Spaces:
Sleeping
Sleeping
File size: 14,747 Bytes
a74054f f2046b4 a74054f d6b0b7a f2046b4 d6b0b7a a74054f d6b0b7a a74054f d6b0b7a a74054f d6b0b7a a74054f d6b0b7a a74054f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 | """
ERA5 meteorological reanalysis data loader (SAFRAN alternative).
Single Responsibility: Load and parse ERA5 data with spatial interpolation.
"""
import pandas as pd
import xarray as xr
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.dates as mdates
from pathlib import Path
from typing import Optional, List
from .base import BaseDataLoader
class SAFRANLoader(BaseDataLoader):
"""
Loads ERA5 reanalysis data (used as SAFRAN alternative).
Handles actual ERA5 NetCDF structure:
- Gridded at 0.25° resolution
- Variables: t2m, tp, pev, ssrd, u10, v10, sf, ro
- Interpolates to station points
"""
# Variable mapping
VAR_MAP = {
't2m': 'temp_2m_K',
'tp': 'precip_m',
'pev': 'evap_m',
'ssrd': 'solar_Jm2',
'u10': 'wind_u_ms',
'v10': 'wind_v_ms',
'sf': 'snow_m',
'ro': 'runoff_m'
}
def __init__(
self,
data_path: Path,
station_coords: Optional[pd.DataFrame] = None,
interp_method: str = "nearest"
):
"""
Initialize ERA5 loader.
Args:
data_path: Path to datasets/safran directory
station_coords: DataFrame with [station_code, lat, lon]
interp_method: "nearest" or "linear"
"""
super().__init__(data_path)
self.station_coords = station_coords
self.interp_method = interp_method
def load(self) -> pd.DataFrame:
"""
Load ERA5 data and interpolate to stations.
Returns:
DataFrame with [date, station_code, temp_2m_K, precip_m, ...]
"""
nc_files = sorted(Path(self.data_path).glob("era5_*.nc"))
if not nc_files:
raise FileNotFoundError(f"No ERA5 files in {self.data_path}")
all_dfs = []
for nc_file in nc_files:
df = self._load_and_interpolate(nc_file)
all_dfs.append(df)
combined = pd.concat(all_dfs, ignore_index=True)
return combined.sort_values(['date', 'station_code']).reset_index(drop=True)
def load_grid(self, bbox: Optional[tuple] = None) -> pd.DataFrame:
"""
Load ERA5 at its NATIVE grid resolution (0.25 degrees) -- no
per-station interpolation at all, unlike load(). Confirmed
against real data that station-level interpolation duplicates
signal: multiple real stations landed on the exact identical
grid cell (several stations sharing bit-identical wind_speed_ms
values across the whole real climatology test), since ERA5's
0.25 deg cells are coarse relative to how tightly some real
gauges cluster. Modeling at grid-cell resolution avoids this
redundancy directly, rather than training separate models on
what's provably the same input.
Args:
bbox: optional (min_lon, min_lat, max_lon, max_lat) to
subset the grid before loading -- the whole point is to
avoid loading the full national/continental ERA5 extent
when only a couple dozen cells actually cover the study
area. Defaults to no subsetting (the full extent
already present in the downloaded files, which
download_era5_full.py already bounded at fetch time).
Returns:
DataFrame with [date, grid_lat, grid_lon, temp_2m_K,
precip_m, ...] -- one row per (date, cell), long format
(not yet pivoted to wide per-variable, since the number of
variables and their names come from VAR_MAP and shouldn't
be hardcoded here).
"""
nc_files = sorted(Path(self.data_path).glob("era5_*.nc"))
if not nc_files:
raise FileNotFoundError(f"No ERA5 files in {self.data_path}")
all_dfs = []
for nc_file in nc_files:
df = self._load_grid_single_file(nc_file, bbox)
all_dfs.append(df)
combined = pd.concat(all_dfs, ignore_index=True)
return combined.sort_values(['date', 'grid_lat', 'grid_lon']).reset_index(drop=True)
def _load_grid_single_file(self, nc_file: Path, bbox: Optional[tuple]) -> pd.DataFrame:
"""Load a single NetCDF file's full grid (optionally bbox-subset),
no station interpolation. Every (lat, lon, date) triple becomes
one row -- this is real gridded data, not point samples."""
ds = xr.open_dataset(nc_file)
time_coord = 'valid_time' if 'valid_time' in ds.coords else 'time'
if bbox is not None:
min_lon, min_lat, max_lon, max_lat = bbox
# ERA5 latitude is typically stored descending (90 to -90);
# slice() requires matching that order or it silently
# returns nothing, which is exactly the kind of thing that
# produced a wrong-axis-order failure earlier in this
# project (BD TOPO's WFS) -- check the actual coordinate
# direction rather than assume one.
lat_vals = ds['latitude'].values
if lat_vals[0] > lat_vals[-1]:
ds = ds.sel(latitude=slice(max_lat, min_lat), longitude=slice(min_lon, max_lon))
else:
ds = ds.sel(latitude=slice(min_lat, max_lat), longitude=slice(min_lon, max_lon))
if ds.sizes.get('latitude', 0) == 0 or ds.sizes.get('longitude', 0) == 0:
raise ValueError(
f"bbox {bbox} selected zero grid cells from {nc_file.name} -- check the bbox "
f"is within this file's real coordinate range: "
f"lat {ds['latitude'].values.min() if ds.sizes.get('latitude',1) else 'N/A'}.."
f"{ds['latitude'].values.max() if ds.sizes.get('latitude',1) else 'N/A'}, "
f"lon similarly."
)
df = ds.to_dataframe().reset_index()
df = df.rename(columns={time_coord: 'date', 'latitude': 'grid_lat', 'longitude': 'grid_lon', **self.VAR_MAP})
keep_cols = ['date', 'grid_lat', 'grid_lon'] + [v for v in self.VAR_MAP.values() if v in df.columns]
df = df[keep_cols].copy()
df['date'] = pd.to_datetime(df['date']).dt.tz_localize(None)
ds.close()
return df
def _load_and_interpolate(self, nc_file: Path) -> pd.DataFrame:
"""
Load single NetCDF and interpolate to ALL stations in one
vectorized xarray selection, not one .sel() + .to_dataframe()
call per station.
The original per-station loop does real work with meaningful
per-call overhead (an index lookup, then a full DataFrame
conversion) once per station PER FILE. At 27 stations across
~67 year-files (1960-2026) that's ~1,800 calls -- slow but
tolerable. At the reach graph's ~2,900 nodes it's ~193,000
individual calls, which is what was actually hanging, not a
proportionally-worse runtime.
Fix: xarray's vectorized ("pointwise") indexing -- when the
latitude/longitude indexers are DataArrays sharing a common
dimension name ("station" here), a single .sel() call looks up
every point at once instead of one point per call, producing a
result with a "station" dimension instead of separate
latitude/longitude dimensions. This turns ~193,000 calls into
~67 (one per file).
Reasoned from documented xarray API behavior, not verified
against a real NetCDF file with real execution -- test against
a small subset of the era5_*.nc files first, and confirm the
output matches the old per-station-loop version's shape/values
before trusting it for real training data.
"""
ds = xr.open_dataset(nc_file)
time_coord = 'valid_time' if 'valid_time' in ds.coords else 'time'
station_lats = xr.DataArray(self.station_coords['lat'].values, dims='station')
station_lons = xr.DataArray(self.station_coords['lon'].values, dims='station')
station_codes = self.station_coords['station_code'].values
point_ds = ds.sel(latitude=station_lats, longitude=station_lons, method=self.interp_method)
# point_ds now has a 'station' dimension (length n_stations) instead
# of separate latitude/longitude dimensions -- attach real station
# codes as a coordinate so they survive the to_dataframe() below,
# rather than staying as anonymous integer positions.
point_ds = point_ds.assign_coords(station=station_codes)
df = point_ds.to_dataframe().reset_index()
df = df.rename(columns={time_coord: 'date', 'station': 'station_code', **self.VAR_MAP})
keep_cols = ['date', 'station_code'] + [v for v in self.VAR_MAP.values() if v in df.columns]
df = df[keep_cols].copy()
df['date'] = pd.to_datetime(df['date']).dt.tz_localize(None)
ds.close()
return df
@staticmethod
def convert_units(df: pd.DataFrame) -> pd.DataFrame:
"""Convert ERA5 units to standard (°C, mm, W/m²)."""
df = df.copy()
# K → °C
if 'temp_2m_K' in df.columns:
df['temp_C'] = df['temp_2m_K'] - 273.15
# m → mm
for var in ['precip_m', 'evap_m', 'snow_m', 'runoff_m']:
if var in df.columns:
df[var.replace('_m', '_mm')] = df[var] * 1000
# J/m² → W/m²
if 'solar_Jm2' in df.columns:
df['solar_Wm2'] = df['solar_Jm2'] / 86400
# Wind speed
if 'wind_u_ms' in df.columns and 'wind_v_ms' in df.columns:
df['wind_speed_ms'] = np.sqrt(df['wind_u_ms']**2 + df['wind_v_ms']**2)
return df
def get_metadata(self) -> dict:
"""Get ERA5 metadata."""
meta = super().get_metadata()
meta.update({
"data_type": "era5_reanalysis",
"source": "Copernicus CDS",
"resolution": "0.25 degrees",
"interp_method": self.interp_method
})
return meta
def plot_temperature(
self,
df: Optional[pd.DataFrame] = None,
stations: Optional[List[str]] = None,
figsize: tuple = (12, 5),
save_path: Optional[Path] = None,
) -> plt.Axes:
"""
Plot 2m air temperature time series, one line per station.
Args:
df: Optional pre-loaded DataFrame (raw or already converted). If not
provided, data is loaded and unit-converted from disk.
stations: Optional list of station codes to plot. Defaults to all.
figsize: Figure size in inches.
save_path: If provided, saves the figure to this path.
Returns:
The matplotlib Axes object.
"""
if df is None:
df = self.convert_units(self.load())
elif "temp_C" not in df.columns and "temp_2m_K" in df.columns:
df = self.convert_units(df)
plot_df = df.dropna(subset=["temp_C"])
if stations:
plot_df = plot_df[plot_df["station_code"].isin(stations)]
if plot_df.empty:
raise ValueError("No temperature data available to plot")
fig, ax = plt.subplots(figsize=figsize)
for station_code, group in plot_df.groupby("station_code"):
group = group.sort_values("date")
ax.plot(group["date"], group["temp_C"], linewidth=1, label=station_code)
ax.set_title("2m Air Temperature")
ax.set_xlabel("Date")
ax.set_ylabel("Temperature (°C)")
ax.xaxis.set_major_locator(mdates.AutoDateLocator())
ax.xaxis.set_major_formatter(mdates.ConciseDateFormatter(ax.xaxis.get_major_locator()))
ax.legend(title="Station", fontsize=8)
ax.grid(True, alpha=0.3)
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=150, bbox_inches="tight")
print(f"Plot saved to {save_path}")
return ax
def plot_precipitation(
self,
df: Optional[pd.DataFrame] = None,
stations: Optional[List[str]] = None,
cumulative: bool = True,
figsize: tuple = (12, 5),
save_path: Optional[Path] = None,
) -> plt.Axes:
"""
Plot precipitation, one line per station. Cumulative by default,
which makes it easy to compare total rainfall received over the
record and spot where stations diverge.
Args:
df: Optional pre-loaded DataFrame (raw or already converted). If not
provided, data is loaded and unit-converted from disk.
stations: Optional list of station codes to plot. Defaults to all.
cumulative: If True, plot the running cumulative total (mm).
If False, plot the raw daily/timestep values.
figsize: Figure size in inches.
save_path: If provided, saves the figure to this path.
Returns:
The matplotlib Axes object.
"""
if df is None:
df = self.convert_units(self.load())
elif "precip_mm" not in df.columns and "precip_m" in df.columns:
df = self.convert_units(df)
plot_df = df.dropna(subset=["precip_mm"])
if stations:
plot_df = plot_df[plot_df["station_code"].isin(stations)]
if plot_df.empty:
raise ValueError("No precipitation data available to plot")
fig, ax = plt.subplots(figsize=figsize)
for station_code, group in plot_df.groupby("station_code"):
group = group.sort_values("date")
values = group["precip_mm"].cumsum() if cumulative else group["precip_mm"]
ax.plot(group["date"], values, linewidth=1, label=station_code)
ax.set_title("Cumulative Precipitation" if cumulative else "Precipitation")
ax.set_xlabel("Date")
ax.set_ylabel("Cumulative precipitation (mm)" if cumulative else "Precipitation (mm)")
ax.xaxis.set_major_locator(mdates.AutoDateLocator())
ax.xaxis.set_major_formatter(mdates.ConciseDateFormatter(ax.xaxis.get_major_locator()))
ax.legend(title="Station", fontsize=8)
ax.grid(True, alpha=0.3)
plt.tight_layout()
if save_path:
plt.savefig(save_path, dpi=150, bbox_inches="tight")
print(f"Plot saved to {save_path}")
return ax |