Update HF dataset discovery tooling

README.md

@@ -82,6 +82,8 @@ See [docs/dataset_standard.md](docs/dataset_standard.md).
 
 - `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/compose_task_profile.py`: compose Markdown capability cards into a
   JSON task profile.
 - `scripts/infer_whole_scene.py`: sliding-window whole-scene inference with
@@ -102,3 +104,16 @@ The target architecture is a unified, explainable, multi-task model:
 
 Current legacy green-tide weights can still be loaded for candidate generation,
 but final products should come from the registry-driven multi-task framework.
+
+## Hugging Face Dataset Discovery
+
+HF dataset search and standardization results are summarized in
+[docs/hf_dataset_discovery.md](docs/hf_dataset_discovery.md). The generated local
+manifests are under:
+
+```text
+C:\Users\HUAWEI\Documents\New project 2\hf_marine_feature_dataset
+```
+
+The discovery output is conservative: unknown elements and unpaired HF assets are
+kept for review and are not treated as training negatives.

configs/profiles/profiles/gf2_green_tide_stream_fusion_2m.json

@@ -0,0 +1,197 @@
+{
+  "profile_id": "green_tide_GF2_PMS_STREAM_FUSION_2m",
+  "cards": {
+    "elements": {
+      "kind": "elements",
+      "id": "green_tide",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\elements\\green_tide.md",
+      "meta": {
+        "id": "green_tide",
+        "type": "element",
+        "name": "Green tide / Enteromorpha",
+        "task_types": [
+          "semantic_segmentation",
+          "polygon_extraction"
+        ],
+        "preferred_heads": [
+          "validity",
+          "context",
+          "semantic_segmentation"
+        ],
+        "label_formats": [
+          "mask",
+          "polygon",
+          "geojson",
+          "shp"
+        ],
+        "positive_label": "green_tide",
+        "negative_policy": "unlabeled_is_ignore",
+        "typical_sensors": [
+          "PMS",
+          "MUX",
+          "MSS"
+        ],
+        "typical_resolutions_m": [
+          2,
+          8,
+          10
+        ]
+      },
+      "body": "# Green Tide / Enteromorpha\n\nGreen tide is a floating-algae target. It can be confused with land vegetation,\naquaculture structures, cloud edges, turbid water, sunglint, and black/no-data\nregions if the model is trained as a simple foreground/background classifier.\n\n## Required Context\n\n- Validity handling: black borders and no-data pixels must not be interpreted as green tide.\n- Context handling: land and water context should be learned or provided as labels/hard negatives.\n- Do not assume a coastline vector or external land mask exists.\n\n## Recommended Output\n\n- Per-pixel probability map for green tide.\n- Binary candidate mask after validity/context filtering.\n- Optional polygon extraction for product delivery.\n\n## Failure Modes\n\n- Land vegetation is often spectrally similar to algae.\n- Black borders can be falsely predicted when labels do not include invalid regions.\n- Single-element binary models overpredict on unknown ecological elements."
+    },
+    "satellites": {
+      "kind": "satellites",
+      "id": "GF2",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\satellites\\GF2.md",
+      "meta": {
+        "id": "GF2",
+        "type": "satellite",
+        "name": "Gaofen-2",
+        "country": "China",
+        "supported_sensors": [
+          "PMS"
+        ],
+        "typical_products": [
+          "MSS",
+          "PAN",
+          "fused"
+        ]
+      },
+      "body": "# GF2\n\nGF2 PMS scenes commonly provide paired PAN and MSS products. For models targeting\n2 m products, stream PAN+MSS fusion tile-by-tile instead of saving a full fused\nscene to disk.\n\n## Notes\n\n- Keep both `pan_path` and `image_path` in the manifest.\n- Preserve sensor geometry metadata when available."
+    },
+    "sensors": {
+      "kind": "sensors",
+      "id": "PMS",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\sensors\\PMS.md",
+      "meta": {
+        "id": "PMS",
+        "type": "sensor",
+        "name": "Panchromatic and Multispectral Sensor",
+        "modalities": [
+          "optical",
+          "pan",
+          "multispectral"
+        ],
+        "common_bands": [
+          "blue",
+          "green",
+          "red",
+          "nir",
+          "pan"
+        ],
+        "supports_streaming_fusion": true
+      },
+      "body": "# PMS\n\nPMS data may provide separate PAN and MSS products or a fused product. For\nhigh-resolution ecological extraction, prefer stream fusion when PAN and MSS are\nboth available."
+    },
+    "resolutions": {
+      "kind": "resolutions",
+      "id": "2m",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\resolutions\\2m.md",
+      "meta": {
+        "id": "2m",
+        "type": "resolution",
+        "resolution_m": 2,
+        "scale_group": "high",
+        "recommended_patch_sizes": [
+          256,
+          512,
+          1024
+        ]
+      },
+      "body": "# 2 m Resolution\n\nUse for fine optical extraction. Floating algae texture, aquaculture structures,\nand small ships may be visible. Full-scene inference must use tiled/striped IO."
+    },
+    "fusion": {
+      "kind": "sensors",
+      "id": "STREAM_FUSION",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\sensors\\STREAM_FUSION.md",
+      "meta": {
+        "id": "STREAM_FUSION",
+        "type": "sensor",
+        "name": "Streaming PAN+MSS fusion",
+        "modalities": [
+          "optical",
+          "pan",
+          "multispectral",
+          "fused_runtime"
+        ],
+        "common_bands": [
+          "blue",
+          "green",
+          "red",
+          "nir"
+        ],
+        "supports_streaming_fusion": true,
+        "fusion_state": "runtime_fusion",
+        "requires_fusion_metadata": true
+      },
+      "body": "# Streaming PAN+MSS Fusion\n\nStreaming fusion means the fused tile is produced in memory during inference or\ntraining and is not stored as a full-scene raster.\n\n## Required Fusion Metadata\n\n- `fusion_state`: `runtime_fusion`.\n- `fusion_method`: method used by the implementation.\n- `fusion_sources`: at minimum `pan_path` and `mss_path`.\n- `target_resolution_m`: usually the PAN product resolution.\n- `native_resolution_m`: original MSS resolution.\n- `fusion_persisted`: `false`.\n- `fusion_reproducible`: `true` if the source images and code are available.\n- `tile_aligned`: whether PAN and MSS windows are aligned consistently.\n\n## Notes\n\n- Prefer this for very large GF2/GF1-like scenes to avoid writing full fused\n  products.\n- Fusion seams and model tile seams are separate problems; use overlap-weighted\n  inference after fusion."
+    }
+  },
+  "task": {
+    "element": "green_tide",
+    "task_types": [
+      "semantic_segmentation",
+      "polygon_extraction"
+    ],
+    "preferred_heads": [
+      "validity",
+      "context",
+      "semantic_segmentation"
+    ],
+    "label_formats": [
+      "mask",
+      "polygon",
+      "geojson",
+      "shp"
+    ],
+    "negative_policy": "unlabeled_is_ignore"
+  },
+  "input": {
+    "satellite": "GF2",
+    "sensor": "PMS",
+    "resolution_m": 2,
+    "modalities": [
+      "optical",
+      "pan",
+      "multispectral"
+    ],
+    "common_bands": [
+      "blue",
+      "green",
+      "red",
+      "nir",
+      "pan"
+    ],
+    "recommended_patch_sizes": [
+      256,
+      512,
+      1024
+    ]
+  },
+  "fusion": {
+    "card": "STREAM_FUSION",
+    "state": "runtime_fusion",
+    "modalities": [
+      "optical",
+      "pan",
+      "multispectral",
+      "fused_runtime"
+    ],
+    "supports_streaming_fusion": true,
+    "requires_fusion_metadata": true,
+    "required_manifest_fields": [
+      "state",
+      "method",
+      "sources",
+      "target_resolution_m",
+      "native_multispectral_resolution_m",
+      "persisted",
+      "reproducible",
+      "spectral_preservation"
+    ]
+  },
+  "constraints": {
+    "do_not_assume_external_coastline_or_land_mask": true,
+    "unlabeled_elements_are_ignore_not_negative": true
+  }
+}

configs/profiles/profiles/gf6_green_tide_fused_2m.json

@@ -0,0 +1,197 @@
+{
+  "profile_id": "green_tide_GF6_PMS_FUSED_OPTICAL_2m",
+  "cards": {
+    "elements": {
+      "kind": "elements",
+      "id": "green_tide",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\elements\\green_tide.md",
+      "meta": {
+        "id": "green_tide",
+        "type": "element",
+        "name": "Green tide / Enteromorpha",
+        "task_types": [
+          "semantic_segmentation",
+          "polygon_extraction"
+        ],
+        "preferred_heads": [
+          "validity",
+          "context",
+          "semantic_segmentation"
+        ],
+        "label_formats": [
+          "mask",
+          "polygon",
+          "geojson",
+          "shp"
+        ],
+        "positive_label": "green_tide",
+        "negative_policy": "unlabeled_is_ignore",
+        "typical_sensors": [
+          "PMS",
+          "MUX",
+          "MSS"
+        ],
+        "typical_resolutions_m": [
+          2,
+          8,
+          10
+        ]
+      },
+      "body": "# Green Tide / Enteromorpha\n\nGreen tide is a floating-algae target. It can be confused with land vegetation,\naquaculture structures, cloud edges, turbid water, sunglint, and black/no-data\nregions if the model is trained as a simple foreground/background classifier.\n\n## Required Context\n\n- Validity handling: black borders and no-data pixels must not be interpreted as green tide.\n- Context handling: land and water context should be learned or provided as labels/hard negatives.\n- Do not assume a coastline vector or external land mask exists.\n\n## Recommended Output\n\n- Per-pixel probability map for green tide.\n- Binary candidate mask after validity/context filtering.\n- Optional polygon extraction for product delivery.\n\n## Failure Modes\n\n- Land vegetation is often spectrally similar to algae.\n- Black borders can be falsely predicted when labels do not include invalid regions.\n- Single-element binary models overpredict on unknown ecological elements."
+    },
+    "satellites": {
+      "kind": "satellites",
+      "id": "GF6",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\satellites\\GF6.md",
+      "meta": {
+        "id": "GF6",
+        "type": "satellite",
+        "name": "Gaofen-6",
+        "country": "China",
+        "supported_sensors": [
+          "PMS",
+          "MUX",
+          "WFV"
+        ],
+        "typical_products": [
+          "MSS",
+          "PAN",
+          "fused"
+        ]
+      },
+      "body": "# GF6\n\nGF6 fused PMS/MUX products can be used directly for full-scene sliding-window\ninference. Large scenes may exceed tens of GB, so stream inference and output by\nstripe.\n\n## Notes\n\n- Use overlap-weighted windows to reduce tile seams.\n- Do not write full probability rasters unless explicitly needed."
+    },
+    "sensors": {
+      "kind": "sensors",
+      "id": "PMS",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\sensors\\PMS.md",
+      "meta": {
+        "id": "PMS",
+        "type": "sensor",
+        "name": "Panchromatic and Multispectral Sensor",
+        "modalities": [
+          "optical",
+          "pan",
+          "multispectral"
+        ],
+        "common_bands": [
+          "blue",
+          "green",
+          "red",
+          "nir",
+          "pan"
+        ],
+        "supports_streaming_fusion": true
+      },
+      "body": "# PMS\n\nPMS data may provide separate PAN and MSS products or a fused product. For\nhigh-resolution ecological extraction, prefer stream fusion when PAN and MSS are\nboth available."
+    },
+    "resolutions": {
+      "kind": "resolutions",
+      "id": "2m",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\resolutions\\2m.md",
+      "meta": {
+        "id": "2m",
+        "type": "resolution",
+        "resolution_m": 2,
+        "scale_group": "high",
+        "recommended_patch_sizes": [
+          256,
+          512,
+          1024
+        ]
+      },
+      "body": "# 2 m Resolution\n\nUse for fine optical extraction. Floating algae texture, aquaculture structures,\nand small ships may be visible. Full-scene inference must use tiled/striped IO."
+    },
+    "fusion": {
+      "kind": "sensors",
+      "id": "FUSED_OPTICAL",
+      "path": "C:\\Users\\HUAWEI\\Documents\\New project 2\\seaweed-detection\\docs\\registry\\sensors\\FUSED_OPTICAL.md",
+      "meta": {
+        "id": "FUSED_OPTICAL",
+        "type": "sensor",
+        "name": "Fused optical product",
+        "modalities": [
+          "optical",
+          "multispectral",
+          "fused"
+        ],
+        "common_bands": [
+          "blue",
+          "green",
+          "red",
+          "nir"
+        ],
+        "supports_streaming_fusion": false,
+        "fusion_state": "fused_product",
+        "requires_fusion_metadata": true
+      },
+      "body": "# Fused Optical Product\n\nA fused optical product is not simply \"multispectral\". It is a derived product\nwhose spatial grid, spectral consistency, and artifacts depend on the fusion\nsource and method.\n\n## Required Fusion Metadata\n\nRecord these fields whenever a fused image is used:\n\n- `fusion_state`: `fused_product`.\n- `fusion_method`: known method name, or `unknown_vendor_product`.\n- `fusion_sources`: source product roles such as `PAN` and `MSS`.\n- `target_resolution_m`: output resolution after fusion.\n- `native_resolution_m`: original multispectral resolution if known.\n- `fusion_persisted`: `true` when the fused image exists on disk.\n- `fusion_reproducible`: `true` only if the source PAN/MSS and method are available.\n- `spectral_preservation`: `unknown`, `low`, `medium`, or `high`.\n\n## Notes\n\n- Do not assume fused products are radiometrically equivalent across satellites\n  or vendors.\n- If the fusion method is unknown, preserve that uncertainty in the profile and\n  dataset manifest."
+    }
+  },
+  "task": {
+    "element": "green_tide",
+    "task_types": [
+      "semantic_segmentation",
+      "polygon_extraction"
+    ],
+    "preferred_heads": [
+      "validity",
+      "context",
+      "semantic_segmentation"
+    ],
+    "label_formats": [
+      "mask",
+      "polygon",
+      "geojson",
+      "shp"
+    ],
+    "negative_policy": "unlabeled_is_ignore"
+  },
+  "input": {
+    "satellite": "GF6",
+    "sensor": "PMS",
+    "resolution_m": 2,
+    "modalities": [
+      "optical",
+      "pan",
+      "multispectral"
+    ],
+    "common_bands": [
+      "blue",
+      "green",
+      "red",
+      "nir",
+      "pan"
+    ],
+    "recommended_patch_sizes": [
+      256,
+      512,
+      1024
+    ]
+  },
+  "fusion": {
+    "card": "FUSED_OPTICAL",
+    "state": "fused_product",
+    "modalities": [
+      "optical",
+      "multispectral",
+      "fused"
+    ],
+    "supports_streaming_fusion": false,
+    "requires_fusion_metadata": true,
+    "required_manifest_fields": [
+      "state",
+      "method",
+      "sources",
+      "target_resolution_m",
+      "native_multispectral_resolution_m",
+      "persisted",
+      "reproducible",
+      "spectral_preservation"
+    ]
+  },
+  "constraints": {
+    "do_not_assume_external_coastline_or_land_mask": true,
+    "unlabeled_elements_are_ignore_not_negative": true
+  }
+}

docs/hf_dataset_discovery.md

@@ -0,0 +1,81 @@
+# Hugging Face Dataset Discovery
+
+This report summarizes the Hugging Face dataset discovery pass for typical
+marine ecological environment feature recognition.
+
+Generated manifests are stored locally at:
+
+```text
+C:\Users\HUAWEI\Documents\New project 2\hf_marine_feature_dataset\
+  manifests\
+    hf_assets_raw.jsonl
+    samples.jsonl
+    samples_ready.jsonl
+    samples_review.jsonl
+  reports\
+    hf_dataset_inventory.csv
+    hf_dataset_summary.json
+    hf_dataset_discovery.md
+```
+
+The manifests use `hf://datasets/<repo>/<path>` references. Full image files are
+not downloaded by the discovery script.
+
+## Current Discovery Result
+
+- Repositories reviewed: 108
+- HF assets indexed: 18,781
+- Standardized sample references: 6,915
+- Training-ready samples: 0
+- Review samples: 6,915
+
+Samples by inferred element:
+
+| Element | Samples |
+| --- | ---: |
+| aquaculture | 1,869 |
+| ship | 641 |
+| sea_ice | 52 |
+| green_tide | 2 |
+| unknown / needs review | 4,351 |
+
+## Why Training-Ready Is Zero
+
+The script is intentionally conservative. A sample is marked training-ready only
+when all of these are true:
+
+- the target element is known;
+- an image is present;
+- a paired mask is present;
+- the pair can be matched from repository file paths without guessing.
+
+Most Hugging Face repositories found in this pass store labels as CSV, parquet,
+YOLO text, RLE, archives, or custom benchmark structures. Those assets are
+indexed in `hf_assets_raw.jsonl`, but they require dataset-specific adapters
+before they should be used for training.
+
+## High-Value Adapter Targets
+
+- `datadrivenscience/ship-detection`: ship imagery with tabular annotation
+  assets; likely needs CSV/RLE or detection-format conversion.
+- `reglab/aquaculture_detection`: aquaculture orthophoto images; likely needs
+  object-detection label parsing.
+- `paperupload/seaicebench_sample` and `paperupload/SEAICEBENCH`: sea-ice
+  benchmark assets with multi-band image products and tabular data.
+- `cuibinge/marine-ecological-feature-dataset`: current project assets and GF
+  scene inputs; already follows the no-external-coastline assumption.
+- `cuibinge/Multimodal-Sea-Land`: large archive asset useful for land/sea
+  context, but it should be unpacked and mapped by an explicit adapter.
+
+## Exclusions
+
+The discovery script excludes obvious false positives such as medical vessel
+segmentation, retinal images, legal/text datasets, anime image datasets, and
+health/community datasets that happen to contain words like `vessel`, `ocean`,
+or `oil spill`.
+
+## Policy
+
+Unknown elements and unpaired assets are not treated as negatives. They remain in
+`samples_review.jsonl` until a card or adapter gives them explicit semantics.
+No coastline vector, land-mask vector, or external GIS prior is assumed.

docs/registry/registry/README.md

@@ -0,0 +1,95 @@
+# Markdown Capability Registry
+
+This directory stores composable Markdown capability cards. A card is a small
+human-readable configuration unit with YAML front matter and operational notes.
+
+Cards are intentionally separated by concern:
+
+- `elements/`: target feature or object type, such as green tide, ship, oil spill, sea ice.
+- `satellites/`: satellite/platform assumptions.
+- `sensors/`: sensor or product type assumptions.
+- `resolutions/`: spatial-resolution policies.
+
+Extraction and training jobs should compose the cards they need instead of using
+a single fixed class list. For example:
+
+```powershell
+python scripts/compose_task_profile.py `
+  --element green_tide `
+  --satellite GF6 `
+  --sensor PMS `
+  --resolution 2m `
+  --output configs/profiles/gf6_green_tide_2m.json
+```
+
+For already fused GF6-like products, inject a fused product card:
+
+```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
+```
+
+For raw PAN+MSS products where fusion should happen tile by tile during
+inference, inject the streaming fusion card:
+
+```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
+```
+
+## Card Format
+
+Each card starts with YAML front matter:
+
+```markdown
+---
+id: green_tide
+type: element
+task_types: [semantic_segmentation]
+preferred_heads: [validity, context, semantic_segmentation]
+label_formats: [mask, polygon]
+---
+
+# Green Tide
+
+Human-readable notes, rules, constraints, and known failure modes.
+```
+
+The front matter is parsed by `scripts/compose_task_profile.py`. The Markdown body
+is kept in the output profile so that training/inference logs preserve the
+reasoning and caveats behind the selected cards.
+
+## Fused Image Expression
+
+Fused imagery is represented as a derived observation, not as a plain
+multispectral image and not as a boolean flag. The selected profile and dataset
+manifest must keep a `fusion` object with:
+
+- `state`: `none`, `fused_product`, `runtime_fusion`, or `unknown`.
+- `method`: known method, implementation method, or `unknown_vendor_product`.
+- `sources`: source roles and paths such as PAN and MSS.
+- `target_resolution_m`: output grid resolution.
+- `native_multispectral_resolution_m`: original multispectral resolution.
+- `persisted`: whether the fused image exists on disk.
+- `reproducible`: whether the source data and method can reproduce it.
+- `spectral_preservation`: known or estimated spectral preservation risk.
+
+This keeps GF6 supplied fused rasters, GF1/GF2 tile-wise fusion, and future
+fusion algorithms comparable without pretending they are identical inputs.
+
+## Hard Constraint
+
+Do not assume coastline vectors, land-mask vectors, or external GIS layers are
+available. If a task needs land, water, invalid, or cloud handling, it must be
+expressed as context labels, hard-negative samples, validity rules, or model
+heads.

docs/registry/registry/elements/aquaculture.md

@@ -0,0 +1,24 @@
+---
+id: aquaculture
+type: element
+name: Aquaculture area and facilities
+task_types: [semantic_segmentation, instance_segmentation, change_detection, polygon_extraction]
+preferred_heads: [validity, context, semantic_segmentation, instance, change_detection]
+label_formats: [mask, bbox, polygon, coco, geojson, shp]
+positive_label: aquaculture
+negative_policy: unlabeled_is_ignore
+typical_sensors: [PMS, MUX, MSI, SAR]
+typical_resolutions_m: [1, 2, 10]
+---
+
+# Aquaculture
+
+Aquaculture may appear as dense rafts, cages, ponds, lines, grids, or coastal
+facilities. It is not only a semantic segmentation problem; some datasets may
+require instance, polygon, or change-detection heads.
+
+## Required Context
+
+- Ships, docks, land structures, and waves are hard negatives.
+- Scale is critical; preserve resolution and patch-size metadata.
+

docs/registry/registry/elements/golden_tide.md

@@ -0,0 +1,24 @@
+---
+id: golden_tide
+type: element
+name: Golden tide / Sargassum
+task_types: [semantic_segmentation, polygon_extraction]
+preferred_heads: [validity, context, semantic_segmentation]
+label_formats: [mask, polygon]
+positive_label: golden_tide
+negative_policy: unlabeled_is_ignore
+typical_sensors: [PMS, MUX, MSS, MSI]
+typical_resolutions_m: [2, 10]
+---
+
+# Golden Tide / Sargassum
+
+Golden tide is a floating vegetation-like target. It shares confusion sources
+with green tide but may have different color, texture, and morphology.
+
+## Required Context
+
+- Treat unlabeled ecological elements as ignore, not as negatives.
+- Keep original scale metadata because raft-like texture and floating patches
+  change strongly with resolution.
+

docs/registry/registry/elements/green_tide.md

@@ -0,0 +1,37 @@
+---
+id: green_tide
+type: element
+name: Green tide / Enteromorpha
+task_types: [semantic_segmentation, polygon_extraction]
+preferred_heads: [validity, context, semantic_segmentation]
+label_formats: [mask, polygon, geojson, shp]
+positive_label: green_tide
+negative_policy: unlabeled_is_ignore
+typical_sensors: [PMS, MUX, MSS]
+typical_resolutions_m: [2, 8, 10]
+---
+
+# Green Tide / Enteromorpha
+
+Green tide is a floating-algae target. It can be confused with land vegetation,
+aquaculture structures, cloud edges, turbid water, sunglint, and black/no-data
+regions if the model is trained as a simple foreground/background classifier.
+
+## Required Context
+
+- Validity handling: black borders and no-data pixels must not be interpreted as green tide.
+- Context handling: land and water context should be learned or provided as labels/hard negatives.
+- Do not assume a coastline vector or external land mask exists.
+
+## Recommended Output
+
+- Per-pixel probability map for green tide.
+- Binary candidate mask after validity/context filtering.
+- Optional polygon extraction for product delivery.
+
+## Failure Modes
+
+- Land vegetation is often spectrally similar to algae.
+- Black borders can be falsely predicted when labels do not include invalid regions.
+- Single-element binary models overpredict on unknown ecological elements.
+

docs/registry/registry/elements/oil_spill.md

@@ -0,0 +1,25 @@
+---
+id: oil_spill
+type: element
+name: Oil spill
+task_types: [semantic_segmentation, anomaly_detection]
+preferred_heads: [validity, context, semantic_segmentation]
+label_formats: [mask, polygon]
+positive_label: oil_spill
+negative_policy: unlabeled_is_ignore
+typical_sensors: [SAR, MSI, OLI]
+typical_resolutions_m: [10, 30]
+---
+
+# Oil Spill
+
+Oil spill extraction is often sensor-specific. SAR imagery is especially useful,
+but dark slicks can be confused with low wind, biogenic films, rain cells, and
+calm water.
+
+## Required Context
+
+- Sensor type must be explicit.
+- Use hard negatives for natural dark water phenomena.
+- Do not treat unlabeled water anomalies as definite negatives.
+

docs/registry/registry/elements/red_tide.md

@@ -0,0 +1,29 @@
+---
+id: red_tide
+type: element
+name: Red tide / harmful algal bloom
+task_types: [semantic_segmentation, classification]
+preferred_heads: [validity, context, semantic_segmentation]
+label_formats: [mask, polygon, image_label]
+positive_label: red_tide
+negative_policy: unlabeled_is_ignore
+typical_sensors: [PMS, MUX, MSS, MSI, OLI]
+typical_resolutions_m: [10, 30]
+---
+
+# Red Tide
+
+Red tide is often a broad water-color anomaly rather than a sharp object. It may
+need both local texture/color features and broader context.
+
+## Required Context
+
+- Water background and turbid-water hard negatives are important.
+- Clouds, haze, sunglint, and sensor striping should be explicit context or ignored.
+- Do not assume external water/land masks.
+
+## Recommended Output
+
+- Pixel-level red-tide probability for segmentation data.
+- Scene or patch-level likelihood when only weak labels are available.
+

docs/registry/registry/elements/sea_ice.md

@@ -0,0 +1,23 @@
+---
+id: sea_ice
+type: element
+name: Sea ice
+task_types: [semantic_segmentation, classification, change_detection]
+preferred_heads: [validity, context, semantic_segmentation, change_detection]
+label_formats: [mask, polygon, image_label]
+positive_label: sea_ice
+negative_policy: unlabeled_is_ignore
+typical_sensors: [SAR, MSI, OLI]
+typical_resolutions_m: [10, 30]
+---
+
+# Sea Ice
+
+Sea ice can be mapped from optical or SAR data, but the feature appearance changes
+with sensor, illumination, weather, and season.
+
+## Required Context
+
+- Clouds, snow, bright land, and sea ice must not be collapsed into one class.
+- Sensor-specific normalization and augmentations are required.
+

docs/registry/registry/elements/ship.md

@@ -0,0 +1,24 @@
+---
+id: ship
+type: element
+name: Ship
+task_types: [detection, instance_segmentation, oriented_detection]
+preferred_heads: [validity, context, detection, instance]
+label_formats: [bbox, rotated_bbox, polygon, coco, dota]
+positive_label: ship
+negative_policy: unlabeled_is_ignore
+typical_sensors: [PMS, PAN, SAR]
+typical_resolutions_m: [0.5, 1, 2, 3, 10]
+---
+
+# Ship
+
+Ships are object targets and should not be forced into a pure semantic
+segmentation class. Use detection, oriented detection, or instance segmentation
+when annotations support it.
+
+## Required Context
+
+- Ports, aquaculture grids, bridges, and bright wave wakes are hard negatives.
+- SAR and optical ship extraction should be sensor-aware.
+

docs/registry/registry/resolutions/10m.md

@@ -0,0 +1,13 @@
+---
+id: 10m
+type: resolution
+resolution_m: 10
+scale_group: medium
+recommended_patch_sizes: [128, 256, 512]
+---
+
+# 10 m Resolution
+
+Common for Sentinel-2 optical products and some ecological element monitoring
+tasks. Good for broad blooms and sea-ice context; limited for small ships.
+

docs/registry/registry/resolutions/2m.md

@@ -0,0 +1,13 @@
+---
+id: 2m
+type: resolution
+resolution_m: 2
+scale_group: high
+recommended_patch_sizes: [256, 512, 1024]
+---
+
+# 2 m Resolution
+
+Use for fine optical extraction. Floating algae texture, aquaculture structures,
+and small ships may be visible. Full-scene inference must use tiled/striped IO.
+

docs/registry/registry/resolutions/30m.md

@@ -0,0 +1,13 @@
+---
+id: 30m
+type: resolution
+resolution_m: 30
+scale_group: coarse
+recommended_patch_sizes: [128, 256]
+---
+
+# 30 m Resolution
+
+Use for broad-scale monitoring. Object-level targets such as small ships or
+aquaculture details are usually unsuitable at this scale.
+

docs/registry/registry/resolutions/8m.md

@@ -0,0 +1,13 @@
+---
+id: 8m
+type: resolution
+resolution_m: 8
+scale_group: medium
+recommended_patch_sizes: [128, 256, 512]
+---
+
+# 8 m Resolution
+
+Use for multispectral products where spatial detail is coarser. Small objects may
+be missed, but broad ecological elements can still be detected.
+

docs/registry/registry/satellites/GF1.md

@@ -0,0 +1,19 @@
+---
+id: GF1
+type: satellite
+name: Gaofen-1
+country: China
+supported_sensors: [PMS, WFV]
+typical_products: [MSS, PAN, fused]
+---
+
+# GF1
+
+GF1 data may include multispectral and panchromatic products depending on the
+sensor and acquisition package. Do not assume a PAN image exists for every scene.
+
+## Notes
+
+- Preserve the original product metadata and RPB/XML side files when available.
+- If PAN is missing, do not run pan-sharpened inference.
+

docs/registry/registry/satellites/GF2.md

@@ -0,0 +1,20 @@
+---
+id: GF2
+type: satellite
+name: Gaofen-2
+country: China
+supported_sensors: [PMS]
+typical_products: [MSS, PAN, fused]
+---
+
+# GF2
+
+GF2 PMS scenes commonly provide paired PAN and MSS products. For models targeting
+2 m products, stream PAN+MSS fusion tile-by-tile instead of saving a full fused
+scene to disk.
+
+## Notes
+
+- Keep both `pan_path` and `image_path` in the manifest.
+- Preserve sensor geometry metadata when available.
+

docs/registry/registry/satellites/GF6.md

@@ -0,0 +1,20 @@
+---
+id: GF6
+type: satellite
+name: Gaofen-6
+country: China
+supported_sensors: [PMS, MUX, WFV]
+typical_products: [MSS, PAN, fused]
+---
+
+# GF6
+
+GF6 fused PMS/MUX products can be used directly for full-scene sliding-window
+inference. Large scenes may exceed tens of GB, so stream inference and output by
+stripe.
+
+## Notes
+
+- Use overlap-weighted windows to reduce tile seams.
+- Do not write full probability rasters unless explicitly needed.
+

docs/registry/registry/satellites/SAR_Generic.md

@@ -0,0 +1,14 @@
+---
+id: SAR_Generic
+type: satellite
+name: Generic SAR platform
+supported_sensors: [SAR]
+typical_products: [GRD, SLC, calibrated_backscatter]
+---
+
+# Generic SAR Platform
+
+Use this card when the platform is not yet modeled as a dedicated satellite card
+but the source is SAR. SAR-specific normalization and speckle handling are
+required.
+

docs/registry/registry/satellites/Sentinel2.md

@@ -0,0 +1,19 @@
+---
+id: Sentinel2
+type: satellite
+name: Sentinel-2
+country: EU
+supported_sensors: [MSI]
+typical_products: [L1C, L2A]
+---
+
+# Sentinel-2
+
+Sentinel-2 MSI provides multispectral optical data with mixed 10 m, 20 m, and
+60 m bands. Resampling policy must be explicit in the task profile.
+
+## Notes
+
+- Record which bands are used and their native resolutions.
+- Cloud/haze handling is important for optical ecological elements.
+

docs/registry/registry/sensors/FUSED_OPTICAL.md

@@ -0,0 +1,37 @@
+---
+id: FUSED_OPTICAL
+type: sensor
+name: Fused optical product
+modalities: [optical, multispectral, fused]
+common_bands: [blue, green, red, nir]
+supports_streaming_fusion: false
+fusion_state: fused_product
+requires_fusion_metadata: true
+---
+
+# Fused Optical Product
+
+A fused optical product is not simply "multispectral". It is a derived product
+whose spatial grid, spectral consistency, and artifacts depend on the fusion
+source and method.
+
+## Required Fusion Metadata
+
+Record these fields whenever a fused image is used:
+
+- `fusion_state`: `fused_product`.
+- `fusion_method`: known method name, or `unknown_vendor_product`.
+- `fusion_sources`: source product roles such as `PAN` and `MSS`.
+- `target_resolution_m`: output resolution after fusion.
+- `native_resolution_m`: original multispectral resolution if known.
+- `fusion_persisted`: `true` when the fused image exists on disk.
+- `fusion_reproducible`: `true` only if the source PAN/MSS and method are available.
+- `spectral_preservation`: `unknown`, `low`, `medium`, or `high`.
+
+## Notes
+
+- Do not assume fused products are radiometrically equivalent across satellites
+  or vendors.
+- If the fusion method is unknown, preserve that uncertainty in the profile and
+  dataset manifest.
+

docs/registry/registry/sensors/MSS.md

@@ -0,0 +1,14 @@
+---
+id: MSS
+type: sensor
+name: Multispectral scanner/product
+modalities: [optical, multispectral]
+common_bands: [blue, green, red, nir]
+supports_streaming_fusion: true
+---
+
+# MSS
+
+MSS products are usually lower resolution than PAN. If the target model expects a
+PAN-scale product, pair with a PAN card/product and fuse during tile inference.
+

docs/registry/registry/sensors/MUX.md

@@ -0,0 +1,14 @@
+---
+id: MUX
+type: sensor
+name: Multispectral camera
+modalities: [optical, multispectral]
+common_bands: [blue, green, red, nir]
+supports_streaming_fusion: false
+---
+
+# MUX
+
+MUX-like products are multispectral. If already fused to a high-resolution grid,
+they can be used directly; otherwise record the native spatial resolution.
+

docs/registry/registry/sensors/PAN.md

@@ -0,0 +1,15 @@
+---
+id: PAN
+type: sensor
+name: Panchromatic product
+modalities: [optical, panchromatic]
+common_bands: [pan]
+supports_streaming_fusion: true
+---
+
+# PAN
+
+PAN products provide spatial detail but not the multispectral information needed
+by most ecological element heads. Use as a fusion source, not as a direct
+four-band model input.
+

docs/registry/registry/sensors/PMS.md

@@ -0,0 +1,15 @@
+---
+id: PMS
+type: sensor
+name: Panchromatic and Multispectral Sensor
+modalities: [optical, pan, multispectral]
+common_bands: [blue, green, red, nir, pan]
+supports_streaming_fusion: true
+---
+
+# PMS
+
+PMS data may provide separate PAN and MSS products or a fused product. For
+high-resolution ecological extraction, prefer stream fusion when PAN and MSS are
+both available.
+

docs/registry/registry/sensors/SAR.md

@@ -0,0 +1,15 @@
+---
+id: SAR
+type: sensor
+name: Synthetic Aperture Radar
+modalities: [radar]
+common_bands: [vv, vh, hh, hv]
+supports_streaming_fusion: false
+---
+
+# SAR
+
+SAR is useful for ships, oil spills, sea ice, and all-weather monitoring. It
+requires sensor-specific normalization and is not interchangeable with optical
+four-band products.
+

docs/registry/registry/sensors/STREAM_FUSION.md

@@ -0,0 +1,34 @@
+---
+id: STREAM_FUSION
+type: sensor
+name: Streaming PAN+MSS fusion
+modalities: [optical, pan, multispectral, fused_runtime]
+common_bands: [blue, green, red, nir]
+supports_streaming_fusion: true
+fusion_state: runtime_fusion
+requires_fusion_metadata: true
+---
+
+# Streaming PAN+MSS Fusion
+
+Streaming fusion means the fused tile is produced in memory during inference or
+training and is not stored as a full-scene raster.
+
+## Required Fusion Metadata
+
+- `fusion_state`: `runtime_fusion`.
+- `fusion_method`: method used by the implementation.
+- `fusion_sources`: at minimum `pan_path` and `mss_path`.
+- `target_resolution_m`: usually the PAN product resolution.
+- `native_resolution_m`: original MSS resolution.
+- `fusion_persisted`: `false`.
+- `fusion_reproducible`: `true` if the source images and code are available.
+- `tile_aligned`: whether PAN and MSS windows are aligned consistently.
+
+## Notes
+
+- Prefer this for very large GF2/GF1-like scenes to avoid writing full fused
+  products.
+- Fusion seams and model tile seams are separate problems; use overlap-weighted
+  inference after fusion.
+

typical-marine-ecological-feature-recognition-source.zip

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:9bdb0199f1f362178a058128bb6d577612bd3a5600ce5083cb563bbf3acdd81f
-size 426619
+oid sha256:9289bdcbd466df83de2b014697ced6d50bfbb2123ecf074449e5ade894602939
+size 434189