Many people thought Earth observation was expensive or exclusive. Having a view of your city or neighborhood was more like an episode from a spy movie, not an ordinary feature in your smartphone. But times have changed. Today we can get tons of satellite data in a matter of minutes on any device you have, even at no cost. 

Copernicus Sentinel-2 is one of its most popular sources. It revisits our planet every five days at 10-meter resolution, giving an ordinary user with internet access the possibility to get a fresh view from above. So, actually, we have a lot of raw material that is current and cheap. But what matters most is how and by whom this data is used, what decisions it can offer, and how anyone who can pull free satellite images from Copernicus, USGS, or NASA can benefit. Agriculture, insurance, urban planning, and credit analysis each get a worked example.

The Data Landscape: What’s Available and Why It Matters

The images we see in various map applications are the tip of the iceberg. Satellites offer us many more options for personal and commercial use:

• Optical (multispectral, hyperspectral): these sensors record sunlight bouncing off the surface, split into visible, near-infrared and shortwave-infrared bands. Same physics as a camera, with extra colors the eye can’t see. Sentinel-2, Landsat 8 and 9 trade sharpness and speed for a far longer archive — 30 meters, sixteen days. 
• Radar (SAR): These sensors don’t need light. The satellite sends out its own microwave pulses and measures what comes back, so a monsoon season or an Arctic winter night is no obstacle. Sentinel-1 is the free workhorse here, with a 6-to-12-day revisit.
• Thermal infrared — sensors with thermal bands measure heat on the surface. Landsat’s TIRS instrument resolves this at 100 m (delivered resampled to 30 m), while MODIS covers the whole planet twice a day at roughly 1 km. The practical uses are irrigation status through evapotranspiration, urban heat-island mapping, and spotting industrial activity from cooling water and flare stacks.

The Analytics Workflow: From Raw Rasters to Business Metrics

Choosing the way of how you get the images is half the job. But many processes are run behind the scenes between a downloaded scene and a number somebody acts on. Here are some technical details. 

Preprocessing and quality control

The first thing you do with satellite data is delete a lot of it.

Every scene comes with a quality layer telling you which pixels were under cloud or shadow when the picture was taken. That’s a free map of your missing values, and people still ignore it. Don’t. Throw those pixels out before you average anything, because one thin streak of cloud can pull a field’s monthly number further off than whatever change you were trying to find.

Step two is where new people get tripped up. Take a reading over the same field in March, then again in November, and the two numbers won’t match — even if the crop did nothing. That’s the air: haze, dust, water vapour, a few kilometres of it. Sen2Cor takes it out. Without that, your March value and your November value aren’t the same kind of number, and what you’re really tracking is weather.

Then compositing, the step people skip most. Instead of hunting for one nice cloud-free day, take the median of every usable pixel across the whole month and your dataset turns into a stack of sunny days, which is not the same thing as a year.

Feature extraction

The next important step is to figure out for what you need this imagery. Because without knowing the subject matter — farming, forestry, insurance, whatever you’re working on — you produce just a nice-looking picture. With it you produce numbers someone can act on.

• Vegetation indices — NDVI for greenness; NDWI for water content; EVI where dense canopy saturates NDVI and flattens exactly the variation you need.
• Burn indices — NBR for fire scars and severity classes, differenced pre- and post-event.
• Texture and shape — GLCM metrics distinguish built-up surfaces from bare soil, which are close to spectrally identical.
• Temporal statistics — per-pixel mean, standard deviation, trend slope and seasonality coefficients over a full year. These four columns often outperform any single-date image.

Exploratory analysis

Plot it before you model it.

Pull the time series for a few pixels you already know — a wheat field, a patch of forest, a car park — and look at the shape. Cropland should give you one or two clean humps a year, planting through harvest. If it doesn’t, stop there. The bug is somewhere earlier and no model will fix it for you.

In most data-science work you’re taught to treat outliers as noise and strip them out. Here that instinct destroys the thing you’re being paid to find. A sharp drop in NDVI is the event — something happened to that field. Removing it removes the finding, because a sudden drop in NDVI might be drought or harvest, or pests. Sorting them out is your job rather than the model’s, and local rainfall and temperature records usually settle it. 

Visualisation and decision support

Your dashboard or report you build, has to work for a busy non-specialist. Picture the actual reader: a regional operations manager who has about eleven minutes between meetings and has no background in remote sensing. If they can’t get the answer out of your output in that window, the output failed — however good the analysis underneath it was. 

Dashboards help — Plotly Dash, R Shiny, Kepler.gl — provided a regional manager can filter down to their own district and dates without phoning you first. Static maps still earn their place in board papers.