File size: 9,921 Bytes
463c895 | 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 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 | """
aggregation.py -> data/ingest/
Phase 2 ingestion, LOCKED to the real HuggingFace schemas (audit 2026-07).
raw source (CSV/parquet) -> grid-aligned table:
node_id | date | <value cols> | n_obs | centroid_lat | centroid_lon
Audit-driven fixes baked in:
- per-source MISSING SENTINELS (VIIRS -999, S5P 9.969e36) -> NaN before aggregating
- per-source lat/lon column names (MODIS/VIIRS use lat/lon)
- timestamp -> date derivation (S5P/ERA5 use time_utc)
- ERA5 hourly -> daily collapse, + a separate 13:30-IST wind spec
- nonneg filter (MODIS uncertainty had 67 negative rows)
"""
from __future__ import annotations
import sys, pathlib
sys.path.insert(0, str(pathlib.Path(__file__).resolve().parents[1] / "grid"))
import numpy as np
import pandas as pd
from dataclasses import dataclass, field
from scipy.spatial import cKDTree
from pyproj import Transformer
from grid_index import GridIndex
S5P_FILL_VALUE = 9.969209968386869e+36
IST_OFFSET_H = 5.5 # UTC -> IST
@dataclass
class SourceSpec:
name: str
mode: str # "containment" | "centroid"
value_cols: list
lat_col: str = "latitude"
lon_col: str = "longitude"
date_col: str | None = "date" # used if timestamp_col is None
timestamp_col: str | None = None # if set, date derived from this (UTC)
agg: str = "mean" # mean|median|max|min|sum|count
sentinels: list = field(default_factory=list) # values -> NaN (e.g. -999)
qa_col: str | None = None
qa_min: float | None = None
nonneg_cols: list = field(default_factory=list) # drop rows where these < 0
ist_target_h: float | None = None # select row nearest this IST hour (13.5 = 13:30)
pre_daily_mean: bool = False # collapse (lat,lon,date) before centroid sample
max_sample_km: float | None = None
rename: dict = field(default_factory=dict)
def _to_date(df: pd.DataFrame, spec: SourceSpec) -> pd.DataFrame:
df = df.copy()
if spec.timestamp_col is not None:
ts = pd.to_datetime(df[spec.timestamp_col], utc=True, errors="coerce")
ist = ts + pd.Timedelta(hours=IST_OFFSET_H)
df["date"] = ist.dt.strftime("%Y-%m-%d")
df["_ist_h"] = ist.dt.hour + ist.dt.minute / 60.0 # fractional IST hour
else:
df["date"] = pd.to_datetime(df[spec.date_col], errors="coerce").dt.strftime("%Y-%m-%d")
return df.dropna(subset=["date"])
def _select_nearest_ist(df: pd.DataFrame, spec: SourceSpec) -> pd.DataFrame:
"""Keep, per (lat, lon, date), the single row whose IST time is nearest the
target (e.g. 13:30 for S5P-aligned wind). Robust to hourly/3-hourly cadence."""
if spec.ist_target_h is None or "_ist_h" not in df:
return df
d = df.assign(_gap=(df["_ist_h"] - spec.ist_target_h).abs())
d = d.sort_values("_gap").drop_duplicates([spec.lat_col, spec.lon_col, "date"], keep="first")
return d.drop(columns=["_gap"])
def _clean(df: pd.DataFrame, spec: SourceSpec) -> pd.DataFrame:
df = df.copy()
# 1. sentinels -> NaN (CRITICAL: VIIRS -999, S5P fill). Must run BEFORE dropna.
sents = list(spec.sentinels) + [S5P_FILL_VALUE]
for c in spec.value_cols:
df[c] = df[c].replace(sents, np.nan)
# 2. nonneg filter (MODIS uncertainty)
for c in spec.nonneg_cols:
if c in df: df = df[(df[c] >= 0) | df[c].isna()]
# 3. drop rows with no valid value
df = df.dropna(subset=spec.value_cols, how="all") # only drop if ALL value cols null
# 4. QA gate
if spec.qa_col and spec.qa_min is not None:
df = df[df[spec.qa_col] >= spec.qa_min]
return df
def _finalize(out, gi, spec, dropped=None):
out = out.rename(columns=spec.rename)
out = out.merge(gi.centroids(), on="node_id", how="left").reset_index(drop=True)
if dropped is not None: out.attrs["dropped_outside_grid"] = int(dropped)
return out
def _containment(df, gi, spec):
df = df.copy()
df["node_id"] = gi.assign_node_id(df[spec.lat_col].values, df[spec.lon_col].values)
dropped = int(df["node_id"].isna().sum())
df = df[df["node_id"].notna()]
if len(df) == 0:
raise ValueError(f"[{spec.name}] nothing landed in grid")
g = df.groupby(["node_id", "date"])
if spec.agg == "count":
out = g.size().to_frame("count")
for c in spec.value_cols: out[c] = g[c].mean()
else:
out = g.agg({c: spec.agg for c in spec.value_cols})
out["n_obs"] = g.size()
out = out.reset_index()
return _finalize(out, gi, spec, dropped)
def _centroid(df, gi, spec):
if spec.pre_daily_mean: # collapse hourly -> one row per (lat,lon,date)
keys = [spec.lat_col, spec.lon_col, "date"]
df = df.groupby(keys, as_index=False)[spec.value_cols].mean()
fwd = Transformer.from_crs("EPSG:4326", gi.proj4, always_xy=True)
cent = gi.centroids()
cx, cy = fwd.transform(cent["centroid_lon"].values, cent["centroid_lat"].values)
cxy = np.column_stack([cx, cy])
frames = []
for date, sub in df.groupby("date"):
sx, sy = fwd.transform(sub[spec.lon_col].values, sub[spec.lat_col].values)
tree = cKDTree(np.column_stack([sx, sy]))
dist_m, idx = tree.query(cxy, k=1)
rec = {"node_id": cent["node_id"].values, "date": date, "n_obs": 1}
for c in spec.value_cols:
v = sub[c].values[idx].astype(float)
if spec.max_sample_km is not None:
v[dist_m > spec.max_sample_km * 1000.0] = np.nan
rec[c] = v
frames.append(pd.DataFrame(rec))
if not frames:
raise ValueError(f"[{spec.name}] no rows survived filtering (check ist_target_h / sentinels / qa)")
return _finalize(pd.concat(frames, ignore_index=True), gi, spec)
def ingest(raw, gi, spec):
df = _clean(_to_date(raw, spec), spec)
df = _select_nearest_ist(df, spec)
if len(df) == 0:
raise ValueError(f"[{spec.name}] empty after clean/filter")
if spec.mode == "containment": return _containment(df, gi, spec)
if spec.mode == "centroid": return _centroid(df, gi, spec)
raise ValueError(f"bad mode {spec.mode!r}")
# ==========================================================================
# REAL SPECS — locked to the HuggingFace audit schemas.
# S5P: one spec PER pollutant (each folder has a generic `value` column).
# ==========================================================================
SPECS = {
# ---- S5P TROPOMI (per-pollutant files, col `value`, time_utc, qa_value) ----
"s5p_no2": SourceSpec("s5p_no2","containment", value_cols=["value"],
timestamp_col="time_utc", qa_col="qa_value", qa_min=0.75,
rename={"value":"s5p_no2_column"}),
"s5p_co": SourceSpec("s5p_co","containment", value_cols=["value"],
timestamp_col="time_utc", qa_col="qa_value", qa_min=0.5,
rename={"value":"s5p_co_column"}),
"s5p_so2": SourceSpec("s5p_so2","containment", value_cols=["value"],
timestamp_col="time_utc", qa_col="qa_value", qa_min=0.5,
rename={"value":"s5p_so2_column"}),
"s5p_ch4": SourceSpec("s5p_ch4","containment", value_cols=["value"],
timestamp_col="time_utc",
rename={"value":"s5p_ch4_column"}), # ch4 qa_threshold null (guide)
"s5p_o3": SourceSpec("s5p_o3","containment", value_cols=["value"],
timestamp_col="time_utc", qa_col="qa_value", qa_min=0.5,
rename={"value":"s5p_o3_column"}), # OUTPUT pollutant; S5P-O3 input OPTIONAL (labels from CPCB)
"s5p_hcho": SourceSpec("s5p_hcho","containment", value_cols=["value"],
timestamp_col="time_utc", qa_col="qa_value", qa_min=0.5,
rename={"value":"s5p_hcho_column"}), # INPUT-ONLY (blueprint Group A, biomass signal) — REQUIRED input
"s5p_cloud": SourceSpec("s5p_cloud","containment", value_cols=["value"],
timestamp_col="time_utc",
rename={"value":"s5p_cloud_fraction"}), # QA / cloud-fraction input
"s5p_aer_ai": SourceSpec("s5p_aer_ai","containment", value_cols=["value"],
timestamp_col="time_utc", qa_col="qa_value", qa_min=0.80,
rename={"value":"s5p_aer_ai"}),
# ---- MODIS AOD (lat/lon, ~70% null, negative uncertainty) ----
"modis": SourceSpec("modis","containment", value_cols=["modis_aod"],
lat_col="lat", lon_col="lon", date_col="date",
nonneg_cols=["modis_aod_uncertainty"]),
# ---- VIIRS AOD (lat/lon, -999 SENTINEL) ----
"viirs": SourceSpec("viirs","containment", value_cols=["viirs_aod"],
lat_col="lat", lon_col="lon", date_col="date",
sentinels=[-999.0]),
# ---- ERA5 (hourly 0.25deg grid) -> daily-mean node features, centroid sample ----
"era5_daily": SourceSpec("era5_daily","centroid",
value_cols=["era5_temperature","era5_boundary_layer_height",
"era5_relative_humidity","era5_surface_pressure",
"era5_cloud_cover","era5_wind_u","era5_wind_v","era5_rainfall"],
# era5_solar_radiation EXCLUDED: all-NaN in data (would empty every row). Re-add if fixed.
timestamp_col="time_utc", pre_daily_mean=True, max_sample_km=30),
# ---- ERA5 13:30-IST wind ONLY (edge-term input; u/v kept as components) ----
"era5_wind_1330": SourceSpec("era5_wind_1330","centroid",
value_cols=["era5_wind_u","era5_wind_v"],
timestamp_col="time_utc", ist_target_h=13.5, max_sample_km=30, # nearest-to-13:30 from RAW hourly
rename={"era5_wind_u":"u_1330ist","era5_wind_v":"v_1330ist"}),
# ---- FIRMS fire -> count per cell/day (acq_date) ----
"firms": SourceSpec("firms","containment", value_cols=["frp"],
lat_col="latitude", lon_col="longitude", date_col="acq_date",
agg="count", rename={"count":"fire_count"}),
}
# DEM (Copernicus GLO-30 GeoTIFF): NOT here — raster reproject to LAEA in static_gis.py. |