Nightlights — mapping power outages from space · Learn

In one lineCompare a clear night before a storm with a clear night after. Neighbourhoods whose light dropped are the likely outage areas. Simple in principle — the skill is in the caveats.

What Black Marble measures

The VIIRS Day/Night Band on the Suomi-NPP and NOAA satellites is sensitive enough to see moonlit clouds and city lights. NASA's Black Marble (product VNP46) cleans that raw radiance — correcting for the angle of view, moonlight, and stray light — into a nightly, comparable measure of how much light each ~500 m patch of ground emits. Black Marble HD sharpens it to neighbourhood scale.

Because grid power is most of what lights a city at night, a sudden, widespread drop in radiance after a disaster is a strong proxy for power loss — the analysis behind the Disasters Portal's Hurricane-Helene maps.

Play with it

Each cell is a city block, lit at night. Knock out power and watch the radiance drop. Then flip on thin cloud and moonlight — and see why this is a proxy you read carefully, not a meter.

Power lost Sky thin cloud + moonlight

measured radiance drop:

Do it yourself

editable · runs in your browser

import numpy as np
rng = np.random.default_rng(3)
pre  = rng.uniform(20, 60, 500)                  # Black Marble night-light radiance, before
post = pre * 0.55 + rng.normal(0, 3, 500)        # a ~45% outage, after
qa_before = rng.integers(0, 3, 500); qa_after = rng.integers(0, 3, 500)   # 0 = good pixel
good = (qa_before == 0) & (qa_after == 0)
drop = 1 - post[good].sum() / pre[good].sum()
print("radiance drop:", str(round(drop * 100)) + "%", "over", int(good.sum()), "clear pixels")
print("MASK on the QA flags FIRST - cloud/bad pixels would fake a drop.")
print("and a drop is NOT a confirmed outage - cross-check it (lesson 4.1).")

The honest caveats

It's a proxy, not a utility feed. Light loss correlates with outages; it doesn't count customers.
Cloud, moonlight, snow and even wildfire smoke distort nighttime radiance — always use the QA flags and compare clear nights, or you'll mistake weather for an outage.
A storm blinds the sensor. Thick storm cloud can hide the surface for days — read the first clear night after.
Single nights are noisy — a multi-night composite is steadier (the signal-vs-noise idea again).