EnvShip-Bench v2 — Cross-Domain Ship Trajectory Prediction
Four maritime jurisdictions and two prediction horizons, packaged in a single shape. Each sample carries an OpenStreetMap raster, a signed-distance field, a 3 km social-neighbour context, a unified ship-class label, and an inline OSM-temporal-consistency flag.
| Provider | AIS span | Track A (10/10) | Track B (30/60) | |
|---|---|---|---|---|
| DMA | Danish Maritime Authority | 2025-09 | 150,000 | 58,158 |
| NOAA | U.S. MarineCadastre | 2025-03 | 60,000 | 44,766 |
| Piraeus | Tritsarolis et al. (Zenodo 6323416) | 2019 | 60,000 | 1,154 |
| Norway | Kystverket / Kystdatahuset | 2025-08–09 | 60,000 | 2,779 |
| Combined | 330,000 | 106,857 |
- Track A — 10-min observation, 10-min prediction (30 + 30 points at 20 s).
- Track B — 30-min observation, 60-min prediction (90 + 180 points at 20 s).
Layout
README.md LICENSE NOTICE.md DATA_CARD.md CITATION.cff CHANGELOG.md SUMMARY_v2.md
checkpoints/ 9 DMA Track A baselines (unchanged from v1)
scripts/extras/ stage_17, PBF parser, taxonomy unifier, verifier
track_a_short-term_Cross-domain_Datasets/ # 10-min observation / 10-min prediction
├── dma_track_v1/
├── noaa_track_v1/
├── piraeus_track_v1/ # contains both context_v1/ (2026 OSM) and context_v1_2019osm/
└── norway_track_v1/
track_b_medium-term_Cross-domain_Datasets/ # 30-min observation / 60-min prediction
├── dma/standard_track_v1/
├── noaa/standard_track_v1/
├── piraeus/standard_track_v1/ # also has context_v1_2019osm/
└── norway/standard_track_v1/
Inside every leaf directory (e.g. track_a_short-term_Cross-domain_Datasets/dma_track_v1/
or track_b_medium-term_Cross-domain_Datasets/dma/standard_track_v1/):
train/ val/ test/ part-000.csv.gz — benchmark windows + 4 inline OSM-flag columns
context_v1/ env-SDF + social context
context_v1_2019osm/ Piraeus only — env-SDF built with 2020-01-01 OSM
osm_temporal_consistency/ per-sample flag CSV side-car (full numeric details)
osm_temporal_consistency_2019osm/ Piraeus only — same against 2020-01-01 OSM
reports/ per-split selection metadata
sample_ids/ deterministic sample-id catalogues
summary.json
Inside each context_v1/:
augmented/{train,val,test}/part-000.csv.gz benchmark + env + social merged
environment/
rasters/{split}/masks.npz per-sample 128×128×6 binary masks (uint8, compressed; key='masks')
rasters/{split}/{sample_ids,
signed_dist_shore,signed_dist_nav}.npy sample-id catalogue + 2-channel SDF (float16)
vectors/{split}/vectors.jsonl.gz OSM polylines, lossless
features/{split}/environment_descriptors.csv
anchors/{split}_anchors.csv all_anchors.csv
all_environment_descriptors.csv
osm_cache/tiles/*.json Overpass-format OSM tile cache (0.25°)
summary.json feature_stats.json failed_tiles.json
social/
features/{split}/social_descriptors.csv
snapshot_buckets/ compact AIS snapshots for neighbour lookup
OSM-temporal-consistency flag (inline in main CSVs)
Each row of train/val/test/part-000.csv.gz carries 4 extra columns:
| column | meaning |
|---|---|
osm_temporal_consistent |
true / false / empty (paper-default filter: true) |
osm_max_inland_depth_m |
deepest signed-distance into land along trajectory |
osm_n_inland_points |
trajectory points where SDF < 0 |
osm_max_consec_inland_run |
longest run of consecutive inland points |
The default rule is osm_temporal_consistent = true iff
max_inland_depth_m ≤ 30 m AND max_consec_inland_run < 3
(positional jitter within one SDF cell is tolerated). Full per-sample
numerics (including any_anchor_inland, anchor_inland_depth_m,
n_uncheckable_points) live in the side-car
osm_temporal_consistency/{split}_flags.csv.
For Piraeus, the inline flag is computed against the 2020-01-01 OSM
snapshot (greece-200101.osm.pbf, Geofabrik) because the AIS year is
2019; this avoids false positives from port piers built between 2020
and 2026. The 2026-OSM flag for ablation is preserved in the side-car
osm_temporal_consistency_2019osm/ (the directory keeps the
historical filename; the file inside holds the 2019-OSM flags that
the inline column already reflects).
Consistency rates per subset (Track A):
| Subset | Train consistent | Val consistent | Test consistent |
|---|---|---|---|
| DMA | 99.26 % | 99.19 % | 99.25 % |
| NOAA | 99.38 % | 99.17 % | 98.77 % |
| Piraeus | 99.59 % (2019 OSM) | 99.80 % | 99.63 % |
| Norway | 96.63 % | 97.43 % | 97.55 % |
Track B rates are in SUMMARY_v2.md.
Anchor-time weather / sea-state / port / TSS columns (Phase 1+2, 2026-06-19)
In addition to the OSM-temporal-consistency flag, every main CSV row carries 15 anchor-time scalars merged inline:
| Group | Column | Unit | Source |
|---|---|---|---|
| Weather | met_wind_speed_mps |
m s⁻¹ | Open-Meteo Archive (ERA5 reanalysis) |
met_wind_dir_deg |
deg, met. (FROM) | same | |
met_wind_rel_heading_deg |
deg | derived (wind dir − vessel COG) | |
met_temperature_c |
°C | same | |
met_pressure_hpa |
hPa | same | |
met_cloud_cover_pct |
% | same | |
| Sea state | sea_wave_height_m |
m | Open-Meteo Marine (ECMWF WAM) |
sea_wave_dir_deg |
deg, oceanographic (TO) | same | |
sea_wave_period_s |
s | same | |
sea_swell_wave_height_m |
m | same | |
| Port | port_nearest_dist_km |
km | OSM harbour=* / seamark:type=harbour |
port_nearest_name |
str | OSM name tag |
|
| TSS / fairway | in_fairway |
bool (≤100 m centreline) | OSM seamark:type=fairway |
dist_to_fairway_m |
m | same | |
in_tss |
bool | OSM seamark:type ∈ {separation_zone, separation_line, separation_boundary} |
Coverage caveats (see SUMMARY_v2.md):
- Piraeus wave columns are empty because Open-Meteo Marine begins on 2022-01-01 and Piraeus AIS is from 2019. The wind/temperature/pressure columns (ERA5 reanalysis archive) are fully populated.
- NOAA open-Pacific samples can have null wave entries where the model grid does not resolve a wave field (1.5°× 1.5° west of -150° E).
- Empty cell = no source data; absence is documented, not silent imputation.
Loader recipe (paper default):
df = pd.read_csv("…/dma_track_v1/train/part-000.csv.gz")
df["met_wind_speed_mps"] = df["met_wind_speed_mps"].astype(float)
# Filter for env-aware models: drop OSM-inconsistent + missing-wave rows
df_clean = df[(df["osm_temporal_consistent"] == "true") &
df["sea_wave_height_m"].notna()]
Quickstart
from huggingface_hub import snapshot_download
import pandas as pd
snapshot_download(
repo_id="mark000071/envship_v2_datasets",
repo_type="dataset",
local_dir="data/envship_v2",
allow_patterns=[
"track_a_short-term_Cross-domain_Datasets/*/train/**",
"track_a_short-term_Cross-domain_Datasets/*/val/**",
"track_a_short-term_Cross-domain_Datasets/*/test/**",
"*.md", "LICENSE", "CITATION.cff",
],
)
# Load DMA + Piraeus Track A (flags inline)
dma = pd.read_csv("data/envship_v2/track_a_short-term_Cross-domain_Datasets/dma_track_v1/train/part-000.csv.gz")
pir = pd.read_csv("data/envship_v2/track_a_short-term_Cross-domain_Datasets/piraeus_track_v1/train/part-000.csv.gz")
print(dma.shape, pir.shape) # ~120k and ~48k rows respectively
# Paper-default filter: drop OSM-temporal-inconsistent windows
dma_clean = dma[dma["osm_temporal_consistent"] == "true"].reset_index(drop=True)
pir_clean = pir[pir["osm_temporal_consistent"] == "true"].reset_index(drop=True)
print(dma_clean.shape, pir_clean.shape) # ~119k and ~47.8k rows
# Decode a Track A trajectory (30 history + 30 future, 20-s step)
import json, numpy as np
row = dma_clean.iloc[0]
hist_xy = np.column_stack([json.loads(row["hist_x_json"]), json.loads(row["hist_y_json"])])
fut_xy = np.column_stack([json.loads(row["fut_x_json"]), json.loads(row["fut_y_json"])])
print(hist_xy.shape, fut_xy.shape) # (30, 2) (30, 2)
For Track B load from
track_b_medium-term_Cross-domain_Datasets/<jurisdiction>/standard_track_v1/.
For env-aware models, also download the rasters under
<subset>/context_v1/environment/rasters/{split}/. They are NumPy
arrays aligned with the CSV rows via sample_ids.npy:
masks = np.load("rasters/train/masks.npz")["masks"] # (N, 6, 128, 128) uint8
sd_shore = np.load("rasters/train/signed_dist_shore.npy") # (N, 128, 128) float16, metres
sd_nav = np.load("rasters/train/signed_dist_nav.npy") # (N, 128, 128) float16, metres
sample_ids = np.load("rasters/train/sample_ids.npy", allow_pickle=True) # (N,) object
Licence
Composite — see LICENSE and NOTICE.md. Sub-licences:
| Subset | Upstream licence |
|---|---|
| DMA | CC BY 4.0 |
| NOAA | U.S. public domain |
| Piraeus | CC BY 4.0 (Zenodo 6323416) |
| Norway | NLOD 2.0 (Kystverket) |
| OSM rasters / SDFs | ODbL |
The processed benchmark and the pipeline code are released under CC BY 4.0. Attribution is required for every upstream subset used.
Citing
@dataset{envship_bench_v2_2026,
author = {Ma, Kun},
title = {EnvShip-Bench v2: A Cross-Domain Multi-Scale Benchmark for Context-Aware Ship Trajectory Prediction},
year = 2026,
publisher = {Hugging Face},
doi = {10.57967/hf/envship_v2_datasets},
url = {https://huggingface.co/datasets/mark000071/envship_v2_datasets}
}
When publishing results, cite the upstream AIS provider for every
subset used. See CITATION.cff and NOTICE.md for the recommended
attribution lines.
What's new in v2 (high level)
- Three new jurisdictions — NOAA (cross-domain transfer), Piraeus (port + ferry), Norway (fjord + coast).
- Track B added for all four jurisdictions.
- Stage 17 OSM-temporal-consistency — every sample carries an inline flag indicating whether its trajectory stays in water on the OSM snapshot used to build the env context.
- Historical OSM for Piraeus — Piraeus AIS is from 2019; the
inline flag uses the 2020-01-01 Geofabrik snapshot
(
context_v1_2019osm/) so port construction after 2020 does not produce false positives. - DMA + Norway env rebuilt after refilling 188 + 184 originally failed OSM tiles from Geofabrik archives.
CHANGELOG.md has the full version history; SUMMARY_v2.md walks
through the v2 methodology in narrative form.
- Downloads last month
- 1,858