DISFOR-S1-Companion / docs /DATA_DICTIONARY.md
grujdin's picture
Stage DISFOR S1 v1.0.0 batch 4
4dac7ae verified
|
Raw History Blame Contribute Delete
17.9 kB

DISFOR Sentinel-1 companion data dictionary

This document describes release schema 1.0.0, as produced by builder package 0.1.0. The generated directory is a local publication candidate. Its exact scope, source hashes, processing policy and counts are recorded in release.json; completion is recorded in RELEASE_COMPLETE.json.

1. Units of observation and scope

Unit Identity Meaning
Catalogue location sample_id An unchanged explicit identifier from original DISFOR. IDs are not consecutive.
Radar track track, for example R160_ASCENDING Relative orbit number plus ascending/descending direction. Platform and absolute orbit are separate properties.
Physical pass at a location sample_id, pass_id A local acquisition group that can contain multiple contributing source slices. It contributes one patch frame.
Annual block block_id A location, track, annual window and native signature; its cube is stored as one NPZ member.
Frame block_id, frame_index A zero-based time index within that annual cube.
Source-scene association block_id, frame_index, source_scene_index A one-based source-slice entry for a frame. Zero in the pixel array denotes no source.

pass_id alone is not a unique key across locations. A satellite scene can appear at multiple catalogue locations, and source-scene table rows are associations rather than a globally deduplicated scene catalogue. More radar observations do not create more independent locations or reference annotations.

Build scopes are explicit:

scope Payload content Interpretation
full All planned blocks for the complete audited catalogue The candidate can describe the full companion, subject to successful construction and verification.
pilot All planned blocks for the configured pilot locations A technical subset for testing local construction and reading.
collected_examples Only the example NPZ blocks included in a collected evidence ZIP An engineering validation product; it does not contain complete trajectories for its locations.

Original metadata and split files preserve the full source catalogue in every scope. In contrast, coverage/locations.parquet, the radar tables and payloads describe the exported subset. A pilot can therefore carry annotations for locations whose radar payloads are absent. Source-catalogue figures, when included, describe all 3,777 audited source locations and say so explicitly.

2. Release layout

Paths below are relative to the completed release root.

Path Content
README.md Generated dataset card with scope, representation, limitations and explicit Hugging Face configurations. Publisher metadata remains to be finalized.
release.json Release identity, scope, schema/builder versions, source hashes, counts, upstream provenance and validation declarations.
RELEASE_COMPLETE.json Completion identity and hashes of the release metadata and checksum list.
checksums.sha256 SHA256 hashes of manifested release files.
metadata/samples.parquet Original DISFOR sampling metadata, copied byte-for-byte.
metadata/labels.parquet Original DISFOR interval annotations, copied byte-for-byte.
metadata/classes.json, metadata/id_mapping.csv Original class descriptions and source-ID mapping, copied byte-for-byte.
metadata/processing_policy.json The audited Sentinel-1 extraction policy.
coverage/locations.parquet One row for each exported location, with coordinates, counts, quality status and split assignments.
coverage/temporal_coverage/part-*.parquet Monthly counts for exported location/track observations.
coverage/source_catalogue_evidence/ Full-source geographic and quality figures and supporting evidence, included when the source-catalogue checksum matches.
radar/pixel_data/part-*.parquet Centre-observation table with explicit stable field types.
radar/blocks/part-*.parquet Annual-cube index, including relative TAR paths, NPZ member offsets and checksums.
radar/source_scenes/part-*.parquet Source-scene associations and retained scene properties.
radar/grids.parquet Grid geometry and an index to the terrain payload for each exported location.
payloads/s1-*.tar Uncompressed TAR shards containing individually compressed annual NPZ cubes.
payloads/terrain.tar Native terrain NPZs, indexed by radar/grids.parquet.
splits/original_disfor/ Original train_ids.json and val_ids.json, unchanged.
splits/project_first_event_v1.json Separate frozen project split, unchanged.
splits/assignments.parquet Full-catalogue mapping of the original and project assignments.
reader/disfor_release/, examples/, docs/ Portable local reader, example and documentation.

The source extractor's checkpoints, completion markers, native CSV duplicates, preview TIFFs and local account configuration are not public payloads. The source audit and builder work directories remain separate from this release directory.

3. Location catalogue

coverage/locations.parquet includes:

Field Meaning
sample_id Original location identifier.
longitude, latitude Geographic WGS 84 point coordinates decoded from the original GeoParquet geometry.
s2_tile, dataset Retained source-catalogue attributes; these do not declare continuous tile coverage.
scope full, pilot or collected_examples.
annual_blocks, track_count Counts in this exported scope.
observations, accepted_centres Recorded frames and frames accepted at reference pixel [16,16].
accepted_centre_fraction accepted_centres / observations; null if there are no observations.
has_accepted_centre, quality_status Whether any centre is accepted; status is has_accepted_centres or no_accepted_centres.
audit_status Status from the accepted source audit. It is distinct from task eligibility.
project_split, spatial_group Existing project assignment and group identity.
original_disfor_split Original assignment, or unassigned when absent from both original lists.

An accepted centre is not a health label. A location with zero accepted centres remains in the catalogue. Its full patch can still contain accepted contextual pixels.

4. Centre observations

The consolidated centre table uses int64, float64, Boolean and string Arrow fields. These are stable export types; the original patch backscatter arrays remain float32. Null numeric values represent unavailable centre values rather than zero.

Fields Type Meaning
sample_id, year, frame_index int64 Location, native annual-block year and zero-based frame position. The annual group follows the physical-pass start year.
block_id, track, pass_id string Relational identity fields described above.
utc string Original ISO UTC observation time. Usually the centre source scene's start time.
timestamp_basis string centre_source_scene_start, or pass_start_no_centre_data when no centre source exists.
pass_first_utc, pass_last_utc string Earliest/latest source-scene start times within the physical-pass group. They are not annotation interval boundaries.
relative_orbit, absolute_orbit int64 Recorded orbit identifiers.
direction, platform string Ascending/descending direction and recorded platform identifier.
source_scene_count, source_scene_index int64 Number of contributing scenes and the centre's one-based source index; zero means no centre source.
source_image_ids, centre_source_image_id string Pipe-separated frame source IDs and centre source ID; the latter is empty when absent. The source-scene table is preferable for relational access.
label_support string reference_pixel_only.
reference_row, reference_col int64 Both 16, using zero-based array indexing.
grid_hash string Retained native grid identity.
VV_sigma0_db, VH_sigma0_db float64 Centre sigma-zero backscatter in dB.
incidence_deg, local_incidence_deg float64 Recorded incidence and derived local-incidence angle, in degrees.
source_valid, angle_ok, terrain_known, layover_risk, shadow_risk, terrain_ok, recommended_valid Boolean Centre quality flags, defined below.
valid_count_3x3 int64 Number of accepted pixels in rows/columns 15–17 around the reference pixel.
valid_fraction_32x32 float64 Accepted pixel count divided by 1,024 for the full patch.
VV_3x3_power_mean_db, VH_3x3_power_mean_db float64 Average of accepted finite values in linear power, converted back to dB; null if none qualify.

The 3×3 values are separate neighbourhood features; they do not replace the centre values or extend its label to neighbouring pixels.

5. Annual NPZ arrays

Each radar NPZ member contains exactly these array families:

Arrays Dtype Shape Meaning
VV_sigma0_db, VH_sigma0_db float32 (T, 32, 32) Sigma-zero radar backscatter in dB; NaN denotes unavailable values.
incidence_deg, local_incidence_deg float32 (T, 32, 32) Incidence and derived local-incidence angles.
source_scene_index uint16 (T, 32, 32) One-based index into that frame's source-scene list; zero means no source.
source_valid, angle_ok, terrain_known, layover_risk, shadow_risk, terrain_ok, recommended_valid bool (T, 32, 32) Per-pixel quality arrays.
utc, pass_id Unicode string (T,) Aligned observation times and physical-pass IDs. String widths vary by native block.
metadata_json Unicode string Scalar JSON containing identity, grid and ordered pass/source metadata, plus release-portability provenance.

T is the number of acquisitions in the block, not a fixed number of days. Native acquisition intervals and gaps are preserved. Pixel [16,16] is the original annotated reference pixel; it is not a patch segmentation mask.

The release builder rewrites only metadata_json: it removes grid.reference_file, retains the reference checksum, adds the upstream dataset identity and records the source NPZ checksum. Every other array's dtype, shape and content bytes are preserved. Therefore native and exported NPZ file hashes differ, while their nonmetadata array content remains unchanged.

6. Quality rules and terrain

The native policy uses:

recommended_valid = source_valid AND angle_ok AND terrain_ok
Flag Native meaning
source_valid Finite VV, VH and incidence, with a nonzero source-scene index.
angle_ok Source-valid observation with incidence between 30° and 45°, inclusive.
terrain_known Finite derived local incidence and terrain elevation. Missing platform heading can make this false even when DEM and radar values exist.
layover_risk, shadow_risk Local geometric screening from DEM derivatives, source-scene heading and incidence. These are model flags rather than independent ground truth.
terrain_ok Geometry known, neither risk flag set, and local incidence strictly below 85°.

The terrain NPZ has elevation_m, slope_deg and aspect_grid_deg, each float32 with shape (32,32). The recorded DEM source is COPERNICUS/DEM/GLO30_2024_1; its fixed use does not make it a contemporaneous terrain observation for every radar date. Terrain derivatives were calculated on the policy's 30 m grid and sampled onto the S2 chip grid.

No radiometric terrain flattening or temporal speckle filtering was applied. Geometric screening is not a radiometric correction. The 10 m grid spacing does not imply independent 10 m radar resolution. The masks contain no freezing, snow, wetness, phenology or physiological-health model.

7. Block, scene, grid and temporal indices

Annual block index

radar/blocks/part-*.parquet stores sample_id, track, year, block_id, native signature, grid_hash, n_frames, n_valid_centres, first_utc and last_utc. Paths and byte positions are:

Field Meaning
archive_path Portable path from the release root to an uncompressed TAR shard.
member_path Portable NPZ name inside that TAR.
byte_offset, byte_size Beginning of the member's data and its byte length. Offsets refer to the published uncompressed TAR bytes.
payload_sha256 SHA256 of the exported NPZ member bytes.
source_npz_sha256, source_parquet_sha256, source_metadata_sha256 Hashes binding the original source files.
array_content_sha256 Builder-defined digest over sorted nonmetadata array names, dtypes, shapes and C-order bytes. It includes timestamps and pass IDs.
quality_counts_json JSON with true-flag counts for centres and all patch pixels. These flag counts can overlap.

Compressing or repacking the outer TAR changes its byte layout and invalidates its offsets and file checksum. Treat each TAR and its indices as one versioned representation.

Source-scene associations

radar/source_scenes/ stores sample_id, track, year, block_id, frame_index, pass_id, source_scene_index, image_id, nullable platform_heading_deg and source_properties_json. Use the array's source index together with the frame identity to recover the contributing scene at a particular pixel.

Grids and terrain index

radar/grids.parquet stores each exported location's grid_hash, CRS, width/height, reference row/column, affine transform, bounds and complete portable grid_json. The affine list is ordered [a,b,c,d,e,f], so projected coordinates follow x=a*col+b*row+c, y=d*col+e*row+f at pixel corners; add 0.5 to row/column for pixel centres.

The fields terrain_archive_path, terrain_member_path, terrain_byte_offset, terrain_byte_size and terrain_sha256 locate and verify the terrain NPZ. source_grid_sha256 binds the original grid JSON. The current reader's load_block loads radar blocks; it does not expose a separate terrain-loading convenience method.

Monthly coverage

coverage/temporal_coverage/ stores sample_id, track, year, month, observations, accepted_centres, source_valid_centres and terrain_unknown_centres. Month is taken from the recorded frame UTC; year is the native annual-block year. Counts describe exported observations. They do not certify that every possible globally available scene was retrieved, nor that an accepted winter observation is ecologically comparable to a summer observation.

8. Original labels, splits and Sentinel-2 joins

The original labels retain sample_id, original_sample_id, dataset, label, original_label, start, end and start_next_label, with the original field types and timestamp values. Class descriptions come from the original classes.json.

A label interval is evidence concerning the reference location during its annotated interval. Do not automatically extend its end to start_next_label, fill annotation gaps with a healthy label, classify all revegetation as undisturbed forest, or label every patch pixel from the centre. Disturbance detection targets and healthy/reference construction require explicit task definitions and event-history treatment.

The original train/validation lists assign 3,398 and 372 catalogue IDs; 259, 262, 265, 270, 280, 286 and 332 remain unassigned in those lists. The separate project split covers all 3,777 IDs with 2,644 train, 567 validation and 566 test assignments. Earlier experiment eligibility, caches and results are not replaced by these catalogue counts.

Join Sentinel-1, original annotations and original Sentinel-2 using sample_id, never row position. Original S2 imagery remains external to this companion: use the original DISFOR archive/reader or the existing local copy. The sensors retain different observation calendars. No automatic one-to-one date join, interpolation, multimodal training target or S1/S2 fusion model is supplied.

9. Hugging Face configurations and local reading

The generated dataset-card YAML declares four table-loading configurations:

Configuration Loading split name Files
locations (default) catalogue coverage/locations.parquet
centre_observations observations radar/pixel_data/*.parquet
annual_blocks catalogue radar/blocks/*.parquet
label_intervals annotations metadata/labels.parquet

These names group files for loading and preview. They do not apply the scientific split JSONs automatically. A learning task must select the appropriate assignment field or ID list explicitly. The original S2 dataloaders do not automatically understand S1 arrays, masks, tracks or TAR byte offsets. Hugging Face's data-file configuration documentation describes the configuration mechanism.

The included CompanionDataset opens a completed local release. iter_blocks streams index rows, load_block reads and verifies a selected NPZ member, read_centre_rows yields Arrow batches, and read_sequence assembles one location/track in chronological order. Date bounds are start-inclusive/end-exclusive. All frames are retained by default; accepted-centre filtering is explicit. See READER_USAGE.md.

The reader does not download remote files or perform training. Download a pinned repository revision before using it. Full catalogue verification and per-block payload verification are distinct operations; opening the reader does not rehash every release file.