# CPCB Ground-Truth Cleaning Report

- Input: `merged_cpcb_data.csv` — **1,549,637 rows**, 34 columns
- Output: `merged_cpcb_clean.csv` — **972,562 rows**, 20 columns
- Net removed: **577,075 rows** (37.2%)

## Rows removed, by reason
| step | rows |
|---|---|
| dropped rows with no lat/lon (unusable in node graph) | 26,296 |
| negative concentrations set to NULL | 0 |
| dropped rows where ALL 7 target pollutants are null | 546,277 |
| dropped duplicate (station,lat,lon,date) rows | 4,502 |

## Columns dropped (garbage / redundant / met)
`SourceFile`, `FileYear`, `State_Raw`, `AT (°C)`, `RH (%)`, `WS (m/s)`, `WD (deg)`, `RF (mm)`, `TOT-RF (mm)`, `SR (W/mt2)`, `BP (mmHg)`, `VWS (m/s)`, `Year`, `Month`

## Kept columns
`date`, `station`, `city`, `state`, `state_hint`, `state_mismatch_flag`, `geocode_confidence`, `agency`, `lat`, `lon`, `coverage_area_km2`, `pm25`, `pm10`, `no2`, `so2`, `co`, `o3`, `nh3`, `no`, `nox_ppb`

## Pollutant coverage after cleaning (non-null %)
| pollutant | non-null % | target? |
|---|---|---|
| pm25 | 93.2% | yes |
| pm10 | 88.5% | yes |
| no2 | 96.6% | yes |
| so2 | 94.1% | yes |
| co | 94.6% | yes |
| o3 | 91.7% | yes |
| nh3 | 80.1% | yes |
| no | 95.0% | aux |
| nox_ppb | 96.0% | aux |

## Year coverage
| year | rows |
|---|---|
| 2009 | 4,057 |
| 2010 | 4,725 |
| 2011 | 6,596 |
| 2012 | 6,700 |
| 2013 | 6,094 |
| 2014 | 5,553 |
| 2015 | 8,524 |
| 2016 | 12,644 |
| 2017 | 18,444 |
| 2018 | 39,186 |
| 2019 | 60,018 |
| 2020 | 76,712 |
| 2021 | 96,640 |
| 2022 | 115,967 |
| 2023 | 156,676 |
| 2024 | 177,757 |
| 2025 | 176,269 |

## Geocode confidence (rows)
| confidence | rows |
|---|---|
| high | 841,064 |
| medium_state_mismatch | 126,158 |
| low_ambiguous | 5,340 |

> `not_found` cities were dropped (no coordinates). Hand-add lat/lon for any that matter: Bileipada, DU Delhi, Kunjemura, Milupara, Palkalaiperur, Suakati, Tumidih.

## State mismatches
- **9,588 rows** flagged (folder state ≠ geocoded state). `state` = geocoded (primary); `state_hint` = folder; flag kept. **Do not auto-resolve** — mixed cause: some folder-wrong (Bhilai/Bilaspur/Korba → Chhattisgarh, geocode correct), some geocode-wrong (Chhal, Panchgaon). Manual review needed; coordinates (not state) drive the node graph.

## Notes
- Pollutant NULLs are **left null** (missing ≠ 0). Do not impute; they mask in training.
- `nox_ppb` is in **ppb** (all others µg/m³, CO in mg/m³) — unit differs, kept separate.
- Timestamps treated as IST calendar dates (CPCB is local IST) — aligns with `node_observations_daily` date convention.