← Lillie AcademyCourse contentslillieearthintelligence.com

Module 0.2: How Earth Observation Works

Phase: 0, Foundations
Level: Beginner
Estimated time: 6 to 8 hours
Prerequisites: Module 0.1 (your toolkit and eo-portfolio repo). No physics background required.
Portfolio thread: produces a sensor-and-mission cheat sheet and your first real scene pull, both committed to eo-portfolio. It gives you the mental model every later module assumes.

Why this module

You cannot reason about burn severity or radar backscatter until you know what a satellite actually measures. Most beginners jump straight to code and stay confused about why an optical image is empty under clouds, why one sensor sees a fire that another misses, or why "resolution" means four different things. This module builds the mental model of Earth observation so that from Module 1 onward the hard part is method, not confusion about the data. In interviews, the people who can explain sensor trade-offs in plain language read as engineers, not button-pushers.

Learning outcomes

After this module you can:

  1. Explain passive versus active remote sensing and give an example of each.
  2. Describe the regions of the electromagnetic spectrum used in EO and what each one reveals about the ground.
  3. Define the four resolutions (spatial, spectral, temporal, radiometric) and the trade-offs among them.
  4. Name the major open EO missions and say what each is good for.
  5. Explain data processing levels and what "analysis-ready" means.
  6. Register for the main data portals and pull your first scene over your study area.

Concept lessons (about 3.5 hours)

Lesson 1: How a satellite "sees"

Remote sensing measures electromagnetic energy. The Sun floods the Earth with energy across a spectrum; surfaces reflect some of it and emit their own (as heat). A sensor in orbit records how much energy arrives in specific wavelength ranges called bands. Wavelength and frequency are two views of the same wave, linked by c = wavelength x frequency (c is the speed of light), so a shorter wavelength means a higher frequency and more energy. Earth observation uses only a small slice of the full spectrum, shown below.

The electromagnetic spectrum with a zoom on the wavelengths and frequencies Earth observation uses
The electromagnetic spectrum (top) and the Earth-observation working range (bottom), with wavelength and frequency for each band. Optical sensors use reflected sunlight and emitted heat; radar uses microwave bands whose longer wavelengths penetrate cloud and, increasingly, vegetation and soil.

Two families matter:

The spectrum regions you will use, with their wavelength and frequency ranges and what each reveals:

One more reason these particular bands are used: the atmosphere is only transparent in certain "windows." The visible, parts of the infrared, and the microwave region pass through; much of the ultraviolet and many infrared wavelengths are absorbed. Clouds block the visible and much of the infrared but are nearly transparent to microwaves, which is the physical reason radar sees through cloud and works day or night.

A useful habit: whenever you pick an index or a sensor later, ask "which part of the spectrum am I relying on, and does it see the thing I care about?"

Lesson 2: The four resolutions and their trade-offs

"Resolution" means four different things, and they trade against each other.

The trade-off to internalize: you rarely get high spatial, high spectral, and high temporal resolution at once. A sensor with tiny pixels usually images a narrower strip less often; a sensor that sees the whole planet daily has coarse pixels. Choosing a data source is choosing which resolution matters most for the job. For a fast-moving fire you may accept coarse pixels for daily coverage; for mapping a burn scar's edges you want fine pixels and can wait for a clear pass.

Lesson 3: The major data sources

Learn these as a toolkit, not a list:

The skill is matching source to question: daily fire detection points to VIIRS; a sharp burn-severity map points to HLS or Sentinel-2; cloud-proof change points to Sentinel-1; canopy height points to GEDI fused with imagery.

Lesson 4: Processing levels and analysis-ready data

Raw sensor data is not what you model on. Products come in levels:

Two rules of thumb: prefer surface-reflectance (Level 2) products for optical work, and always read and apply the quality or cloud mask that ships with the product. Skipping the mask is the most common source of silently wrong results.

Lesson 5: The application landscape, and where this course goes deep

Earth observation is applied across a huge range of fields, and it helps to see the whole map early, because the skills you build in this course transfer across almost all of it.

A grouped map of Earth-observation application fields across land, water, solid Earth, atmosphere, hazards, and the built environment, with the ones this course teaches highlighted
The application landscape. The Phase 0 to IV core (cloud-native data, SAR, machine learning, honest uncertainty, and deployment) is the same toolkit everywhere; the fields differ mainly in the data and the labels. Filled markers are taught now (as core modules or electives); the rest are natural elective directions.

The important idea is that the core of this course is not "wildfire tricks." It is a transferable method spine: turn a cloud archive into an analysis-ready cube, engineer features or train a network, validate honestly with spatial cross-validation and uncertainty, and ship the result as a service. That spine is what agriculture, water, urban, coastal, air-quality, and disaster teams all use; only the data sources and the labels change.

This course goes deep on wildfire, forestry and carbon, SAR, and floods and hazards, because depth in a real niche is what gets you hired, and it adds optional electives (starting with geology and geophysics) so you can specialize toward the fields you want. When you finish the core, adapting to a new application is a matter of swapping the dataset and the target, not learning a new craft.

Guided lab (about 3 hours): accounts, a cheat sheet, and your first scene

Open in Colab Open in GitHub Codespaces

Step 1: Register for the main data portals

Create free accounts (you will use these throughout the course):

Store any credentials with your password manager, never in the repo.

Step 2: Build a sensor cheat sheet

In eo-portfolio, add docs/sensors.md with a table you will actually reuse:

Mission Type Key bands Spatial Revisit Best for
Sentinel-2 passive optical visible, NIR, SWIR ~10 m frequent burn severity, vegetation
Landsat passive optical visible, NIR, SWIR, thermal ~30 m ~16 days long-term change
VIIRS passive optical/thermal thermal ~375 m sub-daily active fire
Sentinel-1 active radar (C-band) VV, VH ~10 to 20 m frequent all-weather change, flood
GEDI active lidar n/a footprint sparse tracks canopy height, biomass

Fill it in from the lesson and the mission docs, and add a one-line "when I would reach for this" note per row. This becomes a reference you use in every later module.

Step 3: Pull your first scene

Using your Module 0.1 toolkit, search a STAC catalog for a Sentinel-2 scene over your study-area fire AOI and describe it. This is a read-and-inspect exercise, not processing yet:

import pystac_client, planetary_computer

catalog = pystac_client.Client.open(
    "https://planetarycomputer.microsoft.com/api/stac/v1",
    modifier=planetary_computer.sign_inplace,
)
# a small bbox over your fire AOI (lon/lat): west, south, east, north
bbox = [-118.6, 34.0, -118.3, 34.2]
search = catalog.search(
    collections=["sentinel-2-l2a"],
    bbox=bbox,
    datetime="2025-01-01/2025-03-31",
    query={"eo:cloud_cover": {"lt": 20}},
)
items = list(search.items())
print("scenes found:", len(items))
best = min(items, key=lambda it: it.properties["eo:cloud_cover"])
print("id:", best.id)
print("date:", best.properties["datetime"])
print("cloud cover:", best.properties["eo:cloud_cover"])
print("bands:", sorted(best.assets.keys())[:12])

If Planetary Computer is unavailable to you, run the same search against the Copernicus Data Space STAC API. The point is to prove you can find the right scene and read its metadata (date, cloud cover, available bands). Save this as notebooks/01_first_scene.ipynb.

Step 4: Commit

git add docs/sensors.md notebooks/01_first_scene.ipynb
git commit -m "Module 0.2: sensor cheat sheet and first Sentinel-2 scene lookup"
git push

Checkpoint

Self-check (answers below).

  1. Why can Sentinel-1 image a flood at night through thick cloud when Sentinel-2 cannot?
  2. Your task is to detect new fires within hours across a whole country. Which resolution do you prioritize, and which sensor family fits?
  3. Why do burn indices rely on the SWIR band?
  4. You downloaded a Level 1 top-of-atmosphere product and your multi-date comparison looks noisy. What did you likely skip?

Interview-style questions (practice out loud).

Answers. (1) Sentinel-1 is active radar using long microwaves that penetrate cloud and need no sunlight, while Sentinel-2 is passive optical that depends on reflected sunlight and is blocked by cloud.
(2) Prioritize temporal resolution (frequent revisit); the coarse but sub-daily VIIRS/MODIS family fits, accepting large pixels for speed.
(3) SWIR is sensitive to moisture and to char and ash, so it drops sharply where vegetation burned, making the pre-minus-post difference a strong burn signal.
(4) You likely skipped surface-reflectance correction and the cloud/quality mask; use a Level 2 product and apply its mask so dates are comparable.

Deliverable

A committed docs/sensors.md cheat sheet and a notebooks/01_first_scene.ipynb that finds and describes one low-cloud Sentinel-2 scene over your study-area AOI (date, cloud cover, bands). This is the reference and the first data touch the rest of the course builds on.

What a hiring manager sees

Anyone can run a model on a dataset someone handed them. Being able to reason about sensor and resolution trade-offs, and to justify why you chose a data source for a given question, is what distinguishes an EO engineer from a general ML practitioner. A tidy sensor cheat sheet and a clean scene-lookup notebook show that judgment early.

Currency note

The concepts here (passive versus active, the spectrum, the four resolutions, processing levels) are stable. The wavelength and frequency ranges use the IEEE radar band designations (X 8 to 12 GHz, C 4 to 8 GHz, L 1 to 2 GHz, P 0.3 to 1 GHz) and the published centre frequencies of Sentinel-1 (5.405 GHz), NISAR L-band (1.257 GHz), and ESA Biomass (435 MHz), verified August 2026; the electromagnetic-spectrum figure is an original diagram (no third-party image). Mission specifics such as exact revisit, band lists, and which sensors are currently operational do drift as satellites launch and retire, so verify a mission's current status and characteristics from its official documentation at study time, and confirm the STAC endpoints and collection names before running the lab.

Previous0.1 Your Geospatial Python ToolkitNext0.3 Geospatial Data 101