Field notes: three country teams, one chart request, three different answers

Scenic mountain valley with a river flowing through a village on a sunny day.

It came up in the same week, from three task teams that don't talk to each other much. Could we get a light-intensity trend for the programme area, same as last quarter. Simple request. The kind you'd expect to take an analyst an afternoon.

What came back were three charts that didn't agree on what "trend" meant.

The Mozambique team pulled a monthly composite from a sensor product that corrects for moonlight and smooths over cloud cover before anyone sees a pixel value. The Bangladesh team used a raw radiance layer, no lunar correction, pulled on whatever day in the month had the clearest overpass. The Kenya team was comparing two different years of the same dataset, except the vendor had quietly reprocessed its calibration in between, so year one and year two weren't built the same way even though the file names matched.

None of this showed up as an error. Each chart looked fine on its own. Side by side, which is exactly what a portfolio review does, the cracks show. One line sloped up. A second looked flat for the whole quarter. The third had a spike in month two that turned out to be a full moon, not a substation coming online.

Letting each country office pick its own light-intensity source costs you comparability. A single number from any one team can be defensible on its own terms, and the portfolio read still breaks, because "progress" has to mean the same thing across every project in the review, and these three didn't.

A few things worth flagging for anyone building out a multi-country indicator set this way:

Lunar correction is a design choice a team makes, whether or not they realize they're making it. Two products can both claim "night-light radiance" and handle moon glow completely differently. If your teams aren't told which approach is standard, they'll pick whatever the tutorial they found uses.

Reprocessing moves the baseline out from under a time series. Vendors recalibrate their sensors and pixel-value math on their own schedule, and nothing in the file tells you a recalibration happened between your month-1 pull and your month-24 pull. Pin a processing version at the start of the programme, or at minimum log the version in the metadata every quarter, so a shift in the numbers can be traced back to the source instead of read as a change on the ground.

"Monthly" means different things depending on the vendor. A single clearest-night pull and a true monthly composite produce different noise profiles, and noise reads as signal to anyone skimming a chart in a review deck.

Light intensity is still one of the few things you can check between a baseline survey and a midterm review without waiting on a household count. Three teams doing honest, careful work still produced three charts that couldn't sit on the same slide, because nobody had agreed on the source, the resolution, the cadence, or the boundary before anyone started pulling data.

Agree on one source, one resolution, one cadence, and one definition of the programme-area boundary, and write it down somewhere a new analyst will find before they build their own pull. A standardized reporting template closes more of this gap than any amount of reconciliation after the fact.

We built Electrification Map around that constraint: one panchromatic layer, one resolution, updated on the same monthly cadence, so a chart from your Mozambique programme and a chart from your Bangladesh programme are built the same way before anyone sets them side by side.

If your next portfolio review needs a light-intensity trend that means the same thing across every project it covers, that's the problem this is built to remove.

Start a project

Back to the blog