jq5522/aml_indo_fires

Indonesian Wildfire Panel Dataset (2012–2025)

A spatiotemporal panel dataset covering fire activity across Sumatra and Kalimantan, Indonesia, at monthly resolution on a 0.25° grid, spanning January 2012 to December 2025. Built for machine learning research into tropical wildfire prediction and its relationship to land cover, peatlands, and climate.

---

Dataset Summary

Each observation corresponds to one grid cell × one calendar month. Fire detections from NASA FIRMS VIIRS are aggregated to each cell-month, then joined with ERA5 meteorological reanalysis and static land-cover features. Two models are trained on this data — an Elastic Net logistic regression baseline and a LightGBM classifier — and their predictions are stored in outputs/tables/.

Five panel files are provided at increasing levels of processing:

File	Stage	Purpose
panel_core_2012_2025.parquet	Raw	Aggregated fire, climate, and land-cover features
panel_eda_enriched_2012_2025.parquet	Enriched	Adds lags, drought proxies, interactions, spatial neighbours
static_cell_features_2012_2025.parquet	Static	One row per cell: land-cover attributes
modelling_data_linear_2012_2025.parquet	Model-ready	Feature set for Elastic Net (with 14 interaction terms; province/month dummies, drop_first=True)
modelling_data_nonlinear_2012_2025.parquet	Model-ready	Feature set for LightGBM (province/month as raw categoricals for native handling)

---

Geographic Scope

Region	Bounding Box (W, S, E, N)
Sumatra	94.0°E, 6.5°S, 106.5°E, 6.5°N
Kalimantan	108.0°E, 5.0°S, 119.5°E, 8.5°N
Combined study area	94.0°E, 6.5°S, 119.5°E, 8.5°N

• Grid resolution: 0.25° × 0.25° (aligned to ERA5 grid)
• Land cells retained: ~1,839 (ocean cells filtered using province assignment and land-cover overlap)
• Temporal coverage: January 2012 – December 2025 (168 months)

---

Files

Panel data

File	Description
panel_core_2012_2025.parquet	Core cell × month panel: fire counts, ERA5 climate, static land-cover
panel_eda_enriched_2012_2025.parquet	Enriched panel with lagged variables, drought proxies, interactions, and spatial neighbour features
static_cell_features_2012_2025.parquet	Static per-cell features (peat, concessions, roads, province)
modelling_data_linear_2012_2025.parquet	Model-ready data for Elastic Net: curated features + 14 interaction terms + province/month dummies (drop_first=True)
modelling_data_nonlinear_2012_2025.parquet	Model-ready data for LightGBM: province and month as raw categoricals (no dummy encoding)

Visualisation layers

File	Description
visualisation/gadm1_clip.parquet	GADM level-1 administrative boundaries (GeoParquet)
visualisation/indonesia_border.parquet	Dissolved study-area outline (GeoParquet)
visualisation/grid_025deg.parquet	0.25° grid cell polygons (GeoParquet)
visualisation/fire_grid_totals_2012_2025.parquet	Grid with total and log1p-transformed fire counts (GeoParquet)
visualisation/peat_clip.parquet	Peatland polygons clipped to study area (GeoParquet)
visualisation/roads_clip.parquet	Road network clipped to study area (GeoParquet)
visualisation/concessions_oil_palm_clip.parquet	Oil palm concession polygons (GeoParquet)
visualisation/concessions_wood_fibre_clip.parquet	Wood fibre concession polygons (GeoParquet)

Model outputs

File	Description
outputs/tables/enet_final_test.csv	Elastic Net test-set metrics (log-loss, PR-AUC, ROC-AUC, Brier) for 2024 and 2025
outputs/tables/enet_test_predictions.parquet	Elastic Net predicted probabilities for the final test year, with cell_id and actuals
outputs/tables/lgbm_final_test.csv	LightGBM test-set metrics (log-loss, PR-AUC, ROC-AUC, Brier) for 2024 and 2025
outputs/tables/lgbm_test_predictions.parquet	LightGBM predicted probabilities for the final test year, with cell_id and actuals

---

Schema — panel_core_2012_2025.parquet

Identifiers & Time

Column	Type	Description
cell_id	int	Unique grid cell identifier
year	int	Calendar year
month	int	Calendar month (1–12)
time	datetime	First day of the month (timestamp)
lon_centre	float	Longitude of cell centre (°E)
lat_centre	float	Latitude of cell centre (°N)

Fire Features (NASA FIRMS VIIRS SNPP)

Column	Type	Description
fire_count_lnh	int	Total fire detections (low, nominal, high confidence)
fire_count_nh	int	Fire detections at nominal or high confidence only
fire_any_lnh	int	Binary: any fire detected (all confidence levels)
fire_any_nh	int	Binary: any fire detected (nominal/high confidence only)
frp_sum_lnh	float	Sum of Fire Radiative Power — all detections (MW)
frp_sum_nh	float	Sum of Fire Radiative Power — nominal/high confidence only (MW)

ERA5 Climate Features (monthly reanalysis, nearest grid point)

Column	Type	Description
t2m	float	2 m air temperature (K)
tp	float	Total precipitation (m)
u10	float	10 m U-component of wind (m/s, eastward)
v10	float	10 m V-component of wind (m/s, northward)
wind_speed	float	10 m wind speed: sqrt(u10² + v10²) (m/s)
era5_latitude	float	Matched ERA5 grid latitude
era5_longitude	float	Matched ERA5 grid longitude

Static Land-Cover Features

Column	Type	Description
cell_area_km2	float	Cell area in km² (EPSG:6933 equal-area projection)
peat_fraction	float	Fraction of cell covered by peatland [0, 1]
oil_palm_fraction	float	Fraction of cell within oil palm concessions [0, 1]
wood_fibre_fraction	float	Fraction of cell within wood fibre concessions [0, 1]
road_length_km	float	Total road length within cell (km)
road_density_km_per_km2	float	Road density (km per km²)
province	str	Indonesian province name (GADM level 1); NaN for ocean/coastal cells

---

Schema — panel_eda_enriched_2012_2025.parquet

Contains all columns from panel_core_2012_2025.parquet, plus the engineered features below. All lagged variables are computed within each cell_id group to avoid cross-cell leakage.

Derived Climate Variable

Column	Type	Description
t2m_c	float	2 m air temperature in Celsius (t2m − 273.15)

Binary & Derived Land-Cover Indicators

Column	Type	Description
peat_binary	int	1 if cell has any peatland coverage
oil_con_binary	int	1 if cell has any oil palm concession coverage
wood_con_binary	int	1 if cell has any wood fibre concession coverage
concession_fraction_total	float	Combined oil palm + wood fibre concession share [0, 1]
high_road_density	int	1 if cell road density exceeds the sample median

Lagged Fire Variables (lags: 1, 2, 3, 12 months)

Pattern: {var}_lag{n} for n in {1, 2, 3, 12}.

Base variable	Description
fire_count_nh	Nominal/high confidence fire count
fire_any_nh	Binary fire indicator (nominal/high)
frp_sum_nh	Total FRP (nominal/high confidence, MW)
fire_count_lnh	All-confidence fire count
fire_any_lnh	Binary fire indicator (all confidence)
frp_sum_lnh	Total FRP (all confidence, MW)

Lagged Climate Variables (lags: 1, 2, 3, 12 months)

Pattern: {var}_lag{n} for n in {1, 2, 3, 12}.

Base variable	Description
tp	Total precipitation
t2m	2 m temperature (K)
t2m_c	2 m temperature (°C)
wind_speed	Wind speed magnitude
u10	U-component of wind (retained for downstream haze transport analysis)
v10	V-component of wind (retained for downstream haze transport analysis)

Precipitation Anomaly & Drought Proxies

Column	Type	Description
tp_clim_eda	float	Cell-month climatological mean precipitation (m)
tp_anomaly	float	Precipitation anomaly: tp_clim_eda minus tp; positive = drier than normal
tp_deficit_monthly	float	Dry-only monthly deficit: tp_anomaly clipped to >= 0
precip_deficit_3m	float	3-month cumulative net anomaly (shifted 1 month to use past months only)
precip_deficit_3m_dryonly	float	3-month cumulative dry-only deficit (shifted 1 month)
tp_3m_sum_lag1	float	Trailing 3-month total precipitation, lagged 1 month (m)

Wind Direction

Column	Type	Description
wind_dir_rad	float	Wind direction in radians: arctan2(v10, u10)

Climate x Land-Cover Interactions (EDA use)

Column	Type	Description
tp_x_peat	float	tp x peat_fraction
t2m_c_x_peat	float	t2m_c x peat_fraction
wind_speed_x_peat	float	wind_speed x peat_fraction
tp_x_oil_con	float	tp x oil_palm_fraction
tp_x_wood_con	float	tp x wood_fibre_fraction
tp_x_road_density	float	tp x road_density_km_per_km2

Spatial Neighbour Variables (Queen contiguity)

Column	Type	Description
neighbor_fire_count_nh_mean	float	Mean fire count (nh) across neighbours — same month (EDA only)
neighbor_fire_count_nh_sum	float	Sum of fire counts (nh) across neighbours — same month (EDA only)
neighbor_fire_any_nh_mean	float	Mean binary fire indicator (nh) across neighbours — same month (EDA only)
neighbor_frp_sum_nh_mean	float	Mean FRP (nh) across neighbours — same month (EDA only)
neighbor_fire_count_nh_mean_lag1	float	Above, lagged 1 month
neighbor_fire_count_nh_sum_lag1	float	Above, lagged 1 month
neighbor_fire_any_nh_mean_lag1	float	Above, lagged 1 month
neighbor_frp_sum_nh_mean_lag1	float	Above, lagged 1 month

---

Schema — modelling_data_linear_2012_2025.parquet and modelling_data_nonlinear_2012_2025.parquet

Both files share the same base feature set but differ in interaction terms and categorical encoding.

Identifiers and Targets (both files)

Column	Type	Description
cell_id	int	Grid cell identifier
year	int	Calendar year (raw integer; not encoded)
month	int	Calendar month — raw integer in nonlinear file; dummy-encoded in linear file
fire_any_nh	int	Primary target: binary fire occurrence (nominal/high confidence)
fire_count_nh	int	Secondary target: fire count (for regression sensitivity check)

Base Features (both files)

Spatial controls: lon_centre, lat_centre, province (raw string in nonlinear; dummies in linear).

Static land-cover: peat_fraction, oil_palm_fraction, wood_fibre_fraction, road_length_km, road_density_km_per_km2.

Lagged fire history (n/h): fire_count_nh_lag{1/2/3/12}, fire_any_nh_lag{1/2/3/12}, frp_sum_nh_lag{1/2/3/12}.

Lagged climate: tp_lag{1/2/3/12}, t2m_c_lag{1/2/3/12}, wind_speed_lag{1/2/3/12}.

Drought proxies: precip_deficit_3m, precip_deficit_3m_dryonly, tp_3m_sum_lag1.

Spatial neighbour lags: neighbor_fire_count_nh_mean_lag1, neighbor_fire_count_nh_sum_lag1, neighbor_fire_any_nh_mean_lag1, neighbor_frp_sum_nh_mean_lag1.

Interaction Terms (linear file only — 14 terms)

Column	Formula	Motivation
ix_drought_x_peat	precip_deficit_3m x peat_fraction	Drought drives peatland fire disproportionately (Sherwood et al. 2021)
ix_drought_dry_x_peat	precip_deficit_3m_dryonly x peat_fraction	Dry-season-only deficit x peat
ix_frp_lag{1/2/3/12}_x_peat	frp_sum_nh_lag{n} x peat_fraction	Intense recent peat fire indicates persistent subsurface combustion
ix_oil_con_x_peat	oil_palm_fraction x peat_fraction	Oil palm on peat — primary THPA enforcement target
ix_wood_con_x_peat	wood_fibre_fraction x peat_fraction	Wood fibre on peat
ix_concession_x_peat	oil_palm_fraction x wood_fibre_fraction x peat_fraction	Combined concession x peat
ix_wind_lag{1/2/3/12}_x_drought	wind_speed_lag{n} x precip_deficit_3m	Wind amplifies fire risk conditionally on dryness (Abatzoglou & Kolden 2013)
ix_neighbor_lag1_x_peat	neighbor_fire_count_nh_mean_lag1 x peat_fraction	Spatial contagion more persistent on peat

Categorical Encoding

Variable	Linear file	Non-linear file
province	One-hot dummies, drop_first=True	Raw string encoded to integer as province_code; passed via categorical_feature
month	One-hot dummies, drop_first=True	Raw integer; passed via categorical_feature
year	Raw integer	Raw integer

---

Schema — outputs/tables/

Test-set evaluation results and predicted probabilities from the two trained models.

enet_final_test.csv and lgbm_final_test.csv

One row per test year (2024, 2025).

Column	Type	Description
test_year	int	Calendar year being evaluated
n_test	int	Number of cell-month observations in the test set
fire_rate	float	Fraction of test observations with fire
log_loss	float	Binary cross-entropy
pr_auc	float	Area under the precision-recall curve
roc_auc	float	Area under the ROC curve
brier	float	Brier score (mean squared probability error)

lgbm_final_test.csv additionally contains n_trees_used (number of boosting rounds selected by early stopping).

enet_test_predictions.parquet and lgbm_test_predictions.parquet

One row per cell-month observation in the final test year.

Column	Type	Description
cell_id	int	Grid cell identifier
year	int	Calendar year
month	int	Calendar month
fire_any_nh	int	Actual binary fire label
y_prob_enet / y_prob_lgbm	float	Predicted fire probability from the respective model

---

Modelling Design

Target Variable

The primary target is fire_any_nh (binary: >=1 nominal/high-confidence VIIRS detection in a cell-month). At threshold=1 the positive rate is ~33%. 513 cells burn in >50% of months (chronic hotspots). Only 236 cells never burn, meaning the negative class is genuinely informative. Consistent with Sherwood et al. (2021) and Kurniawan et al. (2025).

fire_count_nh is retained as a secondary regression target for Poisson sensitivity checks.

Train / Validation / Test Split

Period	Role
2012–2013	Burn-in (lag initialisation only; excluded from all folds)
2014–2021	Training (expanding window)
2022–2023	Outer validation (model class selection)
2024–2025	Final test (touched once, at reporting)

Inner CV (hyperparameter tuning) uses 4 expanding-window folds:

Fold	Train	Validation
1	2014–2017	2018
2	2014–2018	2019 (El Nino stress test)
3	2014–2019	2020
4	2014–2020	2021

Model Classes

Elastic Net (05A): Logistic regression with elastic net penalty, solved via FISTA on GPU (JAX) or sklearn SAGA as fallback. Hyperparameters C and l1_ratio tuned by inner CV on log-loss. Features include 14 pre-computed interaction terms.

LightGBM (05B): Gradient boosted decision trees with GPU acceleration. Hyperparameters tuned via Optuna or grid search. Province and month passed as native categorical features. Early stopping applied on a held-out fold separate from the scoring fold to prevent leakage. No scale_pos_weight — class imbalance is left unweighted to preserve probability calibration.

---

Data Sources

Source	Description	Coverage
NASA FIRMS VIIRS SNPP	Active fire detections	2012–2025
ERA5 Monthly Reanalysis (CDS)	Temperature, precipitation, wind	2012–2025
Global Peatland Database / GFW	Peatland extent polygons	Static
GFW — Oil Palm Concessions	Oil palm concession boundaries	Static
GFW — Wood Fibre Concessions	Wood fibre concession boundaries	Static
OpenStreetMap / GADM	Road network; administrative boundaries	Static

---

Loading the Data

import pandas as pd
from huggingface_hub import hf_hub_download

REPO_ID = "jq5522/aml_indo_fires"

# Core panel
panel = pd.read_parquet(hf_hub_download(REPO_ID, "panel_core_2012_2025.parquet", repo_type="dataset"))

# Enriched EDA panel
panel_eda = pd.read_parquet(hf_hub_download(REPO_ID, "panel_eda_enriched_2012_2025.parquet", repo_type="dataset"))

# Model-ready: Elastic Net
panel_linear = pd.read_parquet(hf_hub_download(REPO_ID, "modelling_data_linear_2012_2025.parquet", repo_type="dataset"))

# Model-ready: LightGBM
panel_nonlin = pd.read_parquet(hf_hub_download(REPO_ID, "modelling_data_nonlinear_2012_2025.parquet", repo_type="dataset"))
panel_nonlin["province"] = panel_nonlin["province"].astype("category")
panel_nonlin["month"] = panel_nonlin["month"].astype("category")

# Model outputs — test metrics
enet_test = pd.read_csv(hf_hub_download(REPO_ID, "outputs/tables/enet_final_test.csv", repo_type="dataset"))
lgbm_test = pd.read_csv(hf_hub_download(REPO_ID, "outputs/tables/lgbm_final_test.csv", repo_type="dataset"))

# Model outputs — predicted probabilities
enet_preds = pd.read_parquet(hf_hub_download(REPO_ID, "outputs/tables/enet_test_predictions.parquet", repo_type="dataset"))
lgbm_preds = pd.read_parquet(hf_hub_download(REPO_ID, "outputs/tables/lgbm_test_predictions.parquet", repo_type="dataset"))

# GeoParquet visualisation layer (requires geopandas)
import geopandas as gpd

grid = gpd.read_parquet(hf_hub_download(REPO_ID, "visualisation/grid_025deg.parquet", repo_type="dataset"))
grid.plot()

---

Construction Notes

• Ocean cell filtering: Cells are retained if they have a non-null province assignment or a non-zero value for any of peat_fraction, oil_palm_fraction, or wood_fibre_fraction. This preserves ~11 coastal cells with roads but no province assignment.
• Fire confidence: _nh columns restrict to nominal and high VIIRS confidence classes. _lnh columns include low-confidence detections. Only _nh series are retained in the model-ready files.
• ERA5 matching: Each grid cell centre is snapped to its nearest ERA5 gridpoint using argmin distance. ERA5 and the analysis grid share the same 0.25° resolution.
• Panel skeleton: A complete cell × month grid is constructed for all 168 months and all land cells. Months with no fire detections are filled with zeros, not dropped.
• Lag leakage prevention: All lagged and rolling variables are computed within cell_id groups and shifted before rolling.
• Wind direction components (u10/v10): Retained in panel_eda_enriched but excluded from both model-ready files. Reserved for a post-prediction transboundary haze transport layer.
• Interaction terms: Present in the linear file only. LightGBM recovers these through recursive partitioning.
• Categorical encoding: drop_first=True for province and month dummies in the linear file. The non-linear file retains raw values for LightGBM native categorical handling.
• Early stopping (LightGBM): A three-way chronological split is used per fold — fit set, early-stopping set, and scoring set — keeping the validation year strictly untouched during tree selection.
• Class weighting: Neither model uses scale_pos_weight or class weights, preserving probability calibration for log-loss and Brier score evaluation.

---

Citation

If you use this dataset, please cite the upstream data sources (NASA FIRMS, Copernicus ERA5, GFW) and link to this repository.

@dataset{aml_indo_fires_2025,
author    = {Jiaqi Chen},
title     = {Indonesian Wildfire Panel Dataset (2012--2025)},
year      = {2026},
publisher = {Hugging Face},
url       = {https://huggingface.co/datasets/jq5522/aml_indo_fires}
}

---

Licence

Released under Creative Commons Attribution 4.0 (CC BY 4.0). Downstream use of ERA5 data is subject to the Copernicus licence.