Hugging Face's logo

Datasets:
mark000071
/
envship_v2_datasets

Tasks:
Time Series Forecasting
Other
Languages:
English
Size:
100K<n<1M
Tags:
maritime
trajectory-prediction
AIS
ship
environmental-context
social-interaction
License:
Dataset card Files
xet
 Community
envship_v2_datasets / DATA_CARD.md
mark000071's picture
mark000071
rebrand v2.2: EnvShip-Bench v2 → MARIS-Forecast (DATA_CARD.md); add Zenodo DOI 10.5281/zenodo.21224009
23e0e14 verified
preview code
|
Raw
History Blame Contribute Delete
19.3 kB

MARIS-Forecast Cross-Domain Extension — Data Card

Datasheet template adapted from Gebru et al., "Datasheets for Datasets" (2021).

This data card covers the cross-domain extension of MARIS-Forecast. It adds three jurisdictions (NOAA, Piraeus, Norway) on top of the original DMA-only release and introduces a long-horizon track (Track B). For the DMA-only single-track parent benchmark see ../ICDE_conferece_dataset_paper/DATA_CARD.md.

Motivation

Purpose. The cross-domain extension exists to test whether ship trajectory predictors trained on one maritime jurisdiction transfer to another, and whether long-horizon (60 min) forecasts trained on short trajectories generalise. The four jurisdictions span three operational regimes (Northern European coastal traffic, US open-ocean transits, Mediterranean port/ferry mix, Norwegian fjord and coastal traffic) and two distinct AIS time-resolutions / vessel-mix profiles.

Authors. Constructed by the MARIS-Forecast authors (see top-level CITATION.cff).

Funding. No external funding earmarked specifically for cross-domain construction.

Composition

What each instance represents. A single sample is a 20-minute window of a vessel's trajectory under Track A (10-minute observation + 10-minute prediction, 30 + 30 points at 20 s); or a 90-minute window under Track B (30-minute observation + 60-minute prediction, 90 + 180 points). All positions are in a local planar frame centred on the anchor (last history point), with x = east, y = north, units of metres.

Sample counts.

Track A (10/10):

Subset Train Val Test Total
DMA 120,000 15,000 15,000 150,000
NOAA 48,000 6,000 6,000 60,000
Piraeus 48,000 6,000 6,000 60,000
Norway 48,000 6,000 6,000 60,000
Combined 264,000 33,000 33,000 330,000

Track B (30/60):

Subset Train Val Test Total
DMA 46,744 5,414 6,000 58,158
NOAA 35,893 4,725 4,148 44,766
Piraeus 628 394 132 1,154
Norway 2,441 117 221 2,779
Combined 85,706 10,650 10,501 106,857

The Piraeus/Norway Track-B sample counts are intrinsically small because the underlying port/coastal trajectories rarely exceed two hours; this is a domain characteristic, not a defect.

Time spans (verified from first-row hist_end_ts in standard_track_v1 train CSVs).

Subset Min date Max date Days
DMA 2025-09-01 2025-09-30 30
NOAA 2025-03-01 2025-03-31 31
Piraeus 2019-01-01 2019-12-26 360
Norway 2025-08-01 2025-09-30 61

Bounding boxes (verified from anchor coordinates).

Subset lon range lat range
DMA 2.13 – 19.66 °E 52.75 – 59.85 °N
NOAA -159.12 – 144.34 13.70 – 49.45 °N (CONUS + Hawaii + western Pacific)
Piraeus 23.06 – 23.70 °E 37.67 – 38.03 °N
Norway 4.00 – 11.93 °E 57.50 – 60.49 °N (Skagerrak + south Norwegian coast)

What each row contains. Each row of <subset>/multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/{train,val,test}/part-000.csv.gz contains (full schema in summary.json):

  • sample_id, mmsi, segment_id, ship_type, ship_class, ship_type_id, ship_class_unified, ship_type_id_unified (Piraeus
    • Norway only)
  • hist_end_ts, pred_end_ts (UTC ISO timestamps)
  • hist_x_json / hist_y_json — 30 (Track A) or 90 (Track B) points; metres relative to anchor
  • fut_x_json / fut_y_json — 30 or 180 future points
  • hist_sog_json, hist_cog_sin_json, hist_cog_cos_json, hist_heading_sin_json, hist_heading_cos_json — kinematic arrays
  • hist_time_of_day_sin/cos_json, hist_day_of_week_sin/cos_json — periodic time encodings
  • hist_interp_json / fut_interp_json — booleans for interpolated points
  • interp_ratio_*, grid_interp_ratio_* — quality scalars
  • hist_displacement_m, fut_displacement_m
  • core_eligible, full_eligible, quality_tier, split

Per-sample artefacts (context_v1):

  • environment/rasters/{split}/masks.npz (N, 6, 128, 128) uint8 — load via np.load(...)['masks']
  • environment/rasters/{split}/signed_dist_shore.npy (N, 128, 128) float16, metres
  • environment/rasters/{split}/signed_dist_nav.npy (N, 128, 128) float16
  • environment/rasters/{split}/sample_ids.npy (N,) object
  • environment/vectors/{split}/vectors.jsonl.gz exact OSM polylines
  • environment/all_environment_descriptors.csv tabular scene descriptors
  • social/<split>_social_descriptors.csv neighbour-aware scalars

OSM temporal-consistency artefacts (new, see § OSM Temporal Consistency below):

  • osm_temporal_consistency/{train,val,test}_flags.csv per-sample flags
  • osm_temporal_consistency/summary.json aggregate
  • The 4 flag columns are merged inline into <subset>_track_v1/<split>/part-000.csv.gz in the v2-final HF release; paper-default filter is df[df["osm_temporal_consistent"] == "true"].

Phase 1+2 context columns (added 2026-06-19). Each row of the main CSV also carries 15 anchor-time scalar columns covering weather (6 cols), sea state (4 cols), port proximity (2 cols), fairway/TSS proximity (3 cols). See README.md for the full column-by-column schema and the CHANGELOG.md v2.1 entry for coverage and known-gap notes (Piraeus 2019 has empty wave columns).

Splits. Vessel-disjoint. stable_split_from_mmsi(mmsi) = MD5(MMSI) mod 10; bucket 0 = val, bucket 1 = test, otherwise = train. No MMSI appears in more than one split.

Missing / partial fields.

  • Ship dimensions (length, beam, draught) — available only for ~20–60% of vessels depending on subset (Piraeus best, Norway worst before VesselFinder enrichment; see norway_ship_trajectory_datasets/ENRICHMENT_REPORT.md).
  • Bathymetry, weather, currents — not bundled.
  • Visual imagery — not bundled.
  • ~0.2 % of samples have anchor_in_water = 0 (anchor falls on land per OSM raster, typically vessels berthed at piers); retained for completeness.

Class taxonomy.

  • ship_class — raw text class assigned at preprocess time.
  • ship_class_unified — 7+1 canonical classes (cargo, tanker, passenger, fishing, tug, service, sailing_leisure, plus unknown). Implemented in scripts/extras/taxonomy.py; mapping rule table is also documented there.
  • ship_type_id_unified — integer 0…7 matching the canonical class order.

For DMA and NOAA the raw ship_class is reused as-is (the upstream field is already clean). Piraeus and Norway gain the unified columns inline; Norway's unknown rate dropped from ~52 % to ~18 % on the test split after VesselFinder enrichment (scripts/extras/enrich_norway_static.py).

OSM temporal consistency

OSM is a living dataset: ports, piers, breakwaters, and quays added after the AIS date can appear as "land" in the SDF even when the trajectory was genuinely on water. This is highest-risk for the Piraeus 2019 subset (6 — 7 year gap to a 2026 OSM snapshot) and near-zero for the other three subsets (4 — 12 month gap).

Method (Stage 17 — scripts/extras/stage_17_osm_temporal_consistency.py).

For every sample, the 60 (Track A) or 270 (Track B) trajectory points are projected onto the per-sample 128 × 128 signed_dist_shore raster. A point with signed_dist_shore < 0 is inland per the OSM snapshot used for the build. We record max_inland_depth_m, n_inland_points, max_consec_inland_run, and a default flag:

osm_temporal_consistent = (max_inland_depth_m ≤ 30 m) AND (max_consec_inland_run < 3)

Thresholds are configurable; 30 m matches the SDF cell pitch of ~78 m so a single-cell brush is permitted (within-cell positional jitter).

Results — Track A against current 2026 OSM:

Subset Train consistent Val consistent Test consistent
DMA 119,135 / 120,000 (99.28 %) 14,881 / 15,000 (99.21 %) 14,889 / 15,000 (99.26 %)
NOAA 47,704 / 48,000 (99.38 %) 5,950 / 6,000 (99.17 %) 5,926 / 6,000 (98.77 %)
Piraeus 47,797 / 48,000 (99.58 %) 5,988 / 6,000 (99.80 %) 5,974 / 6,000 (99.57 %)
Norway 47,965 / 48,000 (99.93 %) 5,994 / 6,000 (99.90 %) 5,989 / 6,000 (99.82 %)

For Piraeus, we additionally rebuild the env/SDF stack using a contemporaneous OSM snapshot (greece-200101.osm.pbf — Geofabrik historical extract), and publish that as context_v1_2019osm/. This roughly halves the inconsistent-sample count for Piraeus because all new-build piers from 2020 – 2025 are excluded from the SDF. See § Piraeus historical-OSM rebuild.

Two subsets are released:

  • track_a_short-term_Cross-domain_Datasets/dma_track_v1/ — the full curated benchmark with the osm_temporal_consistency/<split>_flags.csv side-car (no rows deleted).
  • <subset>_track_v1/<split>/part-000.csv.gz inline filter — drop rows where osm_temporal_consistent != "true". This is the recommended default for paper main tables.

Pipeline

The cross-domain extension reuses the same 14-stage AIS pipeline from the parent DMA-only release and adds three virtual stages:

Stage Name Output
01 Field standardisation data_interim/01_standardized/
02 Validity filter 02_filtered/
03 Sort + dedup 03_deduped/
04 Ship-type-aware speed filter 04_shiptype_speed_filtered/
05 Trajectory segmentation 05_segmented/
06 Anchorage removal 06_underway_only/
07 Short-gap interpolation 07_gap_imputed/
08 20-s UTC resample 08_resampled_20s/
09 Second-pass anomaly check 09_rechecked/
10 Minimum-length filter 10_minlen_filtered/
11 Sliding window generation benchmark/full/
12 Quality + difficulty labels (in-place)
13 Core/full benchmark export benchmark/{core,full}/
14 Partition summaries benchmark/*.json
15 Stratified standard-track curation multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/
16 Env-SDF + social context (OSM) .../track_a_short-term_Cross-domain_Datasets/dma_track_v1/context_v1/
17 OSM temporal consistency .../osm_temporal_consistency/
10b Track-B segment-duration prefilter track_b_medium-term_Cross-domain_Datasets/<DS>/data_interim/

Track B re-runs stages 10b → 11 → 12 → 13 → 15 → 16 → 17.

Collection process

How acquired.

  • DMA: https://web.ais.dk/aisdata/ (aisdk-2025-09-*.zip, 30 daily ZIPs)
  • NOAA: https://marinecadastre.gov/ (March 2025 daily CSVs)
  • Piraeus: Zenodo 10.5281/zenodo.6323416 (CC BY 4.0; ferry-rich port AIS for 2019)
  • Norway: Kystverket / Kystdatahuset public API (NLOD 2.0; Aug + Sep 2025)

Time frame. Documented per-subset in <subset>/multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/summary.json.

Ethical review. AIS positions are broadcast publicly by vessels per SOLAS Ch. V. MMSI is a vessel identifier, not a personal identifier. No human-subjects review applies.

Preprocessing / cleaning / labelling

See § Pipeline above. Stage code lives under each subset's scripts/ plus shared utilities in scripts/extras/ (taxonomy unification, Norway VesselFinder enrichment, Track-B pre-filter, OSM PBF tile builder, OSM temporal-consistency flagger).

Uses

Baselines reported. See the parent ICDE paper (../ICDE_conferece_dataset_paper/paper_v4.pdf) for the full 25-baseline grid (physics, classical ML, traj-only DL, social-aware, env-aware, joint social+env, Transformer family). Headline numbers:

Track A in-domain ADE (metres, best-of-3 seed):

Model DMA NOAA Piraeus Norway
TCN (traj-only) 85.8 91.2 169.1 122.6
LSTM + Env-SDF 87.3 101.8 151.8 127.5
LSTM + Soc + Env-v2 87.2 — — —
GRU-2L — — — 128.1

Other downstream methods that consume this dataset include EnvSocial-TrAISformer (cross-domain Track B), GeoMode-TKDE (SDF-gradient CVAE), ship_LLM_traj_pred_benchmark (LMTraj adaptation), MFPD-TKDE (diffusion), AnchorDiff (DMA-only diffusion), and M-CTX-ICDE (spatial indexing benchmark).

Other tasks the dataset can support:

  • Anomaly / route-deviation detection
  • Vessel-type classification from kinematics
  • Multimodal stochastic forecasting (mixture / diffusion)
  • Geographic generalisation studies (train DMA, evaluate Piraeus etc.)
  • Interaction-aware planning under COLREGS
  • Long-horizon multi-modal forecasting (Track B)

Not to do.

  • Do not de-anonymise vessels for surveillance. MMSI is public but scaling into continuous tracking crosses ethical lines.
  • Do not deploy navigation-safety-critical predictors trained only on this benchmark without operating-region validation.

Distribution

Code, paper, metadata. Public Git repository (see top-level README once released).

Bulk data. The bulk tensors (vectors.jsonl.gz, masks.npz, signed_dist_*.npy, augmented CSVs) are released separately on Hugging Face. Build pipeline + raw-source pointers are sufficient for reconstruction from scratch.

Licences.

Component Licence
Pipeline code (this repo) CC BY 4.0
DMA AIS subset CC BY 4.0 (upstream)
NOAA AIS subset US public domain (upstream)
Piraeus AIS subset CC BY 4.0 (upstream Zenodo)
Norway AIS subset NLOD 2.0 (Kystverket)
OSM rasters / SDFs ODbL (© OpenStreetMap contributors)
Geofabrik historical PBFs ODbL

Top-level LICENSE and NOTICE.md carry the full text. Each sub-dataset has its own LICENSE matching the upstream provider.

Maintenance

Maintainer. Kun Ma (kunma1220@gmail.com).

Versioning. Semantic-versioned releases. Track A and Track B share the same version (currently standard_track_v1). Future revisions will bump to v2.

Planned additions.

  • Bathymetry channel from EMODnet (Europe) + GEBCO global
  • Tidal heights from FES2014
  • Weather (wind / wave) from ERA5 / NOAA WaveWatch III
  • Additional jurisdictions (Black Sea, East China Sea) pending data availability
  • Pre-computed time-disjoint OOD splits for Piraeus + DMA (full-year coverage)

Reporting issues. Open a GitHub Issue against the public repo.

Piraeus historical-OSM rebuild

For the Piraeus subset only, where the AIS year (2019) is ~6 years earlier than the current OSM snapshot, the SDF and 6-channel rasters are rebuilt with the Geofabrik 2020-01-01 Greece extract (greece-200101.osm.pbf, MD5 9c6da7651e624ab182c20d0d629d5e8c, 185 MB). This snapshot captures Greek coastal infrastructure as it stood at the very end of the AIS year.

The 2026-OSM context is also retained for ablation studies. The two contexts live side by side under:

  • Piraeus_ship_trajectory_datasets/multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/context_v1/ — current OSM
  • Piraeus_ship_trajectory_datasets/multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/context_v1_2019osm/ — 2020-01-01 OSM

Stage 17 is run against both; per-sample flag files end with _2019osm.csv for the historical variant. The default paper-default filter for Piraeus is computed from the historical-OSM consistency, since that is the most accurate ground truth for the 2019 trajectories.

File map 

Cross-domain-datasets/
├── LICENSE                       Top-level CC BY 4.0 + sub-licence index
├── NOTICE.md                     Required attribution lines
├── DATA_CARD.md                  This file
├── CITATION.cff                  Top-level citation
├── REPORT.md                     Build report + cross-domain study
├── DMA_ship_trajectory_datasets/      symlink → upstream DMA
├── NOAA_ship_trajectory_datasets/     symlink → upstream NOAA
├── Piraeus_ship_trajectory_datasets/  Tritsarolis 2022 / Zenodo 6323416
│   ├── LICENSE                        CC BY 4.0
│   ├── data_raw/                      raw zips + historical_osm/
│   ├── data_interim/                  per-stage outputs
│   ├── benchmark/                     stage 13 windowed
│   ├── multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/
│   │   ├── train/ val/ test/          gzipped CSVs (per-row JSON arrays)
│   │   ├── context_v1/                env-SDF + social (current OSM)
│   │   ├── context_v1_2019osm/        env-SDF + social (2020-01-01 OSM)
│   │   └── osm_temporal_consistency/  Stage-17 flag CSVs
│   └── scripts/                       pipeline + preprocess
├── norway_ship_trajectory_datasets/   Kystverket NLOD 2.0
│   └── (same layout)
├── track_b/                           long-horizon variant
│   ├── DMA/ NOAA/ Piraeus/ Norway/    same per-DS layout
│   └── scripts/                       Track-B pre-filter + builders
└── scripts/extras/                    shared utilities
    ├── taxonomy.py                    unified ship-class taxonomy
    ├── apply_unified_taxonomy.py      patch benchmark CSVs with unified cols
    ├── enrich_norway_static.py        VesselFinder static-info enrichment
    ├── stage_10b_track_b_filter.py    Track-B duration prefilter
    ├── stage_17_osm_temporal_consistency.py   Stage 17
    └── pbf_to_tile_cache.py           Geofabrik PBF → Overpass-JSON cache

Reproducibility 

bash scripts/reproduce_all.sh

Reproduces the four-jurisdiction Track A from raw drops, then Track B, then the OSM temporal-consistency stage, then the historical-OSM Piraeus rebuild. Wall time ≈ 5 h on an 8-core CPU host with 32 GB RAM. The build is deterministic for the same upstream snapshot.