Has the monsoon’s arrival date over my region shifted earlier or later?

The monsoon onset is the day each year when the steady summer rains finally arrive and stay. Not a lone teaser shower, but the real wet season switching on. Farmers plant to it, reservoirs fill from it, and a late or early start ripples through a whole region. This page finds that switch-on date from space, one year at a time, and asks whether it’s been creeping earlier or later over the past couple of decades.

Which data to use, and how

Use this dataset: GPM IMERG daily (short name GPM_3IMERGDF), NASA’s satellite rainfall estimate covering 2000 → today at about ~10 km, almost everywhere on Earth. It gives you one rainfall number per grid cell per day, which is exactly what an onset rule needs. The other datasets here (ERA5 winds and MERRA-2) are optional confirmations you can add later. Start with IMERG alone.

Then do four steps (the runnable code further down does all of this; this is just the idea):

1. Pick your region and the season. A latitude/longitude box (onset is a big, smeared-out signal, so use a box at least a degree or so wide) and the months when the monsoon could arrive.
2. One rainfall number per day. Average IMERG’s rainfall over your box for each day, giving a daily time series for each year.
3. Define “onset” and find it. Onset is the first day, on or after an earliest permitted date, where the daily area-average rainfall stays above a threshold for several consecutive days. Record that day-of-year (1–365) for each year.
4. Fit a trend and test it. Line up 20+ years of onset dates, draw the long-term line, and check it’s bigger than the year-to-year wobble. Earlier = onset drifting forward, later = drifting back, flat = no real change.

That’s the whole method. Below, “How a scientist answers this” names the exact tools, and the Run it cell does it live on synthetic data so you can see each step work.

What you can find out

• The onset date for each year over your region, from an explicit, reproducible rainfall rule.
• The multi-year trend in that date: the long-term slope (a robust line called Theil–Sen) and whether it’s a real signal or just noise (a significance test called Mann–Kendall).
• How much the answer depends on your definition. Re-run with a different threshold or consecutive-day count and watch the date move.
• Whether the real monsoon circulation showed up. Optionally check that the monsoon winds and moisture (ERA5) arrived along with the rain, not just a stray downpour.
• Year-to-year variability, meaning how widely onset bounces around its average.

What it can’t tell you

• The official onset date. India’s Met Department (IMD) declares onset using ground stations, winds, and an extra satellite signal (outgoing longwave radiation, a proxy for deep clouds). A satellite-rainfall-only estimate is your own reproducible number, and it can differ from the official call by days. It’s an estimate, not a declaration.
• Trends before the year 2000. IMERG’s record starts in 2000. For a longer history you’d switch to rain-gauge products (like IMD’s gridded data) or reanalysis, each with its own caveats.
• When the monsoon ends or pauses. Withdrawal and “break” spells need their own separate rule; this one only finds the start.
• Why onset shifted. Spotting a trend isn’t explaining it. Attributing a shift to a cause needs climate models and large-scale driver analysis beyond these datasets.

Gotchas to watch for

• The answer wobbles with your definition. Onset can swing by many days if you nudge the rainfall threshold or the number of consecutive wet days, and there’s no single “true” rule to fall back on. State your rule explicitly and re-run with two or three variations so you can show the result holds up.
• Satellite rain has a known bias over mountains. IMERG tends to underestimate rainfall forced upward by terrain (this orographic underestimation is well documented over the Western Ghats) and is shaky for extreme downpours. Your onset date inherits that bias, so keep it in mind and lean on the ground-truth comparison below.
• The record is short. ~25 years of IMERG isn’t a lot of statistical muscle, and a slope can look like a real shift while still carrying a weak p-value (p > 0.05 means “could easily be chance”). Report the confidence interval alongside the slope, never the slope on its own.
• A fake onset. A strong pre-monsoon shower can trip a rainfall-only rule before the real monsoon arrives. To filter out these false starts, optionally require the monsoon winds and moisture (ERA5 low-level westerlies / moisture flux) to be established too.
• Data freshness and version. The final, best IMERG product lags reality by ~3.5 months. Use version V07, and re-download (reprocess) any older granules you cached from an earlier version.

The real-data code

The Run it cell above runs the method on synthetic data with no login. Below is the whole recipe against the real archive: search IMERG, open each year’s daily rainfall, area-average it, find the sustained-rainfall onset day, then fit the Theil–Sen + Mann–Kendall trend and plot it. It needs a free Earthdata Login.

import earthaccess, numpy as np, xarray as xr
import scipy.stats as ss
import matplotlib.pyplot as plt

earthaccess.login(strategy="netrc")

# AOI: Kerala (monsoon onset gateway) — a (West, South, East, North) box
aoi = (74, 8, 78, 13)
years = range(2001, 2025)

# --- Onset rule (IMD-style family — SET AND VERIFY for your region) ---
# Onset = first day on/after EARLIEST where area-mean rainfall stays above
# RAIN_MM for CONSEC consecutive days. Thresholds vary by definition; re-run
# with a couple of choices for robustness.
EARLIEST = "05-10"      # earliest permitted onset (mm-dd)
RAIN_MM  = 5.0          # daily area-mean rainfall threshold (mm/day) — verify
CONSEC   = 5            # consecutive wet days required — verify

def onset_doy(da_daily):
    """da_daily: 1-D xarray of area-mean daily rainfall (mm/day) for one year."""
    wet = (da_daily >= RAIN_MM).astype(int)
    roll = wet.rolling(time=CONSEC).sum()        # wet days in trailing window
    hit = roll.where(roll == CONSEC, drop=True)  # first window that's all-wet
    if hit.size == 0:
        return np.nan
    # the window END is the CONSEC-th wet day; onset is its start day
    return int(hit.time.dt.dayofyear[0]) - (CONSEC - 1)

# 1. One onset date per year: search the year's IMERG granules, open them,
#    area-average rainfall over the box each day, run the onset rule.
onsets = []
for yr in years:
    grans = earthaccess.search_data(
        short_name="GPM_3IMERGDF", bounding_box=aoi,
        temporal=(f"{yr}-{EARLIEST}", f"{yr}-07-31"))
    if not grans:
        onsets.append(np.nan); continue
    ds = xr.open_mfdataset(earthaccess.open(grans), combine="by_coords")
    # IMERG daily 'precipitation' is mm/day on (time, lon, lat)
    box = ds["precipitation"].sel(lon=slice(aoi[0], aoi[2]),
                                  lat=slice(aoi[1], aoi[3]))
    w = np.cos(np.deg2rad(box.lat))              # area weight by latitude
    daily = box.weighted(w).mean(("lon", "lat")).sortby("time").compute()
    onsets.append(onset_doy(daily))
    print(f"{yr}: onset day-of-year = {onsets[-1]}")

# 2. Robust trend in the onset date + significance over the record.
yrs = np.array(list(years), float)
od  = np.array(onsets, float)
m   = ~np.isnan(od)
slope, intercept = ss.theilslopes(od[m], yrs[m])[:2]   # days per year (Theil–Sen)
tau, p = ss.kendalltau(yrs[m], od[m])                  # Mann–Kendall monotonic test

# 3. One-line verdict.
drift = "earlier" if slope < 0 else "later"
sig = "significant" if p < 0.05 else "not distinguishable from noise"
print(f"Monsoon onset is drifting {drift}: {slope*10:+.2f} days/decade, "
      f"p={p:.3f} ({sig})")

# 4. Plot the onset dates and the fitted Theil–Sen line.
plt.scatter(yrs[m], od[m], label="onset day-of-year")
plt.plot(yrs[m], intercept + slope * yrs[m], "r--", label="Theil–Sen trend")
plt.gca().invert_yaxis()        # earlier dates upward
plt.ylabel("onset day-of-year"); plt.xlabel("year")
plt.legend(); plt.tight_layout(); plt.show()

# Before you trust it: compare these dates against IMD's official Kerala onset
# declarations — satellite-rainfall-only onset can differ by days, and IMERG
# underestimates orographic rain over the Western Ghats (see the gotchas).

Where the data comes from

The rainfall comes from NASA through one free login (Earthdata Login): GPM IMERG and MERRA-2 both live in the GES DISC archive, so if you stick to those two you only need a single account. The optional wind/moisture confirmation, ERA5, comes from Europe’s Copernicus / C3S service instead. Only sign up for that second account if you actually want to pull ERA5.

Sources

• GPM IMERG: https://gpm.nasa.gov/data/imerg
• IMERG V07 user guide: https://gpm.nasa.gov/sites/default/files/2023-07/IMERG_V07_ReleaseNotes_230713-signed.pdf
• ERA5 (C3S): https://cds.climate.copernicus.eu/datasets/reanalysis-era5-single-levels
• IMD monsoon onset (Kerala) background: https://mausam.imd.gov.in/