File size: 3,524 Bytes
bdc5bb1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 | """
preprocess.py -> data\ingest\ (run: python data\ingest\preprocess.py)
Turns the raw master into a TRAINING-READY dataset. Per month partition:
1. CLIP each feature to a physical envelope (winsorises bad-pixel tails —
e.g. modis_aod 11 -> 5, s5p_aer_ai 151 -> 20 — without touching real highs).
2. Add <source>_present missingness flags (blueprint Group A) — model must
know which sparse satellite inputs are present (fill ranges 21-89%).
3. LEFT-JOIN CPCB labels (cpcb_labels_daily.parquet) -> 7 targets + `labeled`.
Output: data\ingest\train_ready\<YYYY-MM>.parquet (features + flags + targets)
After this, only training is left.
"""
from __future__ import annotations
import os, sys, pathlib
import numpy as np, pandas as pd
OUT = pathlib.Path(os.environ.get("ATMOS_OUT", "data/ingest"))
MASTER = OUT / "node_observations_daily"
DEST = OUT / "train_ready"; DEST.mkdir(exist_ok=True)
# physical clip envelopes (keep real extremes, cap absurd sensor spikes)
CLIP = {
"s5p_no2_column":(-5e-4,5e-3), "s5p_so2_column":(-0.05,0.5), "s5p_co_column":(0,1.0),
"s5p_o3_column":(0,0.3), "s5p_hcho_column":(-0.05,0.05), "s5p_aer_ai":(-15,20),
"s5p_ch4_column":(1400,2200), "s5p_cloud_fraction":(0,1),
"modis_aod":(0,5), "viirs_aod":(0,5), "fire_count":(0,None),
"era5_temperature":(220,330), "era5_boundary_layer_height":(0,5000),
"era5_relative_humidity":(0,100), "era5_surface_pressure":(40000,105000),
"era5_cloud_cover":(0,1), "era5_wind_u":(-60,60), "era5_wind_v":(-60,60),
"era5_rainfall":(0,None), "u_1330ist":(-60,60), "v_1330ist":(-60,60),
}
# sparse satellite sources -> missingness flags
FLAG_COLS = ["s5p_no2_column","s5p_co_column","s5p_so2_column","s5p_o3_column",
"s5p_hcho_column","s5p_aer_ai","s5p_ch4_column","s5p_cloud_fraction",
"modis_aod","viirs_aod","fire_count"]
TARGETS = ["pm2_5","pm10","no2","so2","co","o3","nh3"]
def main():
cpcb_path = OUT / "cpcb_labels_daily.parquet"
cpcb = pd.read_parquet(cpcb_path) if cpcb_path.exists() else None
if cpcb is not None:
cpcb = cpcb[["node_id","date"] + TARGETS + ["n_stations"]]
print(f"CPCB labels: {len(cpcb):,} labeled node-days joined")
else:
print("!! cpcb_labels_daily.parquet not found — writing features only (run ingest_cpcb.py first)")
parts = sorted(MASTER.glob("*.parquet"))
print(f"{len(parts)} partitions -> {DEST}")
tot=0; labeled=0
for pf in parts:
df = pd.read_parquet(pf)
# 1. clip
for col,(lo,hi) in CLIP.items():
if col in df:
df[col] = df[col].clip(lower=lo, upper=hi)
# 2. missingness flags
for col in FLAG_COLS:
if col in df:
df[col+"_present"] = df[col].notna().astype("int8")
# 3. join labels
if cpcb is not None:
df = df.merge(cpcb, on=["node_id","date"], how="left")
df["labeled"] = df[TARGETS].notna().any(axis=1).astype("int8")
labeled += int(df["labeled"].sum())
df.to_parquet(DEST / pf.name, index=False)
tot += len(df)
print(f" {pf.stem}: {len(df):,} rows", end="\r")
print(f"\ndone. {tot:,} node-days -> {DEST}")
if cpcb is not None:
print(f"labeled node-days (have >=1 CPCB target): {labeled:,} ({100*labeled/tot:.2f}%)")
print(" -> supervised loss uses labeled rows; SSL pretrain uses ALL rows.")
if __name__ == "__main__":
main() |