Upload folder using huggingface_hub

README.md

@@ -0,0 +1,124 @@
+---
+license: cc-by-4.0
+task_categories:
+  - tabular-classification
+  - tabular-regression
+language:
+  - en
+tags:
+  - climate-health
+  - wildfire
+  - smoke
+  - pm2-5
+  - respiratory
+  - synthetic
+  - sub-saharan-africa
+pretty_name: Wildfire Smoke & Respiratory Outcomes (SSA)
+size_categories:
+  - 10K<n<100K
+configs:
+  - config_name: savanna_fire_belt
+    data_files: data/wildfire_savanna_fire_belt.csv
+  - config_name: forest_clearing_burn
+    data_files: data/wildfire_forest_clearing_burn.csv
+    default: true
+  - config_name: urban_peri_urban_haze
+    data_files: data/wildfire_urban_peri_urban_haze.csv
+---
+
+# Wildfire Smoke & Respiratory Outcomes in Sub-Saharan Africa
+
+## Abstract
+
+A synthetic dataset modelling wildfire smoke exposure and respiratory health outcomes across three fire regime scenarios in sub-Saharan Africa. Each record represents an individual-month observation describing smoke PM2.5 exposure, fire proximity, vulnerability, respiratory symptoms, and health service utilisation. Africa accounts for over 70% of global burned area annually, with savanna fires in West/Central Africa and agricultural clearing burns in the Congo Basin driving massive seasonal smoke plumes.
+
+### Scenarios
+
+- **Savanna Fire Belt**: West/Central African savanna fires (Nov-Mar) with widespread seasonal haze.
+- **Forest Clearing Burn**: Congo Basin agricultural/forest clearing fires (Jun-Sep) with intense localised smoke.
+- **Urban Peri-Urban Haze**: Cities downwind of fire zones with combined ambient + smoke pollution.
+
+## Dataset Structure
+
+Each scenario contains 10,000 records (30,000 total). Key columns:
+
+- `year`, `month`, `in_fire_season`, `age`, `sex`, `setting`
+- `is_child_u5`, `is_elderly_60plus`, `pre_existing_asthma`, `pre_existing_copd`, `smoker`
+- `pm25_smoke_ugm3`, `pm25_total_ugm3`, `aqi_category`
+- `fire_proximity_km`, `smoke_days_per_season`, `fire_count`
+- `mask_use`, `stayed_indoors`, `vulnerability_index`, `exposure_risk_score`
+- `cough`, `wheeze`, `dyspnoea`, `eye_irritation`
+- `ari_episode`, `asthma_exacerbation`, `copd_exacerbation`
+- `er_visit`, `hospitalised`, `mortality`, `climate_fire_trend`
+
+## Parameterization Evidence
+
+| Parameter | Value Used | Source | Year |
+| --- | --- | --- | --- |
+| Africa >70% of global burned area | Fire frequency baseline | GFEDv4, van der Werf et al. | 2017 |
+| Savanna fire PM2.5 100-300 µg/m³ during fire season | Smoke exposure | Roberts et al., Atmos Chem Phys | 2009 |
+| Forest clearing fires produce denser, localised plumes | Clearing scenario | Reddington et al. | 2015 |
+| PM2.5 exposure increases ARI risk 1.5-3x | Dose-response | WHO AQG | 2021 |
+
+## Validation Summary
+
+The 8-panel validation report (`validation_report.png`) confirms:
+
+1. **PM2.5 gradient**: Forest clearing has highest smoke PM2.5; urban haze adds ambient baseline.
+2. **Seasonality**: Clear monthly PM2.5 peaks during fire season months per scenario.
+3. **Symptoms**: Cough and eye irritation dominate; rates highest in forest clearing.
+4. **Exposure-ARI**: Higher exposure risk scores correlate with more ARI episodes.
+5. **Vulnerability**: Children <5 and elderly have highest vulnerability indices.
+6. **Health cascade**: ARI > ER visits > hospitalisation > mortality gradient realistic.
+7. **AQI**: Forest clearing has most "hazardous" days; urban haze more "unhealthy" days.
+8. **Fire proximity**: Closer fires produce higher smoke PM2.5 during fire season.
+
+![Validation Report](validation_report.png)
+
+## Usage
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("electricsheepafrica/wildfire-smoke-respiratory", name="forest_clearing_burn")
+df = ds["train"].to_pandas()
+
+# Fire season vs non-fire season ARI rates
+print(df.groupby("in_fire_season")["ari_episode"].mean())
+```
+
+## Intended Uses
+
+- Modelling wildfire smoke health impacts in SSA
+- Evaluating respiratory burden during fire seasons
+- Training exposure-response models for smoke-related respiratory outcomes
+
+## Limitations
+
+- **Synthetic data**: Not from air quality monitoring networks or clinical records.
+- **Simplified fire model**: Fire proximity and PM2.5 modeled independently.
+- **No spatial geocoding**: Scenarios proxy geography, not precise coordinates.
+- **Mask efficacy**: Modeled as binary; real-world compliance varies.
+
+## References
+
+1. van der Werf GR, et al. Global fire emissions estimates during 1997-2016. *Earth Syst Sci Data*, 2017;9:697-720.
+2. Roberts G, et al. African biomass burning emissions. *Atmos Chem Phys*, 2009.
+3. Reddington CL, et al. Air quality and health impacts of vegetation fires. *Nat Geosci*, 2015.
+4. WHO. Global Air Quality Guidelines. WHO, 2021.
+
+## Citation
+
+```bibtex
+@dataset{electricsheepafrica_wildfire_smoke_respiratory_2025,
+  title={Wildfire Smoke and Respiratory Outcomes in Sub-Saharan Africa},
+  author={Electric Sheep Africa},
+  year={2025},
+  publisher={HuggingFace},
+  url={https://huggingface.co/datasets/electricsheepafrica/wildfire-smoke-respiratory}
+}
+```
+
+## License
+
+CC-BY-4.0

generate_dataset.py

@@ -0,0 +1,224 @@
+"""Generate synthetic wildfire smoke & respiratory outcomes dataset for SSA."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+import pandas as pd
+
+SEED = 42
+N_PER_SCENARIO = 10_000
+
+YEAR_RANGE = np.arange(2010, 2025)
+YEAR_WEIGHTS = np.linspace(0.85, 1.3, len(YEAR_RANGE))
+YEAR_WEIGHTS = YEAR_WEIGHTS / YEAR_WEIGHTS.sum()
+
+SCENARIOS = {
+    "savanna_fire_belt": {
+        "fire_season_months": [11, 12, 1, 2, 3],
+        "pm25_smoke_mean": 180,
+        "pm25_smoke_sd": 90,
+        "pm25_baseline": 25,
+        "fire_frequency_mean": 3.5,
+        "fire_proximity_km_mean": 12,
+        "fire_proximity_km_sd": 8,
+        "smoke_days_mean": 45,
+        "smoke_days_sd": 18,
+        "asthma_prev": 0.08,
+        "copd_prev": 0.04,
+        "ari_rate_per1k": 85,
+        "er_visit_rate": 0.12,
+        "mortality_rate_per100k": 4.5,
+        "setting_probs": {"rural": 0.60, "peri_urban": 0.25, "urban": 0.15},
+        "mask_access": 0.05,
+    },
+    "forest_clearing_burn": {
+        "fire_season_months": [6, 7, 8, 9],
+        "pm25_smoke_mean": 250,
+        "pm25_smoke_sd": 120,
+        "pm25_baseline": 20,
+        "fire_frequency_mean": 2.0,
+        "fire_proximity_km_mean": 8,
+        "fire_proximity_km_sd": 5,
+        "smoke_days_mean": 60,
+        "smoke_days_sd": 25,
+        "asthma_prev": 0.07,
+        "copd_prev": 0.03,
+        "ari_rate_per1k": 95,
+        "er_visit_rate": 0.10,
+        "mortality_rate_per100k": 5.0,
+        "setting_probs": {"rural": 0.70, "peri_urban": 0.20, "urban": 0.10},
+        "mask_access": 0.03,
+    },
+    "urban_peri_urban_haze": {
+        "fire_season_months": [12, 1, 2, 3],
+        "pm25_smoke_mean": 120,
+        "pm25_smoke_sd": 55,
+        "pm25_baseline": 40,
+        "fire_frequency_mean": 1.5,
+        "fire_proximity_km_mean": 25,
+        "fire_proximity_km_sd": 15,
+        "smoke_days_mean": 30,
+        "smoke_days_sd": 12,
+        "asthma_prev": 0.10,
+        "copd_prev": 0.05,
+        "ari_rate_per1k": 70,
+        "er_visit_rate": 0.18,
+        "mortality_rate_per100k": 3.5,
+        "setting_probs": {"urban": 0.50, "peri_urban": 0.35, "rural": 0.15},
+        "mask_access": 0.15,
+    },
+}
+
+SCENARIO_FILES = {
+    "savanna_fire_belt": "wildfire_savanna_fire_belt.csv",
+    "forest_clearing_burn": "wildfire_forest_clearing_burn.csv",
+    "urban_peri_urban_haze": "wildfire_urban_peri_urban_haze.csv",
+}
+
+
+def _choice(rng, prob_map):
+    keys = list(prob_map.keys())
+    weights = np.array(list(prob_map.values()), dtype=float)
+    weights = weights / weights.sum()
+    return rng.choice(keys, p=weights)
+
+
+def _simulate_scenario(name, params, seed):
+    rng = np.random.default_rng(seed)
+    records = []
+
+    for idx in range(N_PER_SCENARIO):
+        year = int(rng.choice(YEAR_RANGE, p=YEAR_WEIGHTS))
+        age = int(np.clip(rng.normal(32, 18), 1, 85))
+        sex = rng.choice(["male", "female"], p=[0.48, 0.52])
+        setting = _choice(rng, params["setting_probs"])
+        is_child = int(age < 5)
+        is_elderly = int(age >= 60)
+
+        pre_existing_asthma = int(rng.random() < params["asthma_prev"])
+        pre_existing_copd = int(age > 35 and rng.random() < params["copd_prev"])
+        smoker = int(age > 15 and rng.random() < 0.12)
+
+        month = int(rng.choice(range(1, 13)))
+        in_fire_season = int(month in params["fire_season_months"])
+
+        if in_fire_season:
+            pm25_smoke = float(np.clip(rng.normal(params["pm25_smoke_mean"], params["pm25_smoke_sd"]), 20, 800))
+        else:
+            pm25_smoke = float(np.clip(rng.normal(params["pm25_baseline"], 10), 5, 60))
+
+        pm25_total = float(pm25_smoke + rng.normal(params["pm25_baseline"], 8))
+        pm25_total = float(np.clip(pm25_total, 10, 900))
+
+        fire_proximity_km = float(np.clip(
+            rng.normal(params["fire_proximity_km_mean"], params["fire_proximity_km_sd"]), 0.5, 100
+        ))
+
+        smoke_days = int(np.clip(
+            rng.normal(params["smoke_days_mean"], params["smoke_days_sd"]) * (1.2 if in_fire_season else 0.3),
+            0, 120,
+        ))
+
+        fire_count = int(np.clip(rng.poisson(params["fire_frequency_mean"]), 0, 15))
+
+        aqi_category = (
+            "hazardous" if pm25_total > 250 else
+            "very_unhealthy" if pm25_total > 150 else
+            "unhealthy" if pm25_total > 55 else
+            "moderate" if pm25_total > 35 else
+            "good"
+        )
+
+        mask_use = int(rng.random() < params["mask_access"] * (1.5 if in_fire_season else 0.5))
+        stayed_indoors = int(in_fire_season and rng.random() < 0.20)
+
+        vulnerability = (
+            0.15
+            + is_child * 0.25
+            + is_elderly * 0.20
+            + pre_existing_asthma * 0.15
+            + pre_existing_copd * 0.15
+            + smoker * 0.10
+        )
+        vulnerability = float(np.clip(vulnerability, 0, 1))
+
+        exposure_risk = float(np.clip(
+            (pm25_total / 300) * (1 - mask_use * 0.3) * (1 - stayed_indoors * 0.4) * vulnerability * 2,
+            0, 1,
+        ))
+
+        cough = int(rng.random() < 0.15 + exposure_risk * 0.35)
+        wheeze = int(rng.random() < 0.08 + exposure_risk * 0.25)
+        dyspnoea = int(rng.random() < 0.05 + exposure_risk * 0.20)
+        eye_irritation = int(in_fire_season and rng.random() < 0.20 + exposure_risk * 0.25)
+
+        ari_episode = int(rng.random() < params["ari_rate_per1k"] / 1000 * (1 + exposure_risk * 2))
+        asthma_exacerbation = int(pre_existing_asthma and rng.random() < 0.15 + exposure_risk * 0.3)
+        copd_exacerbation = int(pre_existing_copd and rng.random() < 0.10 + exposure_risk * 0.25)
+
+        er_visit = int(
+            (ari_episode or asthma_exacerbation or copd_exacerbation)
+            and rng.random() < params["er_visit_rate"]
+        )
+        hospitalised = int(er_visit and rng.random() < 0.25)
+
+        mortality = int(
+            hospitalised and rng.random() < params["mortality_rate_per100k"] / 100_000 * 500
+        )
+
+        climate_fire_trend = float(np.clip(0.015 * (year - 2010) + rng.normal(0, 0.005), -0.02, 0.1))
+
+        record = {
+            "record_id": f"{name[:3].upper()}-{idx:05d}",
+            "scenario": name,
+            "year": year,
+            "month": month,
+            "in_fire_season": in_fire_season,
+            "age": age,
+            "sex": sex,
+            "setting": setting,
+            "is_child_u5": is_child,
+            "is_elderly_60plus": is_elderly,
+            "pre_existing_asthma": pre_existing_asthma,
+            "pre_existing_copd": pre_existing_copd,
+            "smoker": smoker,
+            "pm25_smoke_ugm3": round(pm25_smoke, 1),
+            "pm25_total_ugm3": round(pm25_total, 1),
+            "aqi_category": aqi_category,
+            "fire_proximity_km": round(fire_proximity_km, 1),
+            "smoke_days_per_season": smoke_days,
+            "fire_count": fire_count,
+            "mask_use": mask_use,
+            "stayed_indoors": stayed_indoors,
+            "vulnerability_index": round(vulnerability, 2),
+            "exposure_risk_score": round(exposure_risk, 3),
+            "cough": cough,
+            "wheeze": wheeze,
+            "dyspnoea": dyspnoea,
+            "eye_irritation": eye_irritation,
+            "ari_episode": ari_episode,
+            "asthma_exacerbation": asthma_exacerbation,
+            "copd_exacerbation": copd_exacerbation,
+            "er_visit": er_visit,
+            "hospitalised": hospitalised,
+            "mortality": mortality,
+            "climate_fire_trend": round(climate_fire_trend, 4),
+        }
+        records.append(record)
+
+    return pd.DataFrame(records)
+
+
+def main():
+    output_dir = Path("data")
+    output_dir.mkdir(parents=True, exist_ok=True)
+    for idx, (name, params) in enumerate(SCENARIOS.items()):
+        df = _simulate_scenario(name, params, SEED + idx * 211)
+        df.to_csv(output_dir / SCENARIO_FILES[name], index=False)
+        print(f"Saved {name} -> {SCENARIO_FILES[name]}")
+
+
+if __name__ == "__main__":
+    main()

requirements.txt

@@ -0,0 +1,3 @@
+numpy>=1.24
+pandas>=2.0
+matplotlib>=3.7

validate_dataset.py

@@ -0,0 +1,122 @@
+"""Validate synthetic wildfire smoke & respiratory outcomes dataset."""
+
+from __future__ import annotations
+
+from pathlib import Path
+
+import matplotlib.pyplot as plt
+import pandas as pd
+
+SCENARIO_FILES = {
+    "savanna_fire_belt": "wildfire_savanna_fire_belt.csv",
+    "forest_clearing_burn": "wildfire_forest_clearing_burn.csv",
+    "urban_peri_urban_haze": "wildfire_urban_peri_urban_haze.csv",
+}
+
+COLORS = {
+    "savanna_fire_belt": "#e6550d",
+    "forest_clearing_burn": "#31a354",
+    "urban_peri_urban_haze": "#756bb1",
+}
+
+
+def load_data() -> pd.DataFrame:
+    frames = []
+    for scenario, filename in SCENARIO_FILES.items():
+        df = pd.read_csv(Path("data") / filename)
+        frames.append(df)
+    return pd.concat(frames, ignore_index=True)
+
+
+def plot_validation(df: pd.DataFrame, output_path: Path) -> None:
+    fig, axes = plt.subplots(4, 2, figsize=(14, 16))
+    axes = axes.flatten()
+
+    # Panel 1: PM2.5 total by scenario
+    for s in SCENARIO_FILES:
+        subset = df[df["scenario"] == s]
+        axes[0].hist(subset["pm25_total_ugm3"], bins=40, alpha=0.5, color=COLORS[s], label=s)
+    axes[0].set_title("PM2.5 Total Distribution by Scenario")
+    axes[0].set_xlabel("PM2.5 (µg/m³)")
+    axes[0].legend(fontsize=7)
+
+    # Panel 2: Seasonality - PM2.5 by month
+    for s in SCENARIO_FILES:
+        subset = df[df["scenario"] == s]
+        monthly = subset.groupby("month")["pm25_total_ugm3"].mean()
+        axes[1].plot(monthly.index, monthly.values, marker="o", color=COLORS[s], label=s)
+    axes[1].set_title("Monthly PM2.5 Trend")
+    axes[1].set_xlabel("Month")
+    axes[1].set_ylabel("Mean PM2.5")
+    axes[1].legend(fontsize=7)
+
+    # Panel 3: Respiratory symptoms by scenario
+    symptom_cols = ["cough", "wheeze", "dyspnoea", "eye_irritation"]
+    symptom_rates = df.groupby("scenario")[symptom_cols].mean() * 100
+    symptom_rates.plot(kind="bar", ax=axes[2])
+    axes[2].set_title("Respiratory Symptom Prevalence (%)")
+    axes[2].set_ylabel("Percent")
+    axes[2].legend(fontsize=7)
+
+    # Panel 4: Exposure risk vs ARI
+    for s in SCENARIO_FILES:
+        subset = df[df["scenario"] == s]
+        axes[3].scatter(
+            subset["exposure_risk_score"], subset["ari_episode"],
+            s=6, alpha=0.1, color=COLORS[s], label=s,
+        )
+    axes[3].set_title("Exposure Risk Score vs ARI Episode")
+    axes[3].set_xlabel("Exposure risk score")
+    axes[3].set_ylabel("ARI episode")
+    axes[3].legend(fontsize=7)
+
+    # Panel 5: Vulnerability index distribution
+    for s in SCENARIO_FILES:
+        subset = df[df["scenario"] == s]
+        axes[4].hist(subset["vulnerability_index"], bins=20, alpha=0.5, color=COLORS[s], label=s)
+    axes[4].set_title("Vulnerability Index Distribution")
+    axes[4].set_xlabel("Vulnerability index")
+    axes[4].legend(fontsize=7)
+
+    # Panel 6: Health outcomes cascade
+    cascade_cols = ["ari_episode", "er_visit", "hospitalised", "mortality"]
+    cascade = df.groupby("scenario")[cascade_cols].mean() * 100
+    cascade.plot(kind="bar", ax=axes[5])
+    axes[5].set_title("Health Outcomes Cascade (%)")
+    axes[5].set_ylabel("Percent")
+    axes[5].legend(fontsize=7)
+
+    # Panel 7: AQI category distribution
+    aqi_counts = df.groupby(["scenario", "aqi_category"]).size().groupby(level=0).apply(lambda s: s / s.sum())
+    aqi_counts.unstack().plot(kind="bar", stacked=True, ax=axes[6])
+    axes[6].set_title("AQI Category Distribution")
+    axes[6].set_ylabel("Share")
+    axes[6].legend(fontsize=6)
+
+    # Panel 8: Fire proximity vs PM2.5
+    fire_season = df[df["in_fire_season"] == 1]
+    for s in SCENARIO_FILES:
+        subset = fire_season[fire_season["scenario"] == s]
+        axes[7].scatter(
+            subset["fire_proximity_km"], subset["pm25_smoke_ugm3"],
+            s=6, alpha=0.15, color=COLORS[s], label=s,
+        )
+    axes[7].set_title("Fire Proximity vs Smoke PM2.5 (fire season)")
+    axes[7].set_xlabel("Fire proximity (km)")
+    axes[7].set_ylabel("PM2.5 smoke (µg/m³)")
+    axes[7].legend(fontsize=7)
+
+    plt.tight_layout()
+    fig.savefig(output_path, dpi=200)
+    plt.close(fig)
+
+
+def main() -> None:
+    df = load_data()
+    output_path = Path("validation_report.png")
+    plot_validation(df, output_path)
+    print(f"Saved {output_path}")
+
+
+if __name__ == "__main__":
+    main()