docs: refresh README with panel preview, full CDL legend, and coverage

README.md

@@ -1,259 +1,253 @@
 ---
-pretty_name: NAIP 16‑day city cubes
-license: apache-2.0
+pretty_name: NAIP 16-Day City Cubes (materialized tiles)
 tags:
 - geospatial
-- satellite
 - remote-sensing
-- naip
-- sentinel-2
-- dem
+- earth-observation
+- wildfire
+- segmentation
 task_categories:
-- image-classification
 - image-segmentation
-datasets:
-- datasets
-language:
-- en
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-dataset_info:
-  features:
-  - name: tile_id
-    dtype: string
-  - name: city
-    dtype: string
-  - name: west
-    dtype: float64
-  - name: south
-    dtype: float64
-  - name: east
-    dtype: float64
-  - name: north
-    dtype: float64
-  - name: chip_px
-    dtype: int64
-  - name: split
-    dtype: string
-  - name: meta_json
-    dtype: string
-  - name: naip_rgb
-    dtype: image
-  - name: naip_ndvi
-    dtype: image
-  - name: s2_rgb
-    dtype: image
-  - name: s2_pseudo_rgb
-    dtype: image
-  - name: dem_rgb
-    dtype: image
-  - name: labels
-    dtype: image
-  - name: landfire_family
-    dtype: image
-  - name: cdl
-    dtype: image
-  - name: s2_B04
-    dtype: image
-  - name: s2_B03
-    dtype: image
-  - name: s2_B02
-    dtype: image
-  - name: s2_B08
-    dtype: image
-  - name: s2_MSAVI
-    dtype: image
-  - name: s2_NDVI
-    dtype: image
-  - name: s2_NDWI
-    dtype: image
-  - name: s2_SCL
-    dtype: image
-  splits:
-  - name: train
-    num_bytes: 34524111669.0
-    num_examples: 24792
-  download_size: 34373796482
-  dataset_size: 34524111669.0
+- image-classification
+license: cc-by-4.0
 ---
 
-# NAIP 16-day City Cubes — Earth Observation for Fire Risk
-
-High-fidelity **Parquet shards** of multi-sensor tiles grouped by `town`.
-Each row is a single **512×512** tile with synchronized **NAIP**, **Sentinel-2**, **DEM**,
-plus categorical layers (**LANDFIRE**, **CDL**) and a per-pixel **semantic `labels`** mask.
-
-Shards live under:
-
-```
-data/city=<name>/part-*.parquet
-```
-
-> Built to train a **four-headed fire-risk model**: (1) segmentation on `labels`,
-  (2) land-cover family classification, (3) CDL/aux targets, and (4) a risk head you attach downstream.
-
-## Snapshot
-
-- **Shards:** 691
-- **Rows (tiles):** 24,792
-- **On-disk size:** 44.05 GB
-- **Tile size:** 512px × 512px
-- **S2 pseudo-RGB recipe:** e.g., `MSAVI,B08,NDMI`
+# NAIP 16-Day City Cubes (materialized tiles)
 
+![Panel preview](assets/panels/santa_maria_3x3.png)
 
-## Columns
+Each row is a 512×512 chip with **16 layers** exposed through logical groupings:
 
-- `tile_id` — Unique tile identifier.
-- `town` — Town used for grouping on the Hub.
-- `x` — Slippy-tile X (int).
-- `y` — Slippy-tile Y (int).
-- `z` — Slippy-tile zoom level (int).
-- `west` — Tile western bound (lon, float).
-- `south` — Tile southern bound (lat, float).
-- `east` — Tile eastern bound (lon, float).
-- `north` — Tile northern bound (lat, float).
-- `cnt` — Internal count / provenance counter.
-- `chip_px` — Tile side length in pixels (H=W).
-- `s2_year` — Sentinel-2 target year for the 16-day window.
-- `s2_month` — Sentinel-2 target month ('01'..'12').
-- `s2_month_window` — ±days window around month for composing S2.
-- `s2_time_slices` — Number of S2 scenes contributing to composite.
-- `s2_cloud_cap` — Cloud coverage cap (%) used to filter S2.
-- `s2_cloudmask` — Whether cloud-masked composite was used (bool).
-- `s2_channels8` — String listing first 8 S2 channels used in cube.
-- `return_groups` — Provenance grouping tag for tiles.
-- `enable_landfire` — Whether LANDFIRE layers were enabled (bool).
-- `overture_release` — Overture release tag for vector data used.
-- `udf_name` — Name of the UDF recipe used to build the cube.
-- `labels_hist_json` — JSON histogram of semantic label classes for this tile.
-- `cube_npz` — Binary NPZ of the full cube (excluded from export if set False).
+- **RGB composites**: `naip_rgb`, `s2_rgb`, `s2_pseudo_rgb`, `dem_rgb`
+- **Mono layers**: `naip_ndvi` and **leftover S2 bands** as single-channel `s2_*`
+- **Semantic masks**: `labels` (task labels), `landfire_family`, `cdl`
+- **Metadata**: `tile_id`, `city`, `bbox` (`west,south,east,north`), `chip_px`, `split`, `meta_json`
 
-### Semantic `labels` classes (sample)
-- `1`
-- `2`
-- `3`
-- `4`
-- `5`
-- `6`
-- `7`
-- `11`
-
-## How to load
-
-Whole dataset:
+> PNGs are **lossless** encodings of uint8 arrays; the big `cube_npz` payload is omitted to keep the repo small.
 
+## Loading
 ```python
 from datasets import load_dataset
 ds = load_dataset("gdurkin/naip-16d-city-cubes", split="train")
-# Tip: project columns to avoid pulling big arrays when you don't need them
-ds_small = ds.select_columns(["tile_id","town","chip_px","labels_hist_json"])
+ds[0].keys()
 ```
 
-One town only:
-
+To load a single city:
 ```python
 from datasets import load_dataset
-pattern = "hf://datasets/gdurkin/naip-16d-city-cubes/data/city=Calabasas__California__USA/*.parquet"
+city = "Arcadia__California__USA"
+pattern = f"hf://datasets/gdurkin/naip-16d-city-cubes/data/city={city}/*.parquet"
 ds_city = load_dataset("parquet", data_files=pattern, split="train")
 ```
 
-Arrow direct (fast column projection):
-
-```python
-import pyarrow.dataset as ds
-dataset = ds.dataset("hf://datasets/gdurkin/naip-16d-city-cubes/data", format="parquet")
-tbl = dataset.scanner(columns=["tile_id","town","chip_px"]).head(5)
-```
-
-## Training tips (4-headed setup)
-
-- **Inputs**: one or more of `naip_rgb`, `s2_rgb`, `s2_pseudo_rgb`, `dem_rgb`; add scalar/context columns as needed.
-- **Segmentation head**: `labels` (H×W, uint8). Build palette from `labels_hist_json`.
-- **Land-cover head**: `landfire_family` (H×W) or tile-level histogram.
-- **Aux head**: `cdl` histogram or any scalar derived per tile.
-- **Risk head**: your external risk target per tile.
-
-Minimal PyTorch collate (projecting only what you need):
-
-```python
-from datasets import load_dataset
-import numpy as np, torch
-
-ds = load_dataset("gdurkin/naip-16d-city-cubes", split="train")
-
-def to_chw(u8_hw3):
-    a = np.array(u8_hw3, dtype=np.uint8)
-    return torch.from_numpy(a).permute(2,0,1)
-
-def collate(batch):
-    x_rgb = [to_chw(ex["naip_rgb"]) for ex in batch]
-    y_seg = [torch.tensor(ex["labels"], dtype=torch.long) for ex in batch]
-    return {"x_rgb": torch.stack(x_rgb), "y_seg": torch.stack(y_seg)}
-
-loader = torch.utils.data.DataLoader(ds.with_format("python"), batch_size=4, collate_fn=collate)
-```
-
-## Performance
-
-- Use **column projection** (see examples) — e.g., skip `naip_rgb` if you only need metadata.
-- Prefer **per-city paths** for fast subsetting and incremental experiments.
-- Arrays are `uint8` (`labels`/`landfire_family`/`cdl` categorical are uint8). Decode to tensors lazily in your collate.
-
-## Reproducibility
-
-Export configuration (used when materializing view columns from `cube_npz`):
-
-```yaml
-make_columns:
-  - naip_rgb
-  - naip_ndvi
-  - s2_rgb
-  - s2_pseudo_rgb   # recipe: MSAVI,B08,NDMI
-  - dem_rgb
-  - labels
-  - landfire_family
-  - cdl
-train_ratio: 0.90
-split_seed: 13
-include_cube_npz: false
-max_shard_size: 500MB
-```
-
-## Largest groups (top 20)
-
-- **Santa_Cruz__California__USA**: 66 shards
-- **Sonoma__California__USA**: 65 shards
-- **Ventura__California__USA**: 47 shards
-- **Flagstaff__Arizona__USA**: 39 shards
-- **Santa_Rosa__California__USA**: 38 shards
-- **Thousand_Oaks__California__USA**: 28 shards
-- **Pasadena__California__USA**: 28 shards
-- **Napa__California__USA**: 27 shards
-- **Mariposa__California__USA**: 23 shards
-- **Eugene__Oregon__USA**: 18 shards
-- **Paradise__California__USA**: 16 shards
-- **Simi_Valley__California__USA**: 16 shards
-- **Lake_Elsinore__California__USA**: 16 shards
-- **Hemet__California__USA**: 16 shards
-- **Malibu__California__USA**: 14 shards
-- **Payson__Arizona__USA**: 13 shards
-- **Arcadia__California__USA**: 12 shards
-- **Laguna_Niguel__California__USA**: 11 shards
-- **Redding__California__USA**: 10 shards
-- **Auberry__California__USA**: 9 shards
-
-## Sources & licensing
-
-- Packaging, metadata, and annotations: **CC-BY-4.0**.
-- Underlying sources follow their native licenses:
-  - USDA **NAIP** and USGS **DEM** are public-domain.
-  - **Copernicus Sentinel-2** is free & open with attribution.
-  - **LANDFIRE/CDL** are US-Gov data. Cite sources in publications.
-
-## Caveats
-
-- `cube_npz` is excluded to keep the repo light; all published arrays were materialized during export.
-- If you hit an old `datasets` version that struggles with YAML headers, load with the Parquet builder as shown above.
+## Columns & shapes
+- `naip_rgb` _(512×512×3 u8)_: NAIP true color
+- `s2_rgb` _(512×512×3 u8)_: Sentinel-2 true color (B04,B03,B02)
+- `s2_pseudo_rgb` _(512×512×3 u8)_: pseudoRGB (default: MSAVI,B08,NDMI)
+- `dem_rgb` _(512×512×3 u8)_: elevation/gradient/aspect visualization
+- `naip_ndvi` _(512×512 u8)_: NDVI from NAIP
+- `s2_*` _(512×512 u8)_: leftover S2 bands (e.g., B08, MSAVI, NDVI, NDWI, NDMI, SCL)
+- `labels`, `landfire_family`, `cdl` _(512×512 u8)_: categorical masks
+
+## Legends
+
+### Labels (task)
+| id | name | color |
+|--:|:-----|:------|
+| 0 | `background` | `#000000` |
+| 1 | `road_concrete` | `#C8C8C8` |
+| 2 | `pavement` | `#A0A0A0` |
+| 3 | `dirt_gravel` | `#96643C` |
+| 4 | `grass_dry` | `#E6DC78` |
+| 5 | `grass_healthy` | `#50C878` |
+| 6 | `vegetation` | `#147814` |
+| 7 | `building` | `#FF5050` |
+| 8 | `water` | `#4682B4` |
+| 11 | `building_res` | `#FF8C8C` |
+| 12 | `building_com` | `#FF5A1E` |
+
+### LANDFIRE family
+| id | family | color |
+|--:|:-------|:------|
+| 0 | `background` | `#000000` |
+| 1 | `grass` | `#F7E68C` |
+| 2 | `shrub` | `#C69842` |
+| 3 | `timber` | `#1E781E` |
+| 4 | `slash` | `#CE5A32` |
+| 9 | `urban` | `#969696` |
+| 10 | `snow_ice` | `#C8E6FF` |
+| 11 | `agriculture` | `#FFCC66` |
+| 12 | `water` | `#4682B4` |
+| 13 | `barren` | `#C2B280` |
+
+### CDL (full palette)
+The table below lists CDL codes and colors (urban classes are overridden for better visual separation).
+
+<details><summary><strong>Click to expand full CDL legend</strong></summary>
+
+| id | class | color |
+|--:|:------|:------|
+| 0 | `Background` | `000000` |
+| 1 | `Corn` | `ffd400` |
+| 2 | `Cotton` | `ff2626` |
+| 3 | `Rice` | `00a9e6` |
+| 4 | `Sorghum` | `ff9e0f` |
+| 5 | `Soybeans` | `267300` |
+| 6 | `Sunflower` | `ffff00` |
+| 10 | `Peanuts` | `70a800` |
+| 11 | `Tobacco` | `00af4d` |
+| 12 | `Sweet Corn` | `e0a60f` |
+| 13 | `Pop or Orn Corn` | `e0a60f` |
+| 14 | `Mint` | `80d4ff` |
+| 21 | `Barley` | `e2007f` |
+| 22 | `Durum Wheat` | `8a6453` |
+| 23 | `Spring Wheat` | `d9b56c` |
+| 24 | `Winter Wheat` | `a87000` |
+| 25 | `Other Small Grains` | `d69dbc` |
+| 26 | `Dbl Crop WinWht/Soybeans` | `737300` |
+| 27 | `Rye` | `ae017e` |
+| 28 | `Oats` | `a15889` |
+| 29 | `Millet` | `73004c` |
+| 30 | `Speltz` | `d69dbc` |
+| 31 | `Canola` | `d1ff00` |
+| 32 | `Flaxseed` | `8099ff` |
+| 33 | `Safflower` | `d6d600` |
+| 34 | `Rape Seed` | `d1ff00` |
+| 35 | `Mustard` | `00af4d` |
+| 36 | `Alfalfa` | `ffa8e3` |
+| 37 | `Other Hay/Non Alfalfa` | `a5f58d` |
+| 38 | `Camelina` | `00af4d` |
+| 39 | `Buckwheat` | `d69dbc` |
+| 41 | `Sugarbeets` | `a900e6` |
+| 42 | `Dry Beans` | `a80000` |
+| 43 | `Potatoes` | `732600` |
+| 44 | `Other Crops` | `00af4d` |
+| 45 | `Sugarcane` | `b380ff` |
+| 46 | `Sweet Potatoes` | `732600` |
+| 47 | `Misc Vegs & Fruits` | `ff6666` |
+| 48 | `Watermelons` | `ff6666` |
+| 49 | `Onions` | `ffcc66` |
+| 50 | `Cucumbers` | `ff6666` |
+| 51 | `Chick Peas` | `00af4d` |
+| 52 | `Lentils` | `00deb0` |
+| 53 | `Peas` | `55ff00` |
+| 54 | `Tomatoes` | `f5a27a` |
+| 55 | `Caneberries` | `ff6666` |
+| 56 | `Hops` | `00af4d` |
+| 57 | `Herbs` | `80d4ff` |
+| 58 | `Clover/Wildflowers` | `e8beff` |
+| 59 | `Sod/Grass Seed` | `b2ffde` |
+| 60 | `Switchgrass` | `00af4d` |
+| 61 | `Fallow/Idle Cropland` | `bfbf7a` |
+| 63 | `Forest` | `95ce93` |
+| 64 | `Shrubland` | `c7d79e` |
+| 65 | `Barren` | `ccbfa3` |
+| 66 | `Cherries` | `ff00ff` |
+| 67 | `Peaches` | `ff91ab` |
+| 68 | `Apples` | `b90050` |
+| 69 | `Grapes` | `704489` |
+| 70 | `Christmas Trees` | `007878` |
+| 71 | `Other Tree Crops` | `b39c70` |
+| 72 | `Citrus` | `ffff80` |
+| 74 | `Pecans` | `b6705c` |
+| 75 | `Almonds` | `00a884` |
+| 76 | `Walnuts` | `ebd6b0` |
+| 77 | `Pears` | `b39c70` |
+| 81 | `Clouds/No Data` | `f7f7f7` |
+| 82 | `Developed` | `9c9c9c` |
+| 83 | `Water` | `4d70a3` |
+| 87 | `Wetlands` | `80b3b3` |
+| 88 | `Nonag/Undefined` | `e9ffbe` |
+| 92 | `Aquaculture` | `00ffff` |
+| 111 | `Open Water` | `4d70a3` |
+| 112 | `Perennial Ice/Snow` | `d4e3fc` |
+| 121 | `Developed/Open Space` | `#BDBDBD` |
+| 122 | `Developed/Low Intensity` | `#F4B183` |
+| 123 | `Developed/Med Intensity` | `#E06666` |
+| 124 | `Developed/High Intensity` | `#A61C00` |
+| 131 | `Barren` | `ccbfa3` |
+| 141 | `Deciduous Forest` | `95ce93` |
+| 142 | `Evergreen Forest` | `95ce93` |
+| 143 | `Mixed Forest` | `95ce93` |
+| 152 | `Shrubland` | `c7d79e` |
+| 176 | `Grass/Pasture` | `e9ffbe` |
+| 190 | `Woody Wetlands` | `80b3b3` |
+| 195 | `Herbaceous Wetlands` | `80b3b3` |
+| 204 | `Pistachios` | `00ff8c` |
+| 205 | `Triticale` | `d69dbc` |
+| 206 | `Carrots` | `ff6666` |
+| 207 | `Asparagus` | `ff6666` |
+| 208 | `Garlic` | `ff6666` |
+| 209 | `Cantaloupes` | `ff6666` |
+| 210 | `Prunes` | `ff91ab` |
+| 211 | `Olives` | `344a34` |
+| 212 | `Oranges` | `e67525` |
+| 213 | `Honeydew Melons` | `ff6666` |
+| 214 | `Broccoli` | `ff6666` |
+| 215 | `Avocados` | `66994d` |
+| 216 | `Peppers` | `ff6666` |
+| 217 | `Pomegranates` | `b39c70` |
+| 218 | `Nectarines` | `ff91ab` |
+| 219 | `Greens` | `ff6666` |
+| 220 | `Plums` | `ff91ab` |
+| 221 | `Strawberries` | `ff6666` |
+| 222 | `Squash` | `ff6666` |
+| 223 | `Apricots` | `ff91ab` |
+| 224 | `Vetch` | `00af4d` |
+| 225 | `Dbl Crop WinWht/Corn` | `ffd400` |
+| 226 | `Dbl Crop Oats/Corn` | `ffd400` |
+| 227 | `Lettuce` | `ff6666` |
+| 228 | `Dbl Crop Triticale/Corn` | `ffd400` |
+| 229 | `Pumpkins` | `ff6666` |
+| 230 | `Dbl Crop Lettuce/Durum Wht` | `8a6453` |
+| 231 | `Dbl Crop Lettuce/Cantaloupe` | `ff6666` |
+| 232 | `Dbl Crop Lettuce/Cotton` | `ff2626` |
+| 233 | `Dbl Crop Lettuce/Barley` | `e2007f` |
+| 234 | `Dbl Crop Durum Wht/Sorghum` | `ff9e0f` |
+| 235 | `Dbl Crop Barley/Sorghum` | `ff9e0f` |
+| 236 | `Dbl Crop WinWht/Sorghum` | `a87000` |
+| 237 | `Dbl Crop Barley/Corn` | `ffd400` |
+| 238 | `Dbl Crop WinWht/Cotton` | `a87000` |
+| 239 | `Dbl Crop Soybeans/Cotton` | `267300` |
+| 240 | `Dbl Crop Soybeans/Oats` | `267300` |
+| 241 | `Dbl Crop Corn/Soybeans` | `ffd400` |
+| 242 | `Blueberries` | `000099` |
+| 243 | `Cabbage` | `ff6666` |
+| 244 | `Cauliflower` | `ff6666` |
+| 245 | `Celery` | `ff6666` |
+| 246 | `Radishes` | `ff6666` |
+| 247 | `Turnips` | `ff6666` |
+| 248 | `Eggplants` | `ff6666` |
+| 249 | `Gourds` | `ff6666` |
+| 250 | `Cranberries` | `ff6666` |
+| 254 | `Dbl Crop Barley/Soybeans` | `267300` |
+</details>
+
+
+## Coverage
+<!-- COVERAGE-TABLE:BEGIN -->
+
+**Coverage**
+- rows: **24,792**
+- S2 mono policy: `leftover`
+
+| column | rows | pct |
+|---|---:|---:|
+| `cdl` | 24,792 | 100.0% |
+| `dem_rgb` | 24,792 | 100.0% |
+| `labels` | 24,792 | 100.0% |
+| `landfire_family` | 24,792 | 100.0% |
+| `naip_ndvi` | 24,792 | 100.0% |
+| `naip_rgb` | 24,792 | 100.0% |
+| `s2_B02` | 0 |   0.0% |
+| `s2_B03` | 0 |   0.0% |
+| `s2_B04` | 0 |   0.0% |
+| `s2_B08` | 0 |   0.0% |
+| `s2_MSAVI` | 0 |   0.0% |
+| `s2_NDVI` | 24,792 | 100.0% |
+| `s2_NDWI` | 24,792 | 100.0% |
+| `s2_SCL` | 24,792 | 100.0% |
+| `s2_pseudo_rgb` | 0 |   0.0% |
+| `s2_rgb` | 24,792 | 100.0% |
+
+<!-- COVERAGE-TABLE:END -->