ayzk/mosaic-v0

Mirror of the Zenodo release https://doi.org/10.5281/zenodo.23582250 (cite that DOI).

Mosaic v0: project layers, loader, and Croissant metadata for the 446 hand-labeled Sen1Floods11 chips (preview release)

Version 0.1.0 (preview). Authors: Kai Zhao (Florida State University), Xin Zhang (Florida State University), Yafan Huang (University of Maryland, College Park), Qing Yang (University of Maryland, College Park).

What v0 is

Mosaic v0 is a preview of how Mosaic enhances one existing flood dataset. It adds project layers to the 446 hand-labeled chips of Sen1Floods11 v1.1 (Bonafilia et al. 2020; 11 flood events, 512 x 512 px, about 10 m, EPSG:4326). The layers are outputs of the Mosaic pilots:

• Per-pixel posteriors (reliability). Each scored pixel gets P(water), estimated from four label sources that disagree. The model is a Dawid-Skene model with the two Sentinel-1 sources treated as one lineage family. It uses 46 hand-labeled anchor chips, a weighted anchor, and land cover. These are the pilot's follow-up settings (f) and (g).
• The four label sources and their lineage families. S1-Otsu and MMFlood-UNet form the Sentinel-1 family, S2-MNDWI forms the Sentinel-2 optical family, and JRC is permanent water. The table is tables/lineage_families.csv.
• Disagreement flags. A per-pixel bitmask records which sources disagree with the posterior and with each other.
• Splits for event-blocked evaluation. The tables give the event fold (leave one event out), the pilot's anchor/test roles, the official Sen1Floods11 split, and the events that overlap other flood datasets.
• Release provenance. The package records the URL, size and SHA-256 of every original file it builds on. It also includes the code that produced each layer and the pilot results it reproduces.
• Kuro Siwo pilot disagreement flags. These cover a second pilot that compared MMFlood with Kuro Siwo annotations. They lie on MMFlood tile grids for 22 CEMS activations, not on the 446 Sen1Floods11 chips; see "Kuro Siwo pilot layer" below.

v0 does not redistribute the original Sen1Floods11 imagery or labels. The loader downloads them from the dataset's official public bucket (gs://sen1floods11, over HTTPS), checks each file against the SHA-256 recorded here, and returns them together with the project layers.

Quick start

pip install numpy rasterio          # Python >= 3.9
python examples/verify_loader.py    # fetches 3 chips (about 10 MB), checks layers against the originals

import mosaic_v0       # run from this folder; or `pip install .` and set MOSAIC_V0_ROOT=<this folder> (or pass root=...)
chips = mosaic_v0.load_chips(["Bolivia_103757", "USA_1010394"])
c = chips[0]
c["s1"], c["s2"], c["label_hand"], c["label_s1otsu"], c["jrc"]   # originals (fetched, SHA-256 checked, cached)
c["posterior_g"], c["posterior_f"]                                 # P(water), NaN where < 2 valid sources
c["s2_mndwi"], c["mmflood_unet"], c["landcover_group"], c["flags"]
mosaic_v0.flag(c["flags"], 4)                                      # pixels where the valid sources disagree
folds = mosaic_v0.event_folds()                                    # {event: fold} for leave-one-event-out
mosaic_v0.load_chips(event="Spain", originals=False)               # project layers only, no download

The cache directory defaults to ~/.cache/mosaic_v0; set MOSAIC_V0_CACHE to change it.

Files

path	content
layers/sen1floods11/posterior/<chip>_posterior.tif (446)	float32, 2 bands: P(water) under setting (f) and setting (g); NaN = fewer than 2 valid sources
layers/sen1floods11/aux/<chip>_aux.tif (446)	uint8, 4 bands: S2-MNDWI source, MMFlood-UNet source (1 water / 0 not / 255 invalid), land-cover group (0-4), disagreement flags (bitmask)
layers/kurosiwo_pilot/flags/<tile>_ksflags.tif (268)	uint8 bitmask on MMFlood tile grids, 22 activations (see below)
tables/chips.csv	one row per chip: event, bounds, event_fold, pilot_role, sen1floods11_split, layer paths, per-chip summary statistics
tables/events.csv	11 events: dates, orbit, chip counts, event_fold
tables/event_overlaps.csv	Sen1Floods11 events that overlap MMFlood or Kuro Siwo events (same CEMS activation, or within 45 days and 500 km)
tables/lineage_families.csv	the four label sources and the hand reference: definition, producer, lineage family, how the posterior models each
tables/originals.csv	2,230 original files (5 layers x 446 chips): official URL, gs:// URI, bytes, SHA-256
tables/posterior_model.json	fitted parameters of settings (f) and (g); reproduction checks against the pilot's published metrics
tables/kurosiwo_pilot_tiles.csv, tables/kurosiwo_pilot_flags.json	Kuro Siwo pilot tile index, flag thresholds and reproduction checks
mosaic_v0/	Python loader (load_chips, load_kurosiwo_pilot_flags, chips_table, lineage_families, event_folds, flag)
examples/verify_loader.py	end-to-end check of the loader
code/build/	scripts that built the layers and tables from the pilot outputs
code/pilot/	the pilots' code as run (Sen1Floods11 Dawid-Skene pilot; Kuro Siwo pilot)
provenance/	the pilots' write-ups and the result tables the build reproduces; Kuro Siwo annotations license
croissant.json	Croissant 1.0 (JSON-LD) metadata with the Responsible AI fields
zenodo.json	deposit metadata for Zenodo
SHA256SUMS	SHA-256 of every file in the package

Layers in detail

Posterior P(water). The model uses four binary sources: 0 S1-Otsu, 1 S2-MNDWI, 2 JRC and 3 MMFlood-UNet, each 1, 0 or invalid. A pixel is scored when at least 2 sources are valid. The two Sentinel-1 sources share one 2 x 4 joint confusion matrix (the S1 family); S2-MNDWI and JRC each have their own 2 x 2 matrix. On the 46 anchor chips (10% per event, seed 0) the hand labels are clamped and counted lam times in the M-step. lam was chosen by leave-one-event-out cross-validation over the anchor chips only.

• Setting (f): S1 family plus weighted anchor (lam = 300).
• Setting (g): (f) plus separate matrices for five ESA WorldCover 2021 land-cover groups (lam = 1000).

The posterior depends only on the pixel's source pattern and its land-cover group.

On the 400 test chips (about 90.4 M pixels with a valid hand label and at least 2 valid sources), the pilot reported the following for (g):

• ECE 0.013 and Brier 0.031;
• mean absolute error of the 8 estimated source error rates: 1.1 points.

For (f): ECE 0.020, Brier 0.032, mean absolute error 2.1 points.

The build re-fitted both settings with the pilot's unchanged code and matched these metrics to within 1e-6 (tables/posterior_model.json). ECE recomputed directly from the shipped (g) rasters is 0.0131. Hand labels of the 46 anchor chips were used to fit the model, so evaluate posteriors on pilot_role == "test" chips only.

Disagreement flags (aux band 4, Sen1Floods11 chips). Each bit marks one condition:

bit	meaning
0	S1-Otsu disagrees with the MAP of (g)
1	S2-MNDWI disagrees with the MAP of (g)
2	JRC disagrees with the MAP of (g)
3	MMFlood-UNet disagrees with the MAP of (g)
4	the valid sources disagree among themselves
5	the S1 family (where Otsu and U-Net agree) disagrees with S2-MNDWI
6	0.2 < P(g) < 0.8
7	fewer than 2 valid sources

These flags come from this package's sources and posterior. They are not the Kuro Siwo pilot flags.

Land-cover group (aux band 3). The group is cropped from ESA WorldCover v200 (2021), with nearest resampling to the chip grid:

value	group	WorldCover classes
0	Tree/shrub/mangrove	10, 20, 95
1	Grass/wetland	30, 90, 100
2	Cropland	40
3	Built/bare	50, 60, 70
4	Water	80

Splits.

• event_fold assigns each of the 11 events its own fold, for event-blocked (leave-one-event-out) evaluation. The pilot chose lam with these folds.
• pilot_role is the pilot's anchor/test assignment. It is stratified by event, not event-blocked: every event has anchor chips.
• sen1floods11_split is the official v1.1 hand-labeled split (train 252 / valid 89 / test 90 / bolivia 15). Apart from Bolivia, it is not event-blocked.
• To keep evaluation event-blocked across datasets, drop the events in tables/event_overlaps.csv from the other dataset whenever a Sen1Floods11 event is held out.

Kuro Siwo pilot layer. This pilot asked whether models trained on other data find label errors in MMFlood. It compared MMFlood masks with Kuro Siwo annotation polygons on 457 MMFlood tiles of 22 shared CEMS activations. EMSR498 was excluded for misregistration. Rasters are written for the 268 tiles that intersect the Kuro Siwo AOI; the other tiles are listed in tables/kurosiwo_pilot_tiles.csv.

The flag bits are:

bit	meaning
0	V_total: inside the Kuro Siwo AOI and S1 valid
1	V_flood: V_total, with no permanent water in either dataset
2	MMFlood flood mask differs from the Kuro Siwo event layer
3	total-water disagreement
4	within 2 px of a flood-label boundary
5	D1 flag: top 5% of pooled V_flood pixels by the score of a U-Net trained only on the Sen1Floods11 hand-labeled train split
6	D2 flag: top 5% by the confident-learning score of 4-fold cross-fitted MMFlood U-Nets

The exported flags reproduce the pilot's precision and recall at a 5% flag rate:

• D1: precision 0.315, recall 0.753;
• D2: precision 0.304, recall 0.727.

These are measured against the MMFlood-vs-Kuro Siwo flood disagreement (tables/kurosiwo_pilot_flags.json). The pilot's D3 combination is not exported, because its rank-reference sample was not saved and cannot be regenerated exactly.

None of the 22 pilot activations overlaps a Sen1Floods11 event in space and time. tables/event_overlaps.csv lists one Kuro Siwo event (EMSR205) near the Sri-Lanka event, but it is not among the pilot activations. The link to the 446 chips is the D1 detector, which was trained on the Sen1Floods11 hand labels. For 15 of the 22 activations, the Kuro Siwo layers carry the original CEMS delineation, so "disagreement" there mostly reflects processing and version differences, not independent re-annotation.

Known limitations

These are taken from the pilot write-ups in provenance/:

• Only 11 events, so event-level uncertainty is large.
• S2-MNDWI is not independent of the hand labels, which were drawn with Sentinel-2 imagery.
• The MMFlood U-Net source is weak.
• JRC covers permanent water only.
• WorldCover is from 2021, while the chips are from 2016-2019.
• Error rates implied directly by the fitted confusion matrices are no more precise than counts on the 46 anchor chips.
• The S1 family was chosen after the first pilot showed correlated errors, so settings (d)-(g) are not a pre-registered test.

Provenance

item	source
Originals	Sen1Floods11 v1.1, gs://sen1floods11/v1.1/data/flood_events/HandLabeled/ (S1Hand, S2Hand, LabelHand, S1OtsuLabelHand, JRCWaterHand), downloaded with code/pilot/sen1floods11-dawid-skene/00_download.py; file sizes verified against the bucket listing; SHA-256 in tables/originals.csv
S2-MNDWI source and cloud mask	01_prepare.py
MMFlood-UNet source	03_train_unet.py, 04_predict.py; MMFlood from Hugging Face links-ads/mmflood
Land cover	02_worldcover.py, ESA WorldCover v200 2021
Posteriors and flags	code/build/build_s11_layers.py, which calls the pilot's 05_ds_eval.py and 08_family_weighted.py unchanged
Kuro Siwo pilot flags	code/build/build_ks_flags.py from the Kuro Siwo pilot's cached outputs (code/pilot/kurosiwo/). Kuro Siwo annotations: github.com/Orion-AI-Lab/KuroSiwo-annotations, commit b930dd2 (2026-05-22)
Tables	code/build/build_tables.py. The event overlaps come from the Mosaic leave-one-dataset-out flood experiment

Everything was built on the project's lab machine on 2026-10-22 with Python 3.12.3 and numpy 2.5.2, scipy 1.18.1, pandas 3.0.6 and rasterio 1.5.2.

License

• Project layers and tables (everything under layers/, tables/, provenance/ and the metadata files): Creative Commons Attribution 4.0 International (CC BY 4.0), https://creativecommons.org/licenses/by/4.0/. See LICENSE-DATA.
• Code (mosaic_v0/, examples/, code/): BSD 3-Clause. See LICENSE-CODE.
• Third-party data used to build the layers stays under its own terms. See THIRD_PARTY_NOTICES.md. The original Sen1Floods11 files are not included. Users obtain them from the official source and must follow and cite its terms.

Citation

Please cite Mosaic v0 and the datasets it builds on:

@misc{mosaicv0,
  author    = {Zhao, Kai and Zhang, Xin and Huang, Yafan and Yang, Qing},
  title     = {{Mosaic v0: project layers, loader, and Croissant metadata for the 446 hand-labeled Sen1Floods11 chips (preview release)}},
  publisher = {Zenodo}, year = {2026}, version = {0.1.0},
  doi       = {10.5281/zenodo.23582250}
}
@inproceedings{bonafilia2020sen1floods11,
  title     = {{Sen1Floods11: a georeferenced dataset to train and test deep learning flood algorithms for Sentinel-1}},
  author    = {Bonafilia, Derrick and Tellman, Beth and Anderson, Tyler and Issenberg, Erica},
  booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)},
  pages     = {835--845}, year = {2020}, doi = {10.1109/CVPRW50498.2020.00113}
}
@inproceedings{bountos2024kurosiwo,
  title     = {{Kuro Siwo: 33 billion m$^2$ under the water. A global multi-temporal satellite dataset for rapid flood mapping}},
  author    = {Bountos, Nikolaos and Sdraka, Maria and Zavras, Angelos and Karasante, Ilektra and Karavias, Andreas and Herekakis, Themistocles and Thanasou, Angeliki and Michail, Dimitrios and Papoutsis, Ioannis},
  booktitle = {Advances in Neural Information Processing Systems 37}, pages = {38105--38121}, year = {2024}, doi = {10.52202/079017-1204}
}

Also cite MMFlood (Montello et al. 2022, IEEE Access 10, doi:10.1109/ACCESS.2022.3205419) and ESA WorldCover 2021 v200 (Zanaga et al. 2022, doi:10.5281/zenodo.7254221).