy Zoo volunteers

   Votes on Galaxy Zoo DR2 (re-imaged with DECaLS) and the DECaLS campaigns ab and c: about 441,000 unique galaxies

3. DocumentedGalaxy10 DECaLS release

   Leung & Bovy select 17,736 images into ten classes; 256 × 256 × 3, 2.7 GB

4. Not recorded64 × 64 copy

   The file the demo trained on; how it was reduced from 256 pixels is not recorded. Not held

5. ReproducedSplit at seed 42

   12,415 / 1,777 / 3,544 rows; rerun here from row numbers

6. HeldHeld display copies

   3,544 test thumbnails, 129 case tiles, the test labels and the stored counts

How to read the labels: Documented — described by the provider, not rerun here. Not recorded — the step happened, but no file held here says how. Reproduced — rerun here. Held — files this page actually has.

One acquisition unit: one galaxy image

One unit is one galaxy: a three-band colour cutout, 256 × 256 pixels in the release, with one class label decided by the volunteers’ votes. The label is a human judgement of visible shape, not a measurement.

What the release records per galaxy

Pixel size: the release gives it per image in pxscale, which is not held. The 64-pixel copy has 4× fewer pixels along each side than the release; its own pixel scale is not recorded.

The ten classes

The files, opened

Held (the release and the 64 × 64 copy are not)
├─ case sheet           JPEG, quality 90 · 960 × 576 px · 15 × 9 tiles of 64 × 64 · 129 used · 138 kB
├─ test-thumbnail sheet WebP, quality 76 · 2,880 × 2,880 px · 60 × 60 tiles of 48 × 48 · 3,544 used · 651 kB
├─ saved test outputs   NumPy archive · 541 kB
│   ├─ test_indices             3544 int64
│   ├─ truth                    3544 int64
│   ├─ baseline_logits          3544 × 10 float32
│   ├─ augmented_logits         3544 × 10 float32
│   ├─ baseline_probabilities   3544 × 10 float32
│   └─ augmented_probabilities  3544 × 10 float32
└─ stored record inside the demo page
    dataset · split · protocol · model · baseline · augmented · cases · sprites · embeddings

Both sheets tile their images row by row: the thumbnail sheet in test-set order, the case sheet in the order the demo stored its 129 selected test galaxies. test_indices gives each test image’s row in the 64 × 64 copy; truth its class.

One record, drawn

The first stored case: row 3,013 of the copy, a Round smooth galaxy, test position 0. Left, its 64 × 64 case tile; beside it, the same galaxy in the 48 × 48 thumbnail sheet; below, the values of its three channels as decoded from the JPEG tile.

What the held files cover

Each tick is one held test image, placed at its row in the 64 × 64 copy; each lane is one class. The shaded band in each lane spans the rows that class would occupy if the file were sorted by class with the documented counts. Hover over a lane to read it.

What the held record does not give you directly

• The rows are grouped by class. All 3,544 held test labels fall inside their class’s block of rows, and the stored training and validation counts agree with the same blocks. Any split that does not shuffle would put whole classes into one part.
• Only display copies of the pixels are held. The 129 case tiles are JPEG (quality 90) at 64 × 64; the 3,544 thumbnails are WebP (quality 76), reduced to 48 × 48. The values the model read are not held.
• Only the test fifth is held as images. 3,544 of 17,736 images; the 12,415 training and 1,777 validation images are present only as class counts.
• A row number is not a galaxy identifier. The held files carry no sky position, redshift or pixel scale, so a galaxy cannot be matched to the survey from them.
• The 64-pixel copy is not the release. The release is 256 × 256. Who reduced it, with which resampling, and whether from the standard file or from the second, “NoDuplicated”, file on the same Zenodo record, is not recorded.

The provider selects and labels the images. An unrecorded step reduces them to 64 × 64, and the demo’s preprocessing casts them to float32 and cuts one split. The demo adds no screening of its own: nothing is dropped after the provider’s selection, so 17,736 released images become 17,736 matrix rows.

Where the images go

Follow one galaxy

Pick a galaxy to see it at each step, from the held tile to the row the matrix stores. There is one example per class; nothing is dropped, so every example reaches the stored row. All held images are test rows.

Checks against the upstream documentation

The page keeps the documented class counts; the stored record agrees with them. The split is rerun here with scikit-learn 1.9.0. The held test labels are consistent with a file sorted by class. No held file identifies the 64-pixel copy byte for byte with the release.

Labels kept beside the matrix

• The classes are imbalanced: 334 Cigar-shaped smooth images against 2,645 Round smooth, a factor of 7.9.
• The reduction from 256 to 64 pixels is not recorded, and the 64-pixel copy is not held, so its exact identity with the release cannot be checked here.
• The Zenodo record also offers a file with duplicates removed; which of the two files the copy came from is not recorded.
• The labels are volunteers’ judgements of visible shape, selected with filtering whose thresholds are not published in the documentation read for this page.
• The rows are grouped by class. The seed-42 split shuffles before cutting, so its parts are unaffected.

Galaxy augmentation explorer: its Workflow page trains the same small CNN with and without random rotation of the training images and compares the two on the test rows; its Insight page compares two saved runs of that comparison.