jer47/irp-hz2321-data

IRP seabed data and model archive

This public Hugging Face repository is the single large-asset archive for the IRP project:

https://github.com/ese-ada-lovelace-2025/irp-hz2321

It contains the geospatial source rasters and vectors, derived labels, train/validation/test patch arrays, sediment-v2 data, compact evaluation evidence, and the 26 validation-selected PyTorch checkpoints used by the project. Source code, tests, and the report remain in the GitHub repository.

The repository can be browsed and downloaded without a Hugging Face account.

Repository layout

source_tifs/
  input/
  input_parts50/
  label/
source_vectors/
dataset_metadata/
patches_v1/
  train/{images,labels,labels_non_sediment}/
  val/{images,labels,labels_non_sediment}/
  test/{images,labels,labels_non_sediment}/
sediment_v2/
  labels/{train,val,test}/
  rasters/
  metadata/
evaluation/
  final/
models/
  checkpoints/<task>/<run_name>/
  manifests/

source_tifs/input/

Small aligned GeoTIFF inputs that can be stored as complete files, plus GDAL .aux.xml sidecars for source rasters where present. This directory includes the MAG validity mask, sediment class raster, slope-risk classes, and raster metadata sidecars.

source_tifs/input_parts50/

Nine large aligned GeoTIFF inputs split into consecutive 50 MiB parts. The parts preserve the original file bytes and use zero-padded names such as:

backscatter_aligned_to_sediment_0.5m.tif.part.000
backscatter_aligned_to_sediment_0.5m.tif.part.001
...

The represented rasters are MBES depth, MBES slope, three roughness window sizes, backscatter, SSS HF, cleaned MAG value, and the MAG risk-factor raster. To reconstruct one GeoTIFF after downloading every part for that basename:

cat <filename>.tif.part.* > <filename>.tif

Do not mix parts belonging to different basenames.

source_tifs/label/

Full-area raster labels and derived risk components:

• the project risk map;
• redrawn sediment labels;
• non-sediment risk score and class labels;
• sediment-type component labels.

source_vectors/

GeoPackage vector data used to define spatial train/validation/test regions and the redrawn sediment-label polygons. Polygon/vector files are kept separate from the rasterized labels.

dataset_metadata/

Dataset-level compact metadata:

• patch_manifest_v1.csv: one row per version-1 patch, including split, source paths, raster window, spatial bounds, coverage, and class counts;
• component_label_report_v2.json: summary of the derived component-label construction.

patches_v1/

Version-1 512 x 512 patch arrays for the spatially independent train, validation, and test regions.

Split	Patches
Train	3,531
Validation	167
Test	305

Each split contains:

• images/: compressed NumPy .npz files. The image array has shape (4, 512, 512) and dtype float32; the channels are MBES depth, SSS HF, cleaned MAG value, and MAG-validity mask;
• labels/: complete-risk .npy labels with shape (512, 512), dtype uint8, project classes 0--4, and 255 as ignore/NoData;
• labels_non_sediment/: non-sediment-risk .npy labels with classes 0--3 and 255 as ignore/NoData.

non_sediment_label_patch_report_v2.json records the non-sediment label-patch generation summary.

sediment_v2/labels/

Formal sediment-v2 labels aligned to the patch windows. Each file is a (512, 512) uint8 NumPy array with classes 0 and 1, and 255 as ignore/NoData.

Split	Label arrays
Train	963
Validation	168
Test	327

sediment_v2/rasters/

The three full-area rasters defining the sediment-v2 target and evaluation support:

• sediment-v2 class labels;
• common-validity mask;
• split-membership raster.

sediment_v2/metadata/

Formal sediment-v2 preparation metadata, including the patch manifest, training normalisation, class counts, rare-patch inventory, spatial split audit, and alignment audit.

evaluation/final/

Compact report-facing configurations and metrics. This directory intentionally contains no prediction rasters, figures, full console logs, or failed runs.

• direct_risk/: active direct-risk configurations and test metrics;
• non_sediment_base12/: Base12 configuration and validation/test/contact metrics;
• mag_no_input/: strict no-MAG configuration, metrics, confusion matrices, and target audit;
• sediment_v2/: formal multi-seed configurations, run specification, and per-seed/summary metrics;
• fusion/: original/current fusion metrics on the same reference support;
• risk_map/: compact component-label evidence;
• sss_audit/: contact inventory, selection, visibility, and multimodal audit summaries.

models/checkpoints/

The 26 validation-selected PyTorch checkpoints. Checkpoints are grouped by task and then by the exact training run name so identically named files cannot collide.

Task folder	Checkpoints
direct_risk/	4
non_sediment/	2
mag_ablation/	1
original_fusion/	1
sediment_v2/	18

Only the selected best_* or frozen-validation checkpoints are included. No latest.pt, failed-run, SSS-detector, or exploratory-architecture weights are part of the formal archive. The sediment-v2 roughness seed-1 checkpoint is the formal _retry01 run.

These .pt files were produced by this project. PyTorch checkpoint files can contain pickled objects; load checkpoint files only from a trusted source and use the corresponding architecture and configuration from the GitHub project.

models/manifests/

• huggingface_upload_manifest.csv: the 26 remote checkpoint paths, validation epochs, sizes, archive origins, and direct download URLs;
• model_registry.csv: the matching report model registry.

The two manifests deliberately do not contain a project-generated checksum manifest.

Access and selective download

No authentication is required. Inspect the remote tree:

hf datasets list jer47/irp-hz2321-data --tree -R -h

Download the complete repository:

hf download jer47/irp-hz2321-data --repo-type dataset --local-dir irp-hz2321-data

For normal use, download only the required folder. For example:

hf download jer47/irp-hz2321-data \
  --repo-type dataset \
  --include "models/checkpoints/non_sediment/**" \
  --local-dir irp-hz2321-data

Scope and limitations

This repository is a reproducibility archive for one IRP case study and one survey area. It is not an industry-validated engineering risk product. The project's fused output is a relative screening and decision-support layer, not a substitute for expert geophysical or UXO interpretation.

Disposable caches, failed or exploratory runs, prediction rasters, generated figures, full logs, report build products, and local development archives are not included. Their exclusion keeps the archive tied to the final report and its reproducible inputs, labels, configurations, evidence, and selected models.