Dataset Description — Karhutla Riau Early-Warning System

Ingested feature and label data for the peatland-fire early-warning model, all aligned on a fixed 5 km equal-area grid over Riau (and the surrounding bbox). cell_idx is the universal join key across every source.

Repository layout

Local (gitignored, regenerable): data/output/<source>/ per source.

HuggingFace mirror (danishritonga/karhutla, the canonical copy):

raw/<source>/       all per-source CSVs (389 files across 8 sources)
tensors/            pre-built tensors: data.npy [1826,82,85,23] float32,
                    labels.npy [1826,82,85] int8 (-1/0/1), meta.json

• raw/chirpsv3/ — 60 monthly CSVs (CHIRPS v3 SAT daily rainfall)
• raw/dynamic_world/ — 60 monthly CSVs (Dynamic World class probabilities)
• raw/era5land/ — 60 monthly CSVs (ERA5-Land daily; corrected sum-precip version)
• raw/grid/ — grid definition (cells + meta)
• raw/peat/ — static peat depth/extent
• raw/sentinel1/ — raw S1 SAR acquisitions per orbit
• raw/sentinel1_filled/ — forward-filled daily S1 (with filled provenance)
• raw/viirs/ — VIIRS hotspot labels per year

The tensors/ folder is the model input: one data.npy covering 2019-01-01 → 2023-12-31 with 0 NaN (all missing data handled via spatial fill / mask channels — see the per-source notes below).

Grid

Defined once by data/grid/grid_definition.py and frozen before any data pull (sources are resampled onto it, never the reverse).

File	Contents
grid/grid_cells.csv	One row per cell: row, col, cell_idx, x_center_m, y_center_m, lon, lat, is_riau
grid/grid_meta.json	CRS (Albers Indonesia Equal Area Conic, proj4), x0, y0, cols, rows, cell_size_m, bbox_cells, riau_cells
grid/riau_boundary_aea.gpkg	Riau admin boundary, projected
grid/riau_grid.png	Map of the grid

Key numbers:

• 6,970 cells total (85 cols × 82 rows over the bbox), cell_idx 0..6969
• 3,598 cells inside Riau (is_riau = 1) — these are the primary prediction targets
• 6,506 land cells = union of CHIRPS coverage (3,598 Riau + 2,908 non-Riau land); the remaining 464 bbox cells are sea (CHIRPS land-only grid skips water)
• 5,000 m cell size; CRS +proj=aea +lat_1=-5 +lat_2=-1 +lat_0=2 +lon_0=113 ...

is_riau decides which cells are targets (get labels and are evaluated). All 6,970 bbox cells keep features, because neighbouring (even sea) cells carry context for 15×15 patches in the spatiotemporal model.

Sources

All sources are downloaded / rasterized by scripts in data/ingest/.

1. VIIRS hotspots — labels (data/ingest/viirs.py)

NASA FIRMS VIIRS S-NPP 375 m active-fire detections (confidence ∈ {nominal, high}), spatially binned onto the grid with exact Albers floor-division binning (col = floor((x-x0)/cell_size), row = floor((y-y0)/cell_size); out-of-bbox points are dropped, not clipped). The binary label for a cell on day t is 1 iff the cell has ≥ k = 2 detections within the window (t+1, t+7] (persistence filter against smouldering-peat false positives).

• Input: data/raw/viirs/viirs-snpp_{year}_all_countries.zip (kept, ~1.7 GB total)
• Raw extracts (untouched FIRMS Indonesia CSVs): real_data/viirs-snpp/{year}/viirs-snpp_{year}_Indonesia.csv
• Output: viirs/labels_{year}.csv → cell_idx, row, col, date, fire_label

Year	Rows	Days	Cells	Positive labels	Pos. rate
2019	2,374,690	365	6,506	21,652	0.91%
2020	2,381,196	366	6,506	4,411	0.19%
2021	2,374,690	365	6,506	3,345	0.14%
2022	2,374,690	365	6,506	3,192	0.13%
2023	2,374,690	365	6,506	4,941	0.21%

• Train (2019–2022): 32,600 positive · Test (2023): 4,941 positive
• Peak seasons match known Riau fire years (e.g. Sep 2019: 254–277 burning cells/day).

2. CHIRPS v3 SAT rainfall (data/ingest/chirpsv3.py)

UCSB-CHC/CHIRPS/V3/DAILY_SAT (0.05°, daily, mm/day, IMERG-Late-derived daily disaggregation). Mean over each 5 km cell.

• Output: chirpsv3/chirps_v3sat_{YYYYMM}.csv → cell_idx, row, col, date, precip_mm
• 60 files, 11.88M rows, full daily coverage 2019-01-01 → 2023-12-31
• 6,506 cells (all land; sea cells legitimately absent), 0 NaN, 0 negatives
• Values 0–385 mm/day; no gap-fill needed (complete daily product)

3. Sentinel-1 SAR backscatter (data/ingest/sentinel1.py)

COPERNICUS/S1_GRD IW 10 m, VV/VH dB + incidence angle (server-side calibrated, terrain-corrected; median over cell as speckle suppression). Separate files per orbit (ASCENDING / DESCENDING) and month. Revisit 6–12 days, swath-limited — not daily.

• Raw: sentinel1/s1_{ORBIT}_{YYYYMM}.csv → cell_idx, row, col, date, vv_db, vh_db, angle_deg, vh_vv_db
• ASC: 41 files, 1.17M rows · DESC: 41 files, 1.01M rows (0 NaN — only real obs)

Gap-filling (--fill, default on): forward-fill real acquisitions to daily rows, capped at max_gap = 14 days. A cell whose last real obs is older than 14 days has its value invalidated to NaN (a stale radar value would be misleading). A filled column records provenance:

• 0 = real acquisition

• 1 + valid value = forward-filled within the 14-day cap

• 1 + NaN = gap exceeded the cap, value deliberately erased

• Filled: sentinel1_filled/s1_{ORBIT}_{YYYYMM}.csv → date, row, col, vv_db, vh_db, angle_deg, vh_vv_db, cell_idx, filled

• 120 files; ASC: 12.69M rows (97.8% valid, ~10% real) · DESC: 12.72M rows (74.7% valid, ~7% real)

Why NaNs are correct: they are validity sentinels, not missing-data accidents. Tensor assembly converts them into a per-channel validity mask so the model learns where radar coverage is weak. The NaN rate itself is informative — DESC is 25.3% invalid vs ASC 2.2%, i.e. a weaker signal channel the model can down-weight. Forward-fill is causal (no future leakage, unlike backfill/interpolation) and mirrors live BPBD dashboard operation.

4. Peat depth / extent — static (data/ingest/peat.py)

BIG / Satupeta FEG layer 48 "Peta Fungsi Ekosistem Gambut 1:50.000" (PP 57/2016). Peat-thickness classes (e.g. '5,0 - 6,0 meter') parsed to midpoint metres; rasterized onto the grid with area-weighted overlay.

• Output: peat/peat_cell.csv → cell_idx, row, col, is_riau, peat_frac, peat_depth_m
• 6,970 rows (one per grid cell); 722 cells with peat (peat_frac > 0)
• Depth 0–13.32 m (area-weighted class midpoint over the peat-covered part of the cell)
• Static channel — broadcast unchanged across time in tensor assembly

5. ERA5-Land daily (data/ingest/era5land.py)

ECMWF/ERA5_LAND/HOURLY, reduced to daily: state bands (temperature_2m, dewpoint_temperature_2m, u/v_component_of_wind_10m, volumetric_soil_water_layer_1/2) are day-meaned; flux bands (total_precipitation, surface_solar_radiation_downwards) are day-summed (corrected version — an earlier teammate upload used mean for precipitation, which understates daily totals ~24×).

• Output: era5land/era5land_{YYYYMM}.csv → cell_idx, row, col, date, temperature_2m, dewpoint_temperature_2m, u_component_of_wind_10m, v_component_of_wind_10m, volumetric_soil_water_layer_1, volumetric_soil_water_layer_2, total_precipitation, surface_solar_radiation_downwards
• 60 files, 10.83M rows, full coverage; t2m 289.4–304.1 K; swvl1 0.041–0.607
• 1,149 static edge cells fall outside the ERA5 0.1° footprint at scale 9000 — tensor assembly spatially fills them from the nearest valid cell (median distance 10 km ≈ ERA5 native resolution)

6. Dynamic World land cover (data/ingest/dynamic_world.py)

GOOGLE/DYNAMICWORLD/V1 (10 m, Sentinel-2-derived, cloud-masked ≤35%). Mean class probability per cell; sparse revisit (cloud cover) — some months have near-zero Riau coverage. Missing (cell, date) rows get class probs = 0 and a dw_available = 0 mask channel in the tensor (0+mask convention, so the model does not misread absent clouds as "bare/water").

• Output: dynamic_world/dynamic_world_{YYYYMM}.csv → cell_idx, row, col, date, water, trees, grass, flooded_vegetation, crops, shrub_and_scrub, built, bare, snow_and_ice, top1_class, top1_prob
• 60 files; 430 unique dates; all 6,970 cells covered at least 11× (median 50)

Summary

Source	Role	Cadence	Coverage	Rows (total)	Missing
Grid	spatial frame	static	6,970 cells	6,970	—
VIIRS	labels	daily	6,506 land cells	11.89M	none
CHIRPS	rainfall feature	daily	6,506 land cells	11.88M	sea cells only
S1 raw	radar feature	6–12 d	6,970 cells	2.18M	swath gaps (filled)
S1 filled	radar feature (daily)	daily	6,970 cells	25.40M	>14 d gaps → NaN + mask
ERA5-Land	weather + soil moisture	daily	5,821 cells native	10.83M	edge cells → spatial fill
Dynamic World	land cover probs	2–5 d	6,970 cells	0.39M	cloud → 0 + mask
Peat	static feature	static	6,970 cells	6,970	none

Pre-built tensors (tensors/)

Built by data/loader/tensor_assembly.py (all sources → daily [D,H,W,23] float32):

• Channel layout: 0-7 ERA5, 8 chirps_precip, 9-11 sar_vv/sar_vh/sar_available, 12-19 DW class probs, 20 dw_available, 21 peat_depth, 22 hotspot_count_lag
• T_IN = 14 days input, HORIZON = 7-day forecast window, 15×15 patch
• labels.npy: 1 = ≥k=2 hotspots in (t+1, t+7], 0 = no fire, -1 = not a land cell
• 0 NaN (ERA5 spatially filled; S1/DW use mask channels)

Reproduction commands (from repo root)

uv run python data/ingest/viirs.py --years 2019 2023 --keep-raw --raw-csv-dir real_data/viirs-snpp
uv run python data/ingest/chirpsv3.py --start 2019-01-01 --end 2023-12-31 --project <ee-project>
uv run python data/ingest/sentinel1.py --start 2019-01-01 --end 2023-12-31 --project <ee-project>
uv run python data/ingest/era5land.py --start 2019-01-01 --end 2023-12-31 --project <ee-project>
uv run python data/ingest/dynamic_world.py --start 2019-01-01 --end 2023-12-31 --project <ee-project>
uv run python data/ingest/peat.py
uv run python data/loader/tensor_assembly.py --start 2019-01-01 --end 2023-12-31

Large regenerable outputs (data/output/*) are gitignored; canonical copies live in the HuggingFace dataset repo under raw/ and tensors/.