Download update_data.py from idiouf/AfriENSO: direct link, hf CLI and curl.
- Browser
- Download file 53.5 kB
-
https://huggingface.co/spaces/idiouf/AfriENSO/resolve/main/update_data.py
- Command line
-
hf download hf://spaces/idiouf/AfriENSO/update_data.py
-
curl -L -o update_data.py https://huggingface.co/spaces/idiouf/AfriENSO/resolve/main/update_data.py
53.5 kB
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
AfriENSO - data pipeline
========================
Downloads the public datasets used by the AfriENSO dashboard, computes every
statistic shown on the page and writes one file, ``afrienso_data.js``, that the
static page (index.html) reads. Run it on your own computer, then upload
``afrienso_data.js`` to the Hugging Face Space (or run ``publish_hf.py``).
python update_data.py # normal run (uses a 12 h download cache)
python update_data.py --no-c3s # skip the Copernicus C3S forecasts
python update_data.py --refresh # ignore the cache, download everything
Data sources (all public):
* NOAA CPC - ONI (ERSSTv5), RONI, weekly OISST v2.1 Nino indices, monthly
ERSSTv5 Nino indices
* NOAA PSL - ERSSTv5 gridded SST, GPCP v2.3 monthly precipitation,
GHCN_CAMS monthly land air temperature (THREDDS/OPeNDAP)
* Copernicus Climate Pulse - ERA5 daily global mean 2 m temperature
* Copernicus C3S (optional, needs a free CDS account + ~/.cdsapirc) -
multi-system seasonal forecasts of SST and precipitation
October 2026 (v1.1, co-authors' review of the associated paper): adds the West
African land boxes GC, WSHL, CSHL and SHL (units of kind "wa", geo["wa"]) and the
SST indices SENC (Senegal-Mauritania coast), MED (Mediterranean) and AMG (TNA-ATL3).
Author: Dr Ibrahima Diouf - University of Labe / LPAO-SF, ESP, UCAD
"""
import argparse
import datetime as dt
import json
import math
import os
import re
import sys
import time
import traceback
import warnings
import numpy as np
import pandas as pd
import requests
HERE = os.path.dirname(os.path.abspath(__file__))
CACHE = os.path.join(HERE, "_cache")
OUT_JS = os.path.join(HERE, "afrienso_data.js")
GEO_JSON = os.path.join(HERE, "africa_geo.json")
URL = {
"oni": "https://www.cpc.ncep.noaa.gov/data/indices/oni.ascii.txt",
"roni": "https://www.cpc.ncep.noaa.gov/data/indices/RONI.ascii.txt",
"weekly": "https://www.cpc.ncep.noaa.gov/data/indices/wksst9120.for",
"ersst_nino": "https://www.cpc.ncep.noaa.gov/data/indices/ersst5.nino.mth.91-20.ascii",
"era5_daily": "https://sites.ecmwf.int/data/climatepulse/data/series/era5_daily_series_2t_global.csv",
"ersst_dap": "https://psl.noaa.gov/thredds/dodsC/Datasets/noaa.ersst.v5/sst.mnmean.nc",
"gpcp_dap": "https://psl.noaa.gov/thredds/dodsC/Datasets/gpcp/precip.mon.mean.nc",
"ghcn_dap": "https://psl.noaa.gov/thredds/dodsC/Datasets/ghcncams/air.mon.mean.nc",
"ersst_file": "https://downloads.psl.noaa.gov/Datasets/noaa.ersst.v5/sst.mnmean.nc",
"gpcp_file": "https://downloads.psl.noaa.gov/Datasets/gpcp/precip.mon.mean.nc",
"ghcn_file": "https://downloads.psl.noaa.gov/Datasets/ghcncams/air.mon.mean.nc",
}
GPCC_DAP = { # GPCC Full Data (to 2020) + Monitoring (recent months), 1 deg, mm/month
"full": ["https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/full_v2022/precip.mon.total.1x1.v2022.nc",
"https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/full_v2020/precip.mon.total.1x1.v2020.nc"],
"monitor": ["https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/monitor/precip.monitor.mon.total.1x1.v2022.nc",
"https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/monitor/precip.monitor.mon.total.1x1.v2020.nc"],
}
DRY_MM = 30.0 # seasonal (3-month) rainfall below which a season is "dry": not assessed
STRONG_THR = 1.5 # strong El Nino: NDJ ONI >= 1.5 C
BASE0, BASE1 = 1991, 2020 # WMO climatological reference period
PI_OFFSET_DEFAULT = 0.88 # C3S: 1991-2020 is 0.88 C above 1850-1900
AFRICA_BOX = (-37.0, 39.0, -20.0, 53.0) # lat0, lat1, lon0, lon1
FDR_Q = 0.10 # Wilks (2016): alpha_FDR = 2 x alpha_global
SEAS = ["DJF", "JFM", "FMA", "MAM", "AMJ", "MJJ", "JJA", "JAS", "ASO", "SON", "OND", "NDJ"]
# C3S systems tried in order (the most recent first). Edit if C3S upgrades.
C3S_SYSTEMS = {
"ecmwf": ["51"], "meteo_france": ["9", "8"], "ukmo": ["606", "605", "604", "603", "602"],
"dwd": ["22", "21"], "cmcc": ["35", "4"], "ncep": ["2"], "jma": ["4", "3", "2"],
"eccc": ["5", "4"],
}
STATUS = []
warnings.filterwarnings("ignore", category=RuntimeWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)
ARGS = None
def log(msg):
print(time.strftime("%H:%M:%S"), msg, flush=True)
def status(key, ok, msg=""):
STATUS.append({"key": key, "ok": bool(ok), "msg": str(msg)[:300],
"time": dt.datetime.utcnow().strftime("%Y-%m-%d %H:%M UTC")})
log(("OK " if ok else "FAIL ") + key + (" - " + str(msg)[:200] if msg else ""))
# --------------------------------------------------------------------------
# Download helpers
# --------------------------------------------------------------------------
def http_get(url, name, max_age_h=12, binary=False, timeout=120):
os.makedirs(CACHE, exist_ok=True)
path = os.path.join(CACHE, name)
fresh = os.path.exists(path) and (time.time() - os.path.getmtime(path)) < max_age_h * 3600
if os.path.exists(path) and (fresh or ARGS.offline) and not ARGS.refresh:
return path
if ARGS.offline:
raise RuntimeError("offline mode and no cached file " + name)
last = None
for attempt in range(3):
try:
r = requests.get(url, timeout=timeout, stream=True,
headers={"User-Agent": "AfriENSO/1.0 (research dashboard)"})
r.raise_for_status()
tmp = path + ".part"
with open(tmp, "wb") as f:
for chunk in r.iter_content(1 << 20):
f.write(chunk)
os.replace(tmp, path)
return path
except Exception as e: # noqa
last = e
time.sleep(3 * (attempt + 1))
if os.path.exists(path):
log(" download failed, using older cached copy of " + name)
return path
raise RuntimeError("download failed: %s (%s)" % (url, last))
def read_text(path):
with open(path, "r", encoding="utf-8", errors="replace") as f:
return f.read()
def r2(x, nd=2):
if x is None:
return None
try:
xf = float(x)
except Exception:
return None
if not np.isfinite(xf):
return None
return round(xf, nd)
def series_out(s, nd=2):
"""Monthly pandas Series (PeriodIndex or DatetimeIndex) -> {start, v}."""
s = s.copy()
if not isinstance(s.index, pd.PeriodIndex):
s.index = pd.PeriodIndex(s.index, freq="M")
full = pd.period_range(s.index.min(), s.index.max(), freq="M")
s = s.reindex(full)
return {"start": str(full[0]), "v": [r2(v, nd) for v in s.values]}
# --------------------------------------------------------------------------
# 1. ENSO indices (NOAA CPC)
# --------------------------------------------------------------------------
def parse_oni(text):
rows = []
for line in text.splitlines():
m = re.match(r"^\s*([A-Z]{3})\s+(\d{4})\s+(.*)$", line)
if not m or m.group(1) not in SEAS:
continue
nums = re.findall(r"-?\d+\.\d+", m.group(3))
if not nums:
continue
val = float(nums[-1])
if val < -90:
continue
rows.append((int(m.group(2)), SEAS.index(m.group(1)) + 1, val))
idx = pd.PeriodIndex([pd.Period(year=y, month=mo, freq="M") for y, mo, _ in rows], freq="M")
return pd.Series([v for _, _, v in rows], index=idx).sort_index()
def parse_weekly(text):
recs = []
for line in text.splitlines():
m = re.match(r"^\s*(\d{2}[A-Z]{3}\d{4})(.*)$", line)
if not m:
continue
nums = re.findall(r"-?\d+\.\d", m.group(2))
if len(nums) < 8:
continue
d = dt.datetime.strptime(m.group(1).title(), "%d%b%Y").date()
v = [float(x) for x in nums[:8]]
recs.append((d, v[1], v[3], v[5], v[7], v[5 - 1]))
df = pd.DataFrame(recs, columns=["date", "n12", "n3", "n34", "n4", "sst34"])
return df
def parse_ersst_nino(text):
recs = []
for line in text.splitlines():
p = line.split()
if len(p) < 10 or not p[0].isdigit():
continue
try:
y, mo = int(p[0]), int(p[1])
v = [float(x) for x in p[2:10]]
except ValueError:
continue
recs.append((pd.Period(year=y, month=mo, freq="M"), v[1], v[3], v[5], v[7]))
df = pd.DataFrame(recs, columns=["per", "n12", "n3", "n4", "n34"]).set_index("per")
df.index = pd.PeriodIndex(df.index, freq="M")
return df
def enso_episodes(oni, thr=0.5, nmin=5):
"""CPC definition: >= 5 consecutive overlapping seasons above +0.5 / below -0.5."""
eps = []
vals = oni.values
idx = oni.index
for sign, kind in ((1, "EN"), (-1, "LN")):
run = []
for i, v in enumerate(vals):
if np.isfinite(v) and sign * v >= thr:
run.append(i)
else:
if len(run) >= nmin:
eps.append((kind, run[:]))
run = []
if len(run) >= nmin:
eps.append((kind, run[:], ))
out = []
for kind, run in eps:
seg = vals[run]
k = int(np.argmax(np.abs(seg)))
peak = float(seg[k])
a = abs(peak)
strength = "weak" if a < 1.0 else "moderate" if a < 1.5 else "strong" if a < 2.0 else "very strong"
ongoing = run[-1] == len(vals) - 1
out.append({"type": kind, "start": str(idx[run[0]]), "end": str(idx[run[-1]]),
"peak": round(peak, 2), "peak_at": str(idx[run[k]]), "months": len(run),
"strength": strength, "ongoing": bool(ongoing)})
out.sort(key=lambda e: e["start"])
return out
def phase_series(oni, episodes):
"""Monthly (season-centre) phase: 1 El Nino, -1 La Nina, 0 neutral."""
ph = pd.Series(0, index=oni.index)
for e in episodes:
sl = (ph.index >= pd.Period(e["start"], "M")) & (ph.index <= pd.Period(e["end"], "M"))
ph[sl] = 1 if e["type"] == "EN" else -1
return ph
def block_enso():
out = {}
oni = parse_oni(read_text(http_get(URL["oni"], "oni.ascii.txt")))
if len(oni) < 100:
raise RuntimeError("ONI parse failed")
eps = enso_episodes(oni)
out["oni"] = series_out(oni)
out["episodes"] = eps
out["phase"] = {"start": str(oni.index[0]), "v": phase_series(oni, eps).astype(int).tolist()}
status("CPC ONI", True, "%d seasons, last %s = %+.2f" % (len(oni), oni.index[-1], oni.iloc[-1]))
try:
roni = parse_oni(read_text(http_get(URL["roni"], "RONI.ascii.txt")))
out["roni"] = series_out(roni)
status("CPC RONI", True, "last %s = %+.2f" % (roni.index[-1], roni.iloc[-1]))
except Exception as e:
status("CPC RONI", False, e)
try:
wk = parse_weekly(read_text(http_get(URL["weekly"], "wksst9120.for")))
out["weekly"] = {"dates": [d.isoformat() for d in wk["date"]],
"n12": wk["n12"].round(2).tolist(), "n3": wk["n3"].round(2).tolist(),
"n34": wk["n34"].round(2).tolist(), "n4": wk["n4"].round(2).tolist()}
status("CPC weekly OISST Nino", True, "last week %s, Nino3.4 %+.1f" % (wk["date"].iloc[-1], wk["n34"].iloc[-1]))
except Exception as e:
status("CPC weekly OISST Nino", False, e)
mon = None
try:
mon = parse_ersst_nino(read_text(http_get(URL["ersst_nino"], "ersst5.nino.mth.91-20.ascii")))
out["monthly"] = {k: series_out(mon[k]) for k in ["n12", "n3", "n34", "n4"]}
status("CPC monthly ERSSTv5 Nino", True, "last %s" % mon.index[-1])
except Exception as e:
status("CPC monthly ERSSTv5 Nino", False, e)
return out, oni, eps, mon
# --------------------------------------------------------------------------
# 2. Gridded data (NOAA PSL, OPeNDAP with file fallback)
# --------------------------------------------------------------------------
def _runs(idx):
idx = np.asarray(idx)
if idx.size == 0:
return []
cuts = np.where(np.diff(idx) != 1)[0] + 1
return [(int(r[0]), int(r[-1]) + 1) for r in np.split(idx, cuts)]
def subset(da, lat0, lat1, lon0, lon1, t0=None):
"""Box subset that works across the Greenwich meridian with 0-360 grids."""
import xarray as xr
latn = [n for n in da.dims if n.lower().startswith("lat")][0]
lonn = [n for n in da.dims if n.lower().startswith("lon")][0]
if t0 is not None:
da = da.sel(time=slice(t0, None))
lat = da[latn].values
lon = da[lonn].values
lon180 = ((lon + 180.0) % 360.0) - 180.0
ilat = np.where((lat >= lat0) & (lat <= lat1))[0]
ilon = np.where((lon180 >= lon0) & (lon180 <= lon1))[0]
la = _runs(ilat)
parts = []
nt = da.sizes.get("time", 1)
for a, b in _runs(ilon):
sub = da.isel({latn: slice(la[0][0], la[-1][1]), lonn: slice(a, b)})
if "time" in da.dims and nt > 120: # chunked requests: friendlier to OPeNDAP servers
chunks = []
for t0_ in range(0, nt, 120):
print(" months %d-%d of %d" % (t0_ + 1, min(nt, t0_ + 120), nt), end="\r", flush=True)
for attempt in range(3):
try:
chunks.append(sub.isel(time=slice(t0_, t0_ + 120)).load())
break
except Exception:
if attempt == 2:
raise
time.sleep(5)
print(" " * 70, end="\r", flush=True)
blk = xr.concat(chunks, dim="time")
else:
blk = sub.load()
parts.append(blk)
out = xr.concat(parts, dim=lonn) if len(parts) > 1 else parts[0]
out = out.assign_coords({lonn: (((out[lonn] + 180.0) % 360.0) - 180.0)})
out = out.sortby(lonn).sortby(latn)
return out.rename({latn: "lat", lonn: "lon"})
def open_ds(src):
"""Open a NetCDF/OPeNDAP dataset and decode time by hand: some PSL files use
units such as 'days since 1800-01-01 00:00:00 0:00' that pandas cannot parse."""
import xarray as xr
ds = xr.open_dataset(src, decode_times=False)
if "time" in ds.variables:
t = ds["time"]
units = str(t.attrs.get("units", ""))
m = re.match(r"\s*(\w+)\s+since\s+(\d{1,4})-(\d{1,2})-(\d{1,2})", units)
if not m:
raise RuntimeError("unknown time units: " + units)
step = {"days": "D", "day": "D", "hours": "h", "hour": "h", "minutes": "min", "seconds": "s"}[m.group(1).lower()]
base = pd.Timestamp(int(m.group(2)), int(m.group(3)), int(m.group(4)))
times = base + pd.to_timedelta(np.asarray(t.values, dtype=float), unit=step)
ds = ds.assign_coords(time=("time", pd.DatetimeIndex(times).values))
return ds
def open_psl(kind, var):
"""Try OPeNDAP first, then a full-file download (cached)."""
import xarray as xr
errs = []
if not ARGS.offline:
for attempt in range(2):
try:
log(" %s: connecting to NOAA PSL (OPeNDAP)..." % kind)
ds = open_ds(URL[kind + "_dap"])
return ds[var], "OPeNDAP"
except Exception as e:
errs.append(str(e))
time.sleep(5)
try:
log(" %s: OPeNDAP unavailable, downloading the full file (can take 10-30 min)..." % kind)
p = http_get(URL[kind + "_file"], kind + "_" + os.path.basename(URL[kind + "_file"]),
max_age_h=24 * 7, binary=True, timeout=1800)
ds = open_ds(p)
return ds[var], "file"
except Exception as e:
errs.append(str(e))
raise RuntimeError(" | ".join(errs))
def monthly_anom(s):
"""Series with PeriodIndex -> anomaly vs 1991-2020 monthly climatology."""
base = s[(s.index.year >= BASE0) & (s.index.year <= BASE1)]
clim = base.groupby(base.index.month).mean()
return s - clim.reindex(s.index.month).values
def valid(x, lo, hi):
"""Mask fill values and non-physical numbers (e.g. 9.97e36 missing-value codes)."""
a = np.asarray(x, dtype=float)
return np.where(np.isfinite(a) & (a >= lo) & (a <= hi), a, np.nan)
def box_mean(da):
w = np.cos(np.deg2rad(da["lat"]))
return da.weighted(w.fillna(0)).mean(("lat", "lon"), skipna=True)
def to_period_series(da):
t = pd.to_datetime(da["time"].values)
return pd.Series(np.asarray(da.values, dtype=float), index=pd.PeriodIndex(t, freq="M"))
def block_sst_indices():
sst, how = open_psl("ersst", "sst")
boxes = {
"wtio": (-10, 10, 50, 70), "setio": (-10, 0, 90, 110),
"tna": (5.5, 23.5, -57.5, -15), "atl3": (-3, 3, -20, 0), "n34e": (-5, 5, -170, -120),
# added October 2026 (co-authors' review): Senegal-Mauritania coast and Mediterranean
"senc": (10, 20, -24, -16), "med": (32, 44, 0, 30),
}
ser = {}
for k, (a, b, c, d) in boxes.items():
log(" ERSSTv5: box %s" % k)
bx = subset(sst, a, b, c, d, t0="1948-01-01")
bx = bx.copy(data=valid(bx.values, -5, 40))
ser[k] = to_period_series(box_mean(bx))
dmi = monthly_anom(ser["wtio"]) - monthly_anom(ser["setio"])
idx = {"dmi": dmi, "tna": monthly_anom(ser["tna"]), "atl3": monthly_anom(ser["atl3"]),
"n34e": monthly_anom(ser["n34e"]),
"senc": monthly_anom(ser["senc"]), "med": monthly_anom(ser["med"])}
idx["amg"] = idx["tna"] - idx["atl3"] # Atlantic meridional gradient (TNA - ATL3)
# latest SST anomaly map, tropics 30S-30N (2 deg), last 3 months
log(" ERSSTv5: tropical SST map")
trop = subset(sst, -30, 30, -180, 180, t0="1991-01-01")
tt = pd.PeriodIndex(pd.to_datetime(trop["time"].values), freq="M")
arr = valid(trop.values, -5, 40)
base = (tt.year >= BASE0) & (tt.year <= BASE1)
maps = []
for k in range(3):
i = len(tt) - 1 - k
clim = np.nanmean(arr[base & (tt.month == tt[i].month)], axis=0)
an = arr[i] - clim
maps.append({"month": str(tt[i]), "z": [[r2(v, 2) for v in row] for row in an]})
sstmap = {"lat": [float(x) for x in trop["lat"].values], "lon": [float(x) for x in trop["lon"].values],
"maps": maps[::-1]}
status("PSL ERSSTv5 gridded (%s)" % how, True, "DMI/TNA/ATL3 to %s" % dmi.dropna().index[-1])
return idx, sstmap
def land_fraction(lat, lon, polys, dlat, dlon, nsub=6):
"""Fraction of each grid cell inside a set of polygons (list of [ext, holes...])."""
from matplotlib.path import Path
LA, LO = np.meshgrid(lat, lon, indexing="ij")
off = (np.arange(nsub) + 0.5) / nsub - 0.5
oy, ox = np.meshgrid(off * dlat, off * dlon, indexing="ij")
pts = np.stack([(LO[..., None, None] + ox).ravel(), (LA[..., None, None] + oy).ravel()], axis=1)
inside = np.zeros(len(pts), dtype=bool)
for poly in polys:
ext = np.asarray(poly[0])
bb = (pts[:, 0] >= ext[:, 0].min()) & (pts[:, 0] <= ext[:, 0].max()) & \
(pts[:, 1] >= ext[:, 1].min()) & (pts[:, 1] <= ext[:, 1].max())
if not bb.any():
continue
cand = np.where(bb)[0]
ins = Path(ext).contains_points(pts[cand])
for hole in poly[1:]:
ins &= ~Path(np.asarray(hole)).contains_points(pts[cand])
inside[cand[ins]] = True
return inside.reshape(len(lat), len(lon), nsub * nsub).mean(axis=2)
def unit_weights(lat, lon, polys, dlat, dlon):
f = land_fraction(lat, lon, polys, dlat, dlon)
w = f * np.cos(np.deg2rad(lat))[:, None]
if w.sum() <= 0: # very small island: nearest cell to the polygon centroid
ext = np.asarray(polys[0][0])
cy, cx = ext[:, 1].mean(), ext[:, 0].mean()
i = int(np.argmin(np.abs(lat - cy)))
j = int(np.argmin(np.abs(lon - cx)))
w = np.zeros_like(w)
w[i, j] = 1.0
return w, True
return w, False
def weighted_series(arr, w):
"""arr (T, ny, nx) with NaN; w (ny, nx) -> (T,) weighted mean over valid cells."""
m = np.isfinite(arr) & (w[None] > 0)
num = np.where(m, arr, 0.0) * w[None]
den = (m * w[None]).sum(axis=(1, 2))
with np.errstate(invalid="ignore", divide="ignore"):
out = num.sum(axis=(1, 2)) / den
# require at least 50 % of the unit weight to be valid
cover = den / w.sum()
out[cover < 0.5] = np.nan
return out
def seasonal_stack(arr, per, how):
"""Monthly (T,...) -> dict centre_month -> (years, (Y,...)) of 3-month sums/means."""
s = {}
ym = per.year * 12 + per.month - 1
pos = {v: i for i, v in enumerate(ym)}
for c in range(1, 13):
years, vals = [], []
for y in range(per.year.min(), per.year.max() + 1):
k = y * 12 + c - 1
ii = [pos.get(k - 1), pos.get(k), pos.get(k + 1)]
if any(i is None for i in ii):
continue
blk = arr[ii]
v = blk.sum(axis=0) if how == "sum" else blk.mean(axis=0)
if np.ndim(v) == 0 and not np.isfinite(v):
continue
years.append(y)
vals.append(v)
if years:
s[c] = (np.array(years), np.array(vals))
return s
def detrend(y, x):
ok = np.isfinite(y)
if ok.sum() < 3:
return y * np.nan
p = np.polyfit(x[ok], y[ok], 1)
return y - np.polyval(p, x)
def detrend_field(Y, x):
"""Y (n, ny, nx) -> residuals from a per-cell linear trend (NaN-safe)."""
n = Y.shape[0]
X = np.vstack([np.ones(n), x - x.mean()]).T
flat = Y.reshape(n, -1)
out = np.full_like(flat, np.nan)
ok_cols = np.isfinite(flat).all(axis=0)
if ok_cols.any():
beta, *_ = np.linalg.lstsq(X, flat[:, ok_cols], rcond=None)
out[:, ok_cols] = flat[:, ok_cols] - X @ beta
return out.reshape(Y.shape)
def corr_field(x, Y):
"""x (n,), Y (n, ny, nx) both detrended -> r, p."""
from scipy import stats
n = len(x)
xa = x - x.mean()
Ya = Y - Y.mean(axis=0)
with np.errstate(invalid="ignore", divide="ignore"):
r = (xa[:, None, None] * Ya).sum(axis=0) / np.sqrt((xa ** 2).sum() * (Ya ** 2).sum(axis=0))
t = r * np.sqrt((n - 2) / (1 - r ** 2))
p = 2 * stats.t.sf(np.abs(t), n - 2)
return r, p
def fdr_mask(p, q=FDR_Q):
"""Benjamini-Hochberg false discovery rate (Wilks 2016)."""
flat = p[np.isfinite(p)]
m = np.zeros(p.shape, dtype=bool)
if flat.size == 0:
return m
ps = np.sort(flat)
k = np.arange(1, ps.size + 1)
below = ps <= q * k / ps.size
if not below.any():
return m
pthr = ps[np.where(below)[0].max()]
return np.isfinite(p) & (p <= pthr)
def enc_map(z, scale, landmask):
out = []
for row, lrow in zip(z, landmask):
out.append([None if (not l or not np.isfinite(v)) else int(round(v * scale)) for v, l in zip(row, lrow)])
return out
def enc_mask(m, landmask):
return "".join("1" if (a and l) else "0" for a, l in zip(m.ravel(), landmask.ravel()))
def driver_seasonal(driver_monthly, c):
"""3-month mean centred on month c for each year -> dict year->value."""
s = driver_monthly
out = {}
for per in s.index:
if per.month != c:
continue
a, b = per - 1, per + 1
if a in s.index and b in s.index:
v = [s[a], s[per], s[b]]
if all(np.isfinite(v)):
out[per.year] = float(np.mean(v))
return out
def strong_events(oni, thr=STRONG_THR):
"""Strong El Nino events: NDJ ONI >= thr (labelled by the developing year)."""
ev = []
for p, v in oni.items():
if p.month == 12 and np.isfinite(v) and v >= thr:
ev.append({"year": int(p.year), "label": "%d-%s" % (p.year, str(p.year + 1)[2:]), "ndj": round(float(v), 2)})
return ev
def load_gpcc(lat0, lat1, lon0, lon1):
import xarray as xr
parts = {}
for kind in ("full", "monitor"):
srcs = list(GPCC_DAP[kind])
local = os.path.join(CACHE, "gpcc_%s.nc" % kind)
if os.path.exists(local):
srcs.insert(0, local)
if ARGS.offline:
srcs = [x for x in srcs if os.path.exists(x)]
for src in srcs:
try:
log(" GPCC %s: %s" % (kind, os.path.basename(src)))
da = subset(open_ds(src)["precip"], lat0, lat1, lon0, lon1, t0="1951-01-01")
da = da.copy(data=valid(da.values, 0, 3000))
parts[kind] = da
break
except Exception as e: # noqa
log(" not available (%s)" % str(e)[:100])
if "full" not in parts:
raise RuntimeError("GPCC Full Data not reachable")
da = parts["full"]
if "monitor" in parts:
mon = parts["monitor"]
mon = mon.sel(time=mon["time"] > da["time"].values[-1])
if mon.sizes["time"]:
da = xr.concat([da, mon.interp(lat=da["lat"], lon=da["lon"], method="nearest")], dim="time")
return da.coarsen(lat=2, lon=2, boundary="trim").mean(skipna=True)
def strong_maps(arr, per, land, events, phase_m, is_pr):
"""Per season: mean anomaly of strong El Nino seasons relative to ENSO-neutral seasons
(detrended), as % of the 1991-2020 normal (rainfall) or C (temperature), and the number
of events below the neutral baseline."""
stack = seasonal_stack(arr, per, "sum" if is_pr else "mean")
out = []
for c in range(1, 13):
if c not in stack:
out.append(None)
continue
years, Y = stack[c]
resid = detrend_field(Y, years.astype(float))
phs = {p.year: int(v) for p, v in phase_m.items() if p.month == c}
neu = np.array([phs.get(int(y)) == 0 for y in years])
if neu.sum() < 5:
out.append(None)
continue
base = np.nanmean(resid[neu], axis=0)
bsel = (years >= BASE0) & (years <= BASE1)
clim = np.nanmean(Y[bsel] if bsel.sum() >= 10 else Y, axis=0)
an, used = [], []
for e in events:
ty = e["year"] if c >= 6 else e["year"] + 1
k = np.where(years == ty)[0]
if len(k):
an.append(resid[k[0]] - base)
used.append(e["label"])
if len(an) < 3:
out.append(None)
continue
A = np.array(an)
if is_pr:
with np.errstate(invalid="ignore", divide="ignore"):
P = np.where(clim[None] >= DRY_MM, 100.0 * (clim[None] + A) / clim[None], np.nan)
mean, scale = np.nanmean(P, axis=0), 1
else:
mean, scale = np.nanmean(A, axis=0), 100
if is_pr:
A = np.where(clim[None] >= DRY_MM, A, np.nan)
nneg = np.sum(A < 0, axis=0)
nval = np.sum(np.isfinite(A), axis=0)
code = "".join(str(int(min(9, a))) if (l and v > 0) else "." for a, v, l in zip(nneg.ravel(), nval.ravel(), land.ravel()))
out.append({"events": used, "mean": enc_map(mean, scale, land), "nneg": code})
return out
def _box_ring(lat0, lat1, lon0, lon1):
return [[lon0, lat0], [lon1, lat0], [lon1, lat1], [lon0, lat1], [lon0, lat0]]
# West African land boxes (restricted to African land like the AR6 regions); SHL = WSHL + CSHL
WA_BOXES = [
{"code": "GC", "en": "Guinea Coast", "fr": "Côte guinéenne", "ring": _box_ring(4, 8, -15, 10)},
{"code": "WSHL", "en": "Western Sahel", "fr": "Sahel occidental", "ring": _box_ring(11, 18, -18, -5)},
{"code": "CSHL", "en": "Central Sahel", "fr": "Sahel central", "ring": _box_ring(11, 18, -5, 15)},
{"code": "SHL", "en": "Sahel", "fr": "Sahel", "ring": _box_ring(11, 18, -18, 15)},
]
def block_africa(geo, oni, phase_m, sst_idx):
"""Teleconnection maps + unit (country/AR6/Africa) monthly series."""
import xarray as xr # noqa
lat0, lat1, lon0, lon1 = AFRICA_BOX
res = {"units": {}, "maps": {}}
all_polys = [p for c in geo["countries"] for p in c["mask"]]
units = [("AFR", "Africa (land)", "Afrique (terres)", all_polys, "africa")]
units += [(r["code"], r["en"], r["fr"], [[r["ring"]]], "ar6") for r in geo["ar6"]]
units += [(c["iso"], c["en"], c["fr"], c["mask"], "country") for c in geo["countries"]]
# West African land boxes of the associated paper (Guinea Coast, western and central Sahel, Sahel)
units += [(b["code"], b["en"], b["fr"], [[b["ring"]]], "wa") for b in WA_BOXES]
drivers = {"nino34": oni}
for k in ("dmi", "tna", "atl3"):
if sst_idx is not None and k in sst_idx:
drivers[k] = sst_idx[k].dropna()
grids = {}
# ---- precipitation: GPCP v2.3 (2.5 deg, mm/day) -> mm/month
try:
pr, how = open_psl("gpcp", "precip")
pr = subset(pr, lat0, lat1, lon0, lon1, t0="1979-01-01")
per = pd.PeriodIndex(pd.to_datetime(pr["time"].values), freq="M")
arr = valid(pr.values, 0, 200) * per.days_in_month.values[:, None, None] # mm/day -> mm/month
grids["pr"] = (pr["lat"].values, pr["lon"].values, arr, per, 2.5)
status("PSL GPCP v2.3 (%s)" % how, True, "%s to %s" % (per[0], per[-1]))
except Exception as e:
status("PSL GPCP v2.3", False, e)
# ---- temperature: GHCN_CAMS (0.5 deg) -> 1.5 deg blocks
try:
ta, how = open_psl("ghcn", "air")
ta = subset(ta, lat0, lat1, lon0, lon1, t0="1948-01-01")
ta = ta.copy(data=valid(ta.values, 150, 350) if float(np.nanmedian(valid(ta.values, -90, 400))) > 150 else valid(ta.values, -80, 70))
if float(ta.mean(skipna=True)) > 150:
ta = ta - 273.15
ta = ta.coarsen(lat=3, lon=3, boundary="trim").mean(skipna=True)
per = pd.PeriodIndex(pd.to_datetime(ta["time"].values), freq="M")
grids["tas"] = (ta["lat"].values, ta["lon"].values, np.asarray(ta.values, dtype=float), per, 1.5)
status("PSL GHCN_CAMS (%s)" % how, True, "%s to %s" % (per[0], per[-1]))
except Exception as e:
status("PSL GHCN_CAMS", False, e)
# ---- long rainfall record for strong El Nino analogs: GPCC (1 deg -> 2 deg), from 1951
try:
gp = load_gpcc(lat0, lat1, lon0, lon1)
per = pd.PeriodIndex(pd.to_datetime(gp["time"].values), freq="M")
arr = valid(gp.values, 0, 3000) # mm/month
grids["prl"] = (gp["lat"].values, gp["lon"].values, arr, per, 2.0)
status("PSL GPCC (Full + Monitoring)", True, "%s to %s" % (per[0], per[-1]))
except Exception as e:
status("PSL GPCC (Full + Monitoring)", False, str(e) + " - strong El Nino rainfall analogs use GPCP (from 1979)")
events = strong_events(oni)
res["strong_events"] = events
for var, (lat, lon, arr, per, dres) in grids.items():
log(" %s: land fractions and unit weights" % var)
land = land_fraction(lat, lon, all_polys, dres, dres) >= 0.25
raw = arr
arr = np.where(land[None], arr, np.nan)
# unit series (fraction-weighted on the unmasked grid)
landfrac = land_fraction(lat, lon, all_polys, dres, dres)
for code, en, fr, polys, kind in units:
w, nearest = unit_weights(lat, lon, polys, dres, dres)
if kind in ("ar6", "wa") and not nearest:
w = w * landfrac # AR6 land regions restricted to African land
s = weighted_series(raw, w)
u = res["units"].setdefault(code, {"en": en, "fr": fr, "kind": kind})
nd = 1 if var.startswith("pr") else 2
u[var] = {"start": str(per[0]), "v": [r2(v, nd) for v in s]}
if nearest:
u[var + "_nearest"] = True
mp = {"lat": [float(x) for x in lat], "lon": [float(x) for x in lon], "res": dres,
"land": "".join("1" if v else "0" for v in land.ravel()), "corr": {}, "comp": {}}
do_strong = var in ("tas", "prl") or (var == "pr" and "prl" not in grids)
if do_strong and events:
log(" %s: strong El Nino analog maps" % var)
mp["strong"] = strong_maps(arr, per, land, events, phase_m, var.startswith("pr"))
if var == "prl":
res["maps"][var] = mp
continue
# maps
log(" %s: correlation and composite maps" % var)
how = "sum" if var == "pr" else "mean"
stack = seasonal_stack(arr, per, how)
for dname, dser in drivers.items():
mp["corr"][dname] = []
for c in range(1, 13):
if c not in stack:
mp["corr"][dname].append(None)
continue
years, Y = stack[c]
if dname == "nino34":
dv = {p.year: float(v) for p, v in dser.items() if p.month == c}
else:
dv = driver_seasonal(dser, c)
keep = np.array([y in dv for y in years])
yy = years[keep]
if len(yy) < 15:
mp["corr"][dname].append(None)
continue
x = detrend(np.array([dv[y] for y in yy]), yy.astype(float))
Yd = detrend_field(Y[keep], yy.astype(float))
r, p = corr_field(x, Yd)
if var == "pr": # dry season: correlation of a few mm is not meaningful
wet = np.nanmean(Y[keep], axis=0) >= DRY_MM
r = np.where(wet, r, np.nan)
p = np.where(wet, p, np.nan)
r = np.where(land, r, np.nan)
p = np.where(land, p, np.nan)
mp["corr"][dname].append({"n": int(len(yy)), "y0": int(yy.min()), "y1": int(yy.max()),
"r": enc_map(r, 100, land), "sig": enc_mask(fdr_mask(p), land)})
# ENSO composites (El Nino / La Nina seasons, CPC episodes), detrended
from scipy import stats
for ph, lab in ((1, "EN"), (-1, "LN")):
mp["comp"][lab] = []
for c in range(1, 13):
if c not in stack:
mp["comp"][lab].append(None)
continue
years, Y = stack[c]
phs = {p.year: int(v) for p, v in phase_m.items() if p.month == c}
keep = np.array([y in phs for y in years])
yy, Yk = years[keep], Y[keep]
sel = np.array([phs[y] == ph for y in yy])
if sel.sum() < 3:
mp["comp"][lab].append(None)
continue
Yd = detrend_field(Yk, yy.astype(float))
clim = np.nanmean(Yk, axis=0)
comp = np.nanmean(Yd[sel], axis=0)
# manual one-sample t-test (scipy nan_policy="omit" returns masked arrays)
xs = Yd[sel]
nn = np.sum(np.isfinite(xs), axis=0).astype(float)
with np.errstate(invalid="ignore", divide="ignore"):
sd = np.nanstd(xs, axis=0, ddof=1)
tt = comp / (sd / np.sqrt(nn))
p = 2.0 * stats.t.sf(np.abs(tt), np.maximum(nn - 1, 1))
p = np.where(land & (nn >= 3) & np.isfinite(tt), p, np.nan)
if var == "pr":
with np.errstate(invalid="ignore", divide="ignore"):
z = np.where(clim >= DRY_MM, 100.0 * comp / clim, np.nan) # % of normal; dry-season cells masked
scale = 1
else:
z = comp
scale = 100
mp["comp"][lab].append({"n": int(sel.sum()), "years": [int(y) for y in yy[sel]],
"z": enc_map(z, scale, land), "sig": enc_mask(fdr_mask(p), land)})
res["maps"][var] = mp
return res
# --------------------------------------------------------------------------
# 3. Global temperature: ERA5 daily (Copernicus Climate Pulse)
# --------------------------------------------------------------------------
def parse_era5_daily(path):
df = pd.read_csv(path, comment="#")
df.columns = [c.strip().lower() for c in df.columns]
dcol = [c for c in df.columns if "date" in c][0]
df[dcol] = pd.to_datetime(df[dcol])
ano = [c for c in df.columns if c.startswith("ano") and ("91" in c or "1991" in c)]
pi = [c for c in df.columns if "pi" in c and c.startswith("ano")]
tcol = [c for c in df.columns if c in ("2t", "t2m", "2m_temperature")]
s_abs = df[tcol[0]].astype(float) if tcol else None
if ano:
s = df[ano[0]].astype(float)
else: # compute the 1991-2020 day-of-year climatology ourselves
x = df.set_index(dcol)[tcol[0]].astype(float)
doy = x.index.dayofyear.where(~((x.index.month == 2) & (x.index.day == 29)), 59)
base = (x.index.year >= BASE0) & (x.index.year <= BASE1)
clim = x[base].groupby(doy[base]).mean().rolling(31, center=True, min_periods=1).mean()
s = pd.Series(x.values - clim.reindex(doy).values, index=x.index).reset_index(drop=True)
offset = PI_OFFSET_DEFAULT
if pi and ano:
d = (df[pi[0]].astype(float) - df[ano[0]].astype(float)).dropna()
if len(d) > 100:
offset = float(d.median())
out = pd.Series(s.values, index=pd.DatetimeIndex(df[dcol])).dropna()
absval = pd.Series(s_abs.values, index=pd.DatetimeIndex(df[dcol])).dropna() if s_abs is not None else None
return out, offset, absval
def ar1_phi(n34m):
x = n34m.dropna().values
return float(np.corrcoef(x[:-1], x[1:])[0, 1])
def fit_gmst(mon, n34m, lag, t_start="1995-01"):
df = pd.DataFrame({"y": mon})
df["n"] = n34m.reindex(df.index - lag).values if lag else n34m.reindex(df.index).values
df = df[df.index >= pd.Period(t_start, "M")].dropna()
t = df.index.year + (df.index.month - 0.5) / 12.0
X = np.vstack([np.ones(len(df)), t - 2000.0, df["n"].values]).T
beta, *_ = np.linalg.lstsq(X, df["y"].values, rcond=None)
res = df["y"].values - X @ beta
phi = float(np.corrcoef(res[:-1], res[1:])[0, 1])
sig = float(np.std(res[1:] - phi * res[:-1], ddof=3))
r2v = 1 - res.var() / df["y"].values.var()
return beta, res, phi, sig, r2v, df.index
def project_year(mon, n34m, year, k_obs, lag, beta, phi, sig, e_last, n34_fc=None, nsim=4000, seed=1):
"""Monte Carlo projection of the annual mean given k_obs observed months."""
rng = np.random.default_rng(seed)
per = pd.period_range("%d-01" % year, "%d-12" % year, freq="M")
days = per.days_in_month.values.astype(float)
obs = mon.reindex(per[:k_obs]).values
phi_n = ar1_phi(n34m)
last_n = n34m.dropna()
ln_per, ln_val = last_n.index[-1], float(last_n.iloc[-1])
sd_n = float(last_n.std())
sims = np.zeros(nsim)
for s in range(nsim):
e = e_last
tot = np.nansum(obs * days[:k_obs])
memb = None
if n34_fc is not None and len(n34_fc["members"]):
memb = n34_fc["members"][rng.integers(len(n34_fc["members"]))]
for i in range(k_obs, 12):
p = per[i]
src = p - lag
if src in n34m.index and np.isfinite(n34m[src]):
nv = float(n34m[src])
else:
h = (src - ln_per).n
nv = None
if memb is not None:
key = str(src)
if key in n34_fc["months"]:
nv = float(memb[n34_fc["months"].index(key)])
if nv is None:
nv = ln_val * phi_n ** h + rng.normal(0, sd_n * math.sqrt(max(1 - phi_n ** (2 * h), 0)))
e = phi * e + rng.normal(0, sig)
t = p.year + (p.month - 0.5) / 12.0
tot += (beta[0] + beta[1] * (t - 2000.0) + beta[2] * nv + e) * days[i]
sims[s] = tot / days.sum()
return sims
def block_globaltemp(n34m, n34_fc):
p = http_get(URL["era5_daily"], "era5_daily_series_2t_global.csv")
d, offset, absval = parse_era5_daily(p)
out = {"daily": {"start": d.index[0].strftime("%Y-%m-%d"), "v": [r2(v, 3) for v in d.values]},
"pi_offset": round(offset, 3)}
# monthly means (complete months only for fitting)
mm = d.groupby(pd.PeriodIndex(d.index, freq="M")).agg(["mean", "count"])
full = mm["count"] >= mm.index.days_in_month
mon = mm["mean"][full]
out["monthly"] = series_out(mm["mean"], 3)
out["monthly_complete_to"] = str(mon.index[-1])
# annual means (complete years)
yr = d.groupby(d.index.year).agg(["mean", "count"])
yfull = yr[yr["count"] >= 365]["mean"]
out["annual"] = {"years": [int(y) for y in yfull.index], "v": [r2(v, 3) for v in yfull.values]}
if n34m is None:
status("ERA5 daily global (Climate Pulse)", True, "no Nino3.4 -> no projection")
return out
# choose lag 0..6 months maximising R2
best = None
for lag in range(0, 7):
beta, res, phi, sig, r2v, idx = fit_gmst(mon, n34m, lag)
if best is None or r2v > best[4]:
best = (lag, beta, res, phi, sig, r2v, idx)
lag, beta, res, phi, sig, r2v, idx = best
adj = mon - beta[2] * n34m.reindex(mon.index - lag).values
out["enso_fit"] = {"lag": lag, "trend_per_decade": round(beta[1] * 10, 3), "enso_coef": round(beta[2], 3),
"r2": round(r2v, 3), "phi": round(phi, 3), "sigma": round(sig, 3),
"period": "%s to %s" % (idx[0], idx[-1])}
out["enso_adjusted"] = series_out(adj, 3)
# projection of the current year
last = mon.index[-1]
year = last.year if last.month < 12 else last.year + 1
k_obs = last.month if last.year == year else 0
e_last = float(res[-1]) if k_obs else 0.0
sims = project_year(mon, n34m, year, k_obs, lag, beta, phi, sig, e_last, n34_fc)
prev = yfull[yfull.index < year]
rec_y, rec_v = int(prev.idxmax()), float(prev.max())
rank2 = float(prev.sort_values().iloc[-2]) if len(prev) > 1 else None
hist, edges = np.histogram(sims, bins=40)
# hindcast (out-of-sample, persistence Nino3.4, same number of observed months)
errs, cover = [], []
for y in range(2005, year):
if y not in yfull.index or k_obs == 0:
continue
mon_y = mon[mon.index <= pd.Period("%d-%02d" % (y, k_obs), "M")]
n34_y = n34m[n34m.index <= pd.Period("%d-%02d" % (y, k_obs), "M")]
b, rs, ph, sg, _, _ = fit_gmst(mon_y, n34_y, lag)
sm = project_year(mon_y, n34_y, y, k_obs, lag, b, ph, sg, float(rs[-1]), None, nsim=800, seed=y)
errs.append(float(np.mean(sm) - yfull[y]))
lo, hi = np.percentile(sm, [5, 95])
cover.append(lo <= yfull[y] <= hi)
out["projection"] = {
"year": year, "months_observed": k_obs, "mean": r2(np.mean(sims), 3),
"p05": r2(np.percentile(sims, 5), 3), "p95": r2(np.percentile(sims, 95), 3),
"p_record": r2(float(np.mean(sims > rec_v)), 3), "record_year": rec_y, "record_value": r2(rec_v, 3),
"p_top2": r2(float(np.mean(sims > rank2)), 3) if rank2 is not None else None,
"hist": {"counts": hist.tolist(), "edges": [r2(v, 3) for v in edges]},
"hindcast": {"n": len(errs), "rmse": r2(np.sqrt(np.mean(np.square(errs))), 3) if errs else None,
"bias": r2(np.mean(errs), 3) if errs else None,
"coverage90": r2(np.mean(cover), 2) if cover else None,
"years": "2005-%d" % (year - 1)},
"n34_source": "C3S multi-system ensemble" if n34_fc else "damped persistence (AR1)",
}
status("ERA5 daily global (Climate Pulse)", True, "to %s, projection %d = %+.2f C (1991-2020)" % (
d.index[-1].date(), year, np.mean(sims)))
return out
# --------------------------------------------------------------------------
# 4. Copernicus C3S seasonal forecasts (optional)
# --------------------------------------------------------------------------
def c3s_retrieve(client, centre, systems, year, month, variable, area, target):
last = None
for sy in systems:
try:
client.retrieve("seasonal-postprocessed-single-levels", {
"originating_centre": centre, "system": sy, "variable": [variable],
"product_type": ["monthly_mean"], "year": [str(year)], "month": ["%02d" % month],
"leadtime_month": ["1", "2", "3", "4", "5", "6"], "area": area, "data_format": "grib",
}, target)
return sy
except Exception as e: # noqa
last = e
raise RuntimeError(str(last)[:200])
def open_grib_members(path):
import xarray as xr
ds = xr.open_dataset(path, engine="cfgrib", backend_kwargs={"indexpath": ""})
v = list(ds.data_vars)[0]
da = ds[v]
ren = {}
for n in da.dims:
if n.startswith("lat"):
ren[n] = "lat"
elif n.startswith("lon"):
ren[n] = "lon"
elif n in ("forecastMonth", "step"):
ren[n] = "lead"
da = da.rename(ren)
for n in list(da.dims):
if n not in ("number", "lead", "lat", "lon") and da.sizes[n] == 1:
da = da.isel({n: 0})
if "number" not in da.dims:
da = da.expand_dims("number")
extra = [n for n in da.dims if n not in ("number", "lead", "lat", "lon")]
if extra:
# lagged ensembles (e.g. NCEP) come as a sparse hypercube (start dates x members, mostly
# empty). Read one start date at a time and keep only real members (memory safe).
import itertools
mems, nmem = [], 0
for idx in itertools.product(*[range(da.sizes[e]) for e in extra]):
sub = np.asarray(da.isel(dict(zip(extra, idx))).transpose("number", "lead", "lat", "lon").values,
dtype=np.float32)
ok = ~np.all(np.isnan(sub), axis=(1, 2, 3))
if ok.any():
mems.append(sub[ok])
nmem += int(ok.sum())
if nmem >= 200:
break
if not mems:
raise RuntimeError("no valid ensemble member in " + os.path.basename(path))
da = xr.DataArray(np.concatenate(mems, axis=0)[:200], dims=("number", "lead", "lat", "lon"),
coords={"lat": da["lat"].values, "lon": da["lon"].values})
else:
da = da.dropna("number", how="all")
return da.transpose("number", "lead", "lat", "lon")
def block_c3s():
import cdsapi
client = cdsapi.Client(quiet=True)
now = dt.date.today()
inits = [(now.year, now.month)]
pm = now.replace(day=1) - dt.timedelta(days=1)
inits.append((pm.year, pm.month))
os.makedirs(CACHE, exist_ok=True)
result = None
for (y, m) in inits:
systems = {}
pr_sys = {}
for centre, sys_list in C3S_SYSTEMS.items():
f_sst = os.path.join(CACHE, "c3s_sst_%s_%d%02d.grib" % (centre, y, m))
try:
if not os.path.exists(f_sst):
sy = c3s_retrieve(client, centre, sys_list, y, m, "sea_surface_temperature_anomaly",
[5, -170, -5, -120], f_sst)
else:
sy = "cached"
da = open_grib_members(f_sst)
w = np.cos(np.deg2rad(da["lat"]))
n34 = da.weighted(w).mean(("lat", "lon"), skipna=True).values # (member, lead)
if n34.ndim != 2 or n34.shape[1] != 6:
# e.g. NCEP/JMA lagged-start GRIBs: leads not mapped to 6 monthly values -> skip system
log(" C3S %s skipped: unexpected shape %s (expected members x 6 leads)" % (centre, str(n34.shape)))
continue
systems[centre] = {"system": sy, "members": [[r2(v, 2) for v in row] for row in n34]}
log(" C3S %s %s: %d members" % (centre, sy, n34.shape[0]))
except Exception as e:
log(" C3S %s SST failed: %s" % (centre, str(e)[:120]))
continue
f_pr = os.path.join(CACHE, "c3s_pr_%s_%d%02d.grib" % (centre, y, m))
try:
if not os.path.exists(f_pr):
c3s_retrieve(client, centre, sys_list, y, m,
"total_precipitation_anomalous_rate_of_accumulation",
[39, -20, -37, 53], f_pr)
da = open_grib_members(f_pr).sortby("lat")
em = da.mean("number").values * 86400.0 * 1000.0 # m/s -> mm/day
pr_sys[centre] = (da["lat"].values, da["lon"].values, em)
except Exception as e:
log(" C3S %s precip failed: %s" % (centre, str(e)[:120]))
if len(systems) >= 2:
months = [str(pd.Period(year=y, month=m, freq="M") + k) for k in range(6)]
result = {"init": "%d-%02d" % (y, m), "months": months, "systems": systems}
# probabilities for 3-month windows (equal weight per system)
probs = []
for a in range(0, 4):
pen, pln = [], []
for s in systems.values():
mem = np.array([[np.nan if v is None else v for v in row] for row in s["members"]])
sm = np.nanmean(mem[:, a:a + 3], axis=1)
pen.append(np.mean(sm >= 0.5))
pln.append(np.mean(sm <= -0.5))
c = pd.Period(months[a + 1], "M")
probs.append({"season": SEAS[c.month - 1], "centre": str(c), "p_en": r2(np.mean(pen), 3),
"p_ln": r2(np.mean(pln), 3), "p_neu": r2(1 - np.mean(pen) - np.mean(pln), 3)})
result["probs"] = probs
if pr_sys:
lat, lon, _ = next(iter(pr_sys.values()))
wins = []
for a in range(0, 4):
fields = [np.nanmean(em[a:a + 3], axis=0) for (_, _, em) in pr_sys.values()
if em.shape[1:] == (len(lat), len(lon))]
F = np.array(fields)
mme = np.nanmean(F, axis=0)
agree = np.mean(np.sign(F) == np.sign(mme)[None], axis=0)
c = pd.Period(months[a + 1], "M")
wins.append({"season": SEAS[c.month - 1], "centre": str(c),
"z": [[r2(v, 2) for v in row] for row in mme],
"agree": [[r2(v, 2) for v in row] for row in agree]})
result["precip"] = {"lat": [float(v) for v in lat], "lon": [float(v) for v in lon],
"n_systems": len(pr_sys), "windows": wins}
break
if result is None:
raise RuntimeError("fewer than 2 C3S systems available")
status("Copernicus C3S seasonal", True, "init %s, %d systems" % (result["init"], len(result["systems"])))
return result
# --------------------------------------------------------------------------
def main():
global ARGS
ap = argparse.ArgumentParser(description="AfriENSO data pipeline")
ap.add_argument("--no-c3s", action="store_true", help="skip Copernicus C3S seasonal forecasts")
ap.add_argument("--refresh", action="store_true", help="ignore the download cache")
ap.add_argument("--offline", action="store_true", help="use cached files only (testing)")
ARGS = ap.parse_args()
t0 = time.time()
geo = json.load(open(GEO_JSON, encoding="utf-8"))
data = {"meta": {}, "geo": {"countries": [{k: c[k] for k in ("iso", "en", "fr", "draw", "bbox")}
for c in geo["countries"]],
"ar6": geo["ar6"], "wa": WA_BOXES,
"wsahara": geo["wsahara"], "source": geo["source"]}}
oni = eps = mon = None
try:
data["enso"], oni, eps, mon = block_enso()
except Exception as e:
status("CPC ONI", False, e)
traceback.print_exc()
sst_idx = None
try:
sst_idx, sstmap = block_sst_indices()
data["sst"] = {k: series_out(v.dropna()) for k, v in sst_idx.items()}
data["sstmap"] = sstmap
except Exception as e:
status("PSL ERSSTv5 gridded", False, e)
traceback.print_exc()
c3s = None
if not ARGS.no_c3s:
try:
c3s = block_c3s()
data["c3s"] = c3s
except Exception as e:
status("Copernicus C3S seasonal", False, e)
n34m = mon["n34"] if mon is not None else None
n34_fc = None
if c3s:
mem = [row for s in c3s["systems"].values() for row in s["members"]]
n34_fc = {"months": c3s["months"], "members": [[np.nan if v is None else v for v in r] for r in mem]}
try:
data["globaltemp"] = block_globaltemp(n34m, n34_fc)
except Exception as e:
status("ERA5 daily global (Climate Pulse)", False, e)
traceback.print_exc()
if oni is not None:
try:
phase_m = phase_series(oni, eps)
data["africa"] = block_africa(geo, oni, phase_m, sst_idx)
except Exception as e:
status("Africa grids", False, e)
traceback.print_exc()
data["meta"] = {"generated": dt.datetime.utcnow().strftime("%Y-%m-%d %H:%M UTC"), "status": STATUS,
"base": "%d-%d" % (BASE0, BASE1), "fdr_q": FDR_Q, "version": "1.1"}
js = "window.AFRIENSO = " + json.dumps(data, separators=(",", ":"), ensure_ascii=False, allow_nan=False) + ";\n"
with open(OUT_JS, "w", encoding="utf-8") as f:
f.write(js)
log("wrote %s (%.1f MB) in %.0f s" % (OUT_JS, len(js) / 1e6, time.time() - t0))
bad = [s["key"] for s in STATUS if not s["ok"]]
if bad:
log("Blocks that failed (the page hides them): " + ", ".join(bad))
if __name__ == "__main__":
main()
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
AfriENSO - data pipeline
========================
Downloads the public datasets used by the AfriENSO dashboard, computes every
statistic shown on the page and writes one file, ``afrienso_data.js``, that the
static page (index.html) reads. Run it on your own computer, then upload
``afrienso_data.js`` to the Hugging Face Space (or run ``publish_hf.py``).
python update_data.py # normal run (uses a 12 h download cache)
python update_data.py --no-c3s # skip the Copernicus C3S forecasts
python update_data.py --refresh # ignore the cache, download everything
Data sources (all public):
* NOAA CPC - ONI (ERSSTv5), RONI, weekly OISST v2.1 Nino indices, monthly
ERSSTv5 Nino indices
* NOAA PSL - ERSSTv5 gridded SST, GPCP v2.3 monthly precipitation,
GHCN_CAMS monthly land air temperature (THREDDS/OPeNDAP)
* Copernicus Climate Pulse - ERA5 daily global mean 2 m temperature
* Copernicus C3S (optional, needs a free CDS account + ~/.cdsapirc) -
multi-system seasonal forecasts of SST and precipitation
October 2026 (v1.1, co-authors' review of the associated paper): adds the West
African land boxes GC, WSHL, CSHL and SHL (units of kind "wa", geo["wa"]) and the
SST indices SENC (Senegal-Mauritania coast), MED (Mediterranean) and AMG (TNA-ATL3).
Author: Dr Ibrahima Diouf - University of Labe / LPAO-SF, ESP, UCAD
"""
import argparse
import datetime as dt
import json
import math
import os
import re
import sys
import time
import traceback
import warnings
import numpy as np
import pandas as pd
import requests
HERE = os.path.dirname(os.path.abspath(__file__))
CACHE = os.path.join(HERE, "_cache")
OUT_JS = os.path.join(HERE, "afrienso_data.js")
GEO_JSON = os.path.join(HERE, "africa_geo.json")
URL = {
"oni": "https://www.cpc.ncep.noaa.gov/data/indices/oni.ascii.txt",
"roni": "https://www.cpc.ncep.noaa.gov/data/indices/RONI.ascii.txt",
"weekly": "https://www.cpc.ncep.noaa.gov/data/indices/wksst9120.for",
"ersst_nino": "https://www.cpc.ncep.noaa.gov/data/indices/ersst5.nino.mth.91-20.ascii",
"era5_daily": "https://sites.ecmwf.int/data/climatepulse/data/series/era5_daily_series_2t_global.csv",
"ersst_dap": "https://psl.noaa.gov/thredds/dodsC/Datasets/noaa.ersst.v5/sst.mnmean.nc",
"gpcp_dap": "https://psl.noaa.gov/thredds/dodsC/Datasets/gpcp/precip.mon.mean.nc",
"ghcn_dap": "https://psl.noaa.gov/thredds/dodsC/Datasets/ghcncams/air.mon.mean.nc",
"ersst_file": "https://downloads.psl.noaa.gov/Datasets/noaa.ersst.v5/sst.mnmean.nc",
"gpcp_file": "https://downloads.psl.noaa.gov/Datasets/gpcp/precip.mon.mean.nc",
"ghcn_file": "https://downloads.psl.noaa.gov/Datasets/ghcncams/air.mon.mean.nc",
}
GPCC_DAP = { # GPCC Full Data (to 2020) + Monitoring (recent months), 1 deg, mm/month
"full": ["https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/full_v2022/precip.mon.total.1x1.v2022.nc",
"https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/full_v2020/precip.mon.total.1x1.v2020.nc"],
"monitor": ["https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/monitor/precip.monitor.mon.total.1x1.v2022.nc",
"https://psl.noaa.gov/thredds/dodsC/Datasets/gpcc/monitor/precip.monitor.mon.total.1x1.v2020.nc"],
}
DRY_MM = 30.0 # seasonal (3-month) rainfall below which a season is "dry": not assessed
STRONG_THR = 1.5 # strong El Nino: NDJ ONI >= 1.5 C
BASE0, BASE1 = 1991, 2020 # WMO climatological reference period
PI_OFFSET_DEFAULT = 0.88 # C3S: 1991-2020 is 0.88 C above 1850-1900
AFRICA_BOX = (-37.0, 39.0, -20.0, 53.0) # lat0, lat1, lon0, lon1
FDR_Q = 0.10 # Wilks (2016): alpha_FDR = 2 x alpha_global
SEAS = ["DJF", "JFM", "FMA", "MAM", "AMJ", "MJJ", "JJA", "JAS", "ASO", "SON", "OND", "NDJ"]
# C3S systems tried in order (the most recent first). Edit if C3S upgrades.
C3S_SYSTEMS = {
"ecmwf": ["51"], "meteo_france": ["9", "8"], "ukmo": ["606", "605", "604", "603", "602"],
"dwd": ["22", "21"], "cmcc": ["35", "4"], "ncep": ["2"], "jma": ["4", "3", "2"],
"eccc": ["5", "4"],
}
STATUS = []
warnings.filterwarnings("ignore", category=RuntimeWarning)
warnings.filterwarnings("ignore", category=DeprecationWarning)
ARGS = None
def log(msg):
print(time.strftime("%H:%M:%S"), msg, flush=True)
def status(key, ok, msg=""):
STATUS.append({"key": key, "ok": bool(ok), "msg": str(msg)[:300],
"time": dt.datetime.utcnow().strftime("%Y-%m-%d %H:%M UTC")})
log(("OK " if ok else "FAIL ") + key + (" - " + str(msg)[:200] if msg else ""))