⌨ Learn · Code sandbox

Run real analysis in your browser.

This is real Python — xarray, numpy, scipy, pymannkendall — running entirely in your browser (via Pyodide/WebAssembly). No install, no NASA login, no backend. The data is synthetic but mirrors the real product (same dimensions, units, seasonality, noise and gaps) and carries a known planted signal — so when your analysis recovers it, you've verified your own method. That's the whole game.

dataset: runtime not loaded — first run fetches ~20 MB (cached after)

# Synthetic rainfall that MIRRORS GPM IMERG (monthly, mm, lat x lon) over a region,
# with a hidden trend planted in it. Recover it with the real method, then check.
import numpy as np, pymannkendall as mk

ds = load_synthetic("rainfall")
print(ds, "\n")                       # same shape as real IMERG: (time, lat, lon), mm

# area-weight by cos(lat) -> annual TOTAL -> one regional number per year
w = np.cos(np.deg2rad(ds.lat))
annual = ds["precipitation"].groupby("time.year").sum()
series = annual.weighted(w).mean(dim=["lat", "lon"])
years, vals = series["year"].values, series.values

# robust trend + significance (exactly the pipeline: Theil-Sen + Mann-Kendall)
r = mk.original_test(vals)
print(f"Recovered trend: {r.slope:+.2f} mm/yr   (Mann-Kendall p = {r.p:.3f}, "
      f"{'significant' if r.h else 'not significant'})")
print(f"Planted truth:   {reveal('rainfall'):+.2f} mm/yr   <- your analysis should be close!")

import matplotlib.pyplot as plt
plt.figure(figsize=(6, 3))
plt.plot(years, vals, "o-", label="annual rainfall")
plt.plot(years, r.intercept + r.slope * np.arange(len(years)), "--", label="Theil-Sen fit")
plt.title("Synthetic annual rainfall + recovered trend"); plt.xlabel("year"); plt.ylabel("mm/year")
plt.legend()

output

Click ▶ Run. The first run loads the Python runtime (~10–20 s); after that it's instant.

Pilot: 2 synthetic datasets. The plan is one for every variable archetype (LST, NO₂, soil moisture, GRACE, SAR, nightlights…), each mirroring its real product and carrying a planted signal, wired into the matching concept guides so every method is runnable. The synthetic data is generated by load_synthetic() — read it, it's honest about how each product is structured.