# Typical Marine Ecological Environment Feature Recognition

This project is being refactored from a single green-tide detector into a
registry-driven framework for typical marine ecological environment feature
recognition.

The goal is not a closed multiclass model. Each element, satellite, sensor,
resolution, and fused-product policy can be expressed as a small Markdown
capability card, then composed into a task profile for training or inference.
This keeps the system extensible for green tide, red tide, golden tide,
aquaculture, ships, oil spill, sea ice, and future targets.

## Core Constraint

Do not assume coastline vectors, land-mask vectors, or external GIS masks are
available. Land, black borders, no-data areas, water background, cloud, and other
hard negatives must be handled by labels, validity/context heads, hard-negative
samples, or image-derived validity rules.

## Markdown Capability Registry

Cards live under [docs/registry](docs/registry):

- `elements/`: target feature or object cards.
- `satellites/`: satellite/platform cards.
- `sensors/`: sensor, product, and fusion cards.
- `resolutions/`: spatial-resolution policy cards.

Example for an already fused GF6 product:

```powershell
python scripts/compose_task_profile.py `
  --element green_tide `
  --satellite GF6 `
  --sensor PMS `
  --fusion FUSED_OPTICAL `
  --resolution 2m `
  --output configs/profiles/gf6_green_tide_fused_2m.json
```

Example for GF2/GF1-style PAN+MSS imagery where fusion should happen tile by
tile during inference:

```powershell
python scripts/compose_task_profile.py `
  --element green_tide `
  --satellite GF2 `
  --sensor PMS `
  --fusion STREAM_FUSION `
  --resolution 2m `
  --output configs/profiles/gf2_green_tide_stream_fusion_2m.json
```

## Fused Image Expression

Fused imagery is represented as a derived observation, not as a plain
multispectral raster and not as `is_fused=true`.

Every sample that uses a fused raster or runtime fusion should keep a `fusion`
object in the manifest:

- `state`: `none`, `fused_product`, `runtime_fusion`, or `unknown`.
- `method`: vendor method, implemented method, or `unknown_vendor_product`.
- `sources`: PAN/MSS/source product roles, paths, and native resolutions.
- `target_resolution_m`: output grid resolution.
- `native_multispectral_resolution_m`: original multispectral resolution.
- `persisted`: whether the fused raster exists on disk.
- `reproducible`: whether source data and method can reproduce it.
- `spectral_preservation`: known or estimated spectral preservation risk.

This is important because a GF6 fused image, a GF2 tile-wise fusion stream, and a
native multispectral image should not be treated as identical observations.

## Dataset Standard

The normalized dataset is manifest-driven and may retain different patch sizes,
satellites, sensors, resolutions, and fusion states.

See [docs/dataset_standard.md](docs/dataset_standard.md).

For the SAMPoly-style polygon head, the dataset gate is stricter: only real
mask or polygon annotations enter the main polygon training set. Bbox-only
labels are rejected because rectangles cannot supervise true boundaries,
vertices, or polygon ordering. The polygon-ready layout is:

```text
images/{train,val,test}
masks/{train,val,test}
manifests/accepted_polygon_samples.jsonl
manifests/rejected_polygon_samples.jsonl
dataset_card.json
```

## Key Scripts

- `scripts/build_marine_feature_dataset.py`: scan local or server-side data roots
  and write normalized manifests.
- `scripts/search_hf_marine_datasets.py`: search Hugging Face datasets and write
  normalized `hf://` manifest references without downloading full repositories.
- `scripts/prepare_polygon_dataset.py`: build the strict mask/polygon dataset
  for SAMPoly-style training and reject bbox-only samples.
- `scripts/compose_task_profile.py`: compose Markdown capability cards into a
  JSON task profile.
- `scripts/infer_whole_scene.py`: sliding-window whole-scene inference with
  overlap weighting and optional PAN+MSS runtime fusion.
- `scripts/postprocess_mask.py`: temporary cleanup utility for obvious invalid
  areas; it is not a substitute for the final context/validity model heads.

## Model Direction

The target architecture is a unified, explainable, multi-task model:

- shared DINOv3-style visual encoder;
- clear FPN/PAN multi-scale neck;
- validity head for invalid/no-data areas;
- scene-context head for water, land, cloud/shadow, and other;
- element-specific semantic, instance, detection, polygon, or change heads
  selected by Markdown task profiles.

Current legacy green-tide weights can still be loaded for candidate generation,
but final products should come from the registry-driven multi-task framework.

No bbox-only accepted dataset is retained. New training samples should be
imported only when they provide mask or polygon annotations suitable for the
SAMPoly-style head.