What a satellite actually measures · Learn

You're a coder. You know how to manipulate arrays. The thing that's genuinely new here isn't the code — it's what the numbers in the array mean. Almost every mistake beginners make comes from skipping this. So, the mental model first.

Platform, sensor, product — keep them straight

A platform is the satellite (Terra, Sentinel-2, GPM).
A sensor / instrument is the camera or radar bolted onto it (MODIS, OLI, SAR).
A product is the processed data you actually download (e.g. MOD13A2 NDVI). You analyse products, never raw sensors.

In plain EnglishThe satellite is the truck, the sensor is the camera, and the product is the photo album someone already sorted and labelled for you. You work with the album.

Passive vs. active — and why it matters in the monsoon

Passive sensors measure light the Earth reflects (sunlight) or emits (heat). They need daylight and are blocked by cloud — optical and thermal cameras. Active sensors send their own signal and listen for the echo — radar (SAR) and lidar. They work day or night and see through cloud.

Why you careDuring the Indian monsoon, optical satellites are blind exactly when a flood happens — it's cloudy. Radar (Sentinel-1, NISAR) is not. Choosing the wrong sensor type is the difference between "no data" and a flood map.

The four resolutions — always ask all four

Every product is a trade-off across four "resolutions". Before using one, ask each:

Spatial — how big is one pixel? (Sentinel-2 10 m · MODIS 250 m–1 km · IMERG rain ~10 km · GRACE ~300 km.) ⚠️ An Indian farm field is often smaller than one MODIS pixel — so a "field" measurement is really an average of several fields. This is the mixed-pixel problem.
Temporal — how often does it revisit? (Sentinel-2 ~5 days · MODIS daily · Landsat 16 days · a weather satellite like INSAT, every few minutes.)
Spectral — which colours/wavelengths? (visible, near-infrared, shortwave-infrared, thermal, microwave.) Vegetation indices are built from combinations of bands.
Radiometric — how finely it measures brightness. Rarely your bottleneck.

You cannot have fine pixels and frequent revisit from one sensor — a wide camera revisits often but sees coarsely; a sharp camera revisits rarely. That single trade-off explains most of the satellite zoo.

Pixel size 250 m

"Data" is already a model — the processing levels

The number in your array is not raw reality. It has been through a pipeline, labelled by processing level:

L0 raw instrument counts → L1 calibrated radiances, geolocated → L2 the geophysical variable per pixel (NDVI, temperature, NO₂) → L3 gridded & time-composited → L4 model-assimilated / gap-filled.
You almost always want L2 or L3.

The mindset that makes you good at thisBy the time you read a "temperature", a retrieval algorithm, an atmosphere, a grid, and sometimes a model already stood between the sensor and your number. That's not a reason to distrust it — it's the reason you'll always cross-check it. Every step adds error; the whole back half of this course is about measuring that error and refusing when it's too big.

So what is a pixel, really?

One pixel is one estimate of one variable, over one patch of ground, at one time — with error baked in. An "image" is a grid of those estimates. A "time series" is the same patch through time. That's the whole game; everything else is bookkeeping (which the labelled-array tools handle for you).

Next: stop reading and run a real analysis in your browser — no install — and watch the method recover a signal you can check.