## v2.2 rebrand — 2026-07-06 - **Dataset rebranded** from *EnvShip-Bench v2* to **MARIS-Forecast** — *A multi-region AIS trajectory dataset with temporally aligned maritime context for vessel trajectory forecasting.* # Changelog All notable changes to this benchmark are listed here. Dates are UTC. ## v2.1 weather + waves + ports + TSS — 2026-06-19 Each Track A / Track B sample CSV row now carries 15 anchor-time scalars (no row changes, no flag-filter break). Sources: - **Weather** (Open-Meteo Archive, ERA5 reanalysis): 10-m wind speed & direction, 2-m temperature, surface pressure, cloud cover. Wind direction also reported relative to vessel COG. - **Sea state** (Open-Meteo Marine, ECMWF WAM): significant wave height, mean wave direction & period, swell wave height. - **Port proximity** (OSM `harbour=*` / `seamark:type=harbour`): nearest port distance (km) + name. - **Fairway / TSS** (OSM `seamark:type=fairway` / `separation_zone|line|boundary`): in-fairway flag, distance to fairway, in-TSS flag. Coverage: | subset | wind | wave | port | in_tss | |---|---|---|---|---| | DMA Track A | 100 % | 98 % | 100 % | 3 % (Skagerrak) | | NOAA Track A | 100 % | 93 % | 100 % | 1 % | | Piraeus Track A | 100 % | 0 % (see below) | 100 % | 0 % | | Norway Track A | 100 % | 97 % | 100 % | 0 % | **Known gap.** Open-Meteo Marine reanalysis begins 2022-01-01. Piraeus AIS is from 2019, so wave columns there are blank. Other columns (weather, port, TSS) are present for every Piraeus sample. Build wall-clock: ≈ 3.5 h (NOAA OSM cluster-fetch dominates). Code: `scripts/extras/build_meteo_features.py`. ## v2 compact — 2026-06-18 (storage cleanup) The bulk of the HF storage was the uncompressed `masks.npy` rasters (6-channel uint8 binary, 0.5–11 GB each, 27 files total). Switching to `numpy.savez_compressed` is fully lossless on binary masks and shrinks each file 100–1700×: - All 27 `rasters//masks.npy` replaced with `masks.npz` (key=`masks`). - Total mask storage 40.1 GB → 283.6 MB (saved 99.3%). - LFS history super-squashed to drop pre-v2 raster versions and intermediate upload-phase artefacts. Loader change for users: masks = np.load("…/rasters/train/masks.npz")["masks"] # was np.load(…masks.npy) Other rasters (`signed_dist_shore.npy`, `signed_dist_nav.npy`, `sample_ids.npy`) are unchanged. SDFs are float16 metres at 78 m grid pitch — float32 would be excess precision and was never used in the build pipeline. ## v2 final — 2026-06-17 (HF simplification) Following user feedback for an open-source-friendly layout, the HF dataset was simplified to **one directory per (subset, track)**: - The 4 stage-17 OSM-temporal-consistency flag columns (`osm_temporal_consistent`, `osm_max_inland_depth_m`, `osm_n_inland_points`, `osm_max_consec_inland_run`) are merged inline into `_track_v1/{train,val,test}/part-000.csv.gz`. - For Piraeus, the inline flag uses the 2020-01-01 OSM snapshot (matches the 2019 AIS year). The 2026-OSM flag remains in the side-car `osm_temporal_consistency/` for ablation. - The 16 redundant sibling directories (`_track_v1_all/`, `_track_v1_filtered/` × 4 subsets × 2 tracks) were removed from HF. The same row sets are recoverable in one line of pandas: df = pd.read_csv("_track_v1/train/part-000.csv.gz") df_filtered = df[df["osm_temporal_consistent"] == "true"] - `scripts/extras/publish_dual_subset.py` is kept for historical reproducibility but marked deprecated. ## v2 — 2026-06-14 > **Notice to downstream authors using the Norway subset** > > The Norway environmental context (`context_v1/environment/`) has > been substantively updated. If your trained model relies on the > per-sample SDF / raster / scene-type fields for Norway, re-evaluate > against the new context. The corresponding tensors at > `rasters/{split}/{masks,signed_dist_shore,signed_dist_nav}.npy` and > `all_environment_descriptors.csv` have all changed for ~85 % of > Norway samples. Pre-v2 numbers for Norway env-aware models are no > longer comparable to v2 numbers. The Track A trajectory CSVs and > the env-free models are unaffected. ### Corrections to documentation - Time spans in the original 2026-05-21 `REPORT.md` were inaccurate. Verified time spans (from `train/part-000.csv.gz` first-row `hist_end_ts`): - DMA = 2025-09-01 ~ 2025-09-30 (was claimed: 2019) - NOAA = 2025-03-01 ~ 2025-03-31 (was claimed: 2015–2019) - Piraeus = 2019-01-01 ~ 2019-12-26 (was partly claimed: 2019-12 only) - Norway = 2025-08-01 ~ 2025-09-30 (unchanged) - Piraeus upstream license corrected: CC BY 4.0, not CC BY-NC-SA 4.0 (verified from Zenodo 10.5281/zenodo.6323416 metadata). ### New files - `LICENSE`, `NOTICE.md`, `DATA_CARD.md`, `SUMMARY_v2.md` - `Piraeus_ship_trajectory_datasets/LICENSE` (CC BY 4.0) - `norway_ship_trajectory_datasets/LICENSE` (NLOD 2.0) - `NOAA_ship_trajectory_datasets/LICENSE` (US public domain notice) - `scripts/extras/stage_17_osm_temporal_consistency.py` - `scripts/extras/pbf_to_tile_cache.py` - `scripts/extras/publish_dual_subset.py` - `scripts/extras/post_rebuild_refresh.sh` - `scripts/extras/verify_v2_artifacts.py` ### New data artefacts per subset - `multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/osm_temporal_consistency/` — Stage-17 per-sample flag CSVs. - `multi_type_mini_bench_build/standard_track_v1_all/` — Original benchmark rows with 4 extra flag columns. - `multi_type_mini_bench_build/standard_track_v1_filtered/` — Subset where `osm_temporal_consistent = true`. **Recommended default.** ### Piraeus-only - `data_raw/historical_osm/greece-200101.osm.pbf` (185 MB, Geofabrik archive). - Track A: `multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/context_v1_2019osm/` — Env / SDF / raster stack rebuilt against the 2020-01-01 OSM snapshot for temporal alignment with the 2019 AIS year. - Track B: `track_b/Piraeus/multi_type_mini_bench_build/track_a_short-term_Cross-domain_Datasets/dma_track_v1/context_v1_2019osm/` — Same rebuild applied to the long-horizon track (1,154 samples). Build wall time: 10 min. - Stage-17 flag CSVs against 2019-OSM at `osm_temporal_consistency_2019osm/` for both Track A and Track B. These drive Piraeus's default `standard_track_v1_filtered/`. - Per-sample comparison of the 2026-OSM vs 2019-OSM consistency flags: - Track A: train 8 / val 0 / test 4 samples flipped consistent under 2019 OSM (i.e. the 2026 OSM incorrectly flagged them as inland due to post-2020 port construction). - Track B: 0 samples changed (Track B trajectories happen to stay clear of the contested coastal cells; we ship the historical context anyway for ablation symmetry). ### DMA-only - `data_raw/dma/historical_osm/denmark-260101.osm.pbf` (459 MB). - 188 → 2 failed OSM tiles after Geofabrik refill + env rebuild. - Top-level `context_v1/summary.json` now correctly serialises the `environment` section (previously empty `{}` due to a cosmetic bug in the summary writer — the underlying data was always at `environment/summary.json` and was correct). ### Norway-only - `data_raw/historical_osm/norway-260101.osm.pbf` (1.36 GB). - 184 → 2 failed OSM tiles after refill + env rebuild. - Scene-counts on train split rebalanced from `open_water 47,695 / nearshore 168 / harbor 0 / constrained 137` to `open_water 34,888 / nearshore 10,113 / harbor 1,320 / constrained 1,679`, reflecting the true fjord / coastal distribution that had been hidden by the silently-failing OSM tiles. - The pre-rebuild context is no longer kept on disk for Norway (overwritten by the v2 rebuild). The pre-rebuild failed-tile list is preserved at `failed_tiles.pre_refill.json`. ### Known limitations carried into v2 - 2 DMA + 2 Norway OSM tiles still fail Overpass (429/504 throttle). The affected anchors are all in the open North Sea or Baltic with no expected coastline features; their SDFs read as uniformly +5000 m (open water), which is the correct outcome for the anchor locations. - The Track B contexts were built later when Overpass was stable (failed_tiles = 0 for DMA / Piraeus / Norway, = 17 for NOAA). They were NOT rebuilt as part of v2; their existing SDFs and stage-17 flags are kept as-is. ### Build wall times | Job | Elapsed wall | |------------------------------------|---------------:| | DMA env rebuild (150 K samples) | 6.5 h | | Piraeus 2020-01 OSM rebuild (60 K) | 8.6 h | | Norway env rebuild (60 K) | 10.3 h | | Stage 17 + dual-subset publish | < 5 min | Total wall time (3-way parallel + final stage): ~10.5 h. ## v1.x — earlier history See `REPORT.md` and `PIPELINE_NOTES.md` for the pre-v2 history. Note that several specific numeric / license claims in those documents are superseded by v2 — `REPORT.md` carries an inline v2-update banner at the top that lists the corrections.