26 · Satellite I: on your machine

Tuesday, Dec 1, 2026

Materials for this session are not published yet. They appear here before class.

Objectives

By the end of this session you can:

• Say what each sensor type records and which questions it can answer.
• Match a sensor’s spatial, temporal and spectral resolution to a research question.
• Compute NDVI from two bands and say what it proxies.
• Choose between a ready-made product and classifying your own, and measure the accuracy of a map you classified.
• Download one scene and process it in R on your own machine.

What we cover

• Remote sensing (measuring the Earth’s surface from a distance, here from satellites) in economics:
  • Henderson, Storeygard & Weil (AER, 2012) measured economic growth from night lights seen from space.
  • Burchfield et al. (QJE, 2006) measured urban sprawl across the US from land use mapped on 30-metre cells.
  • Jack et al. (AER: Insights, 2025) detected crop-residue burning on farms in a randomized trial in Punjab, India, with a classifier trained on field-labelled pixels.
• What each sensor records. A band is one range of light wavelengths, stored as its own grid of numbers.
  • Visible bands: a true-colour picture, close to what the eye sees.
  • Near-infrared, light just beyond red and invisible to the eye: healthy leaves reflect much of it.
  • Thermal infrared: heat given off by the surface, converted into surface temperature.
  • Radar: the sensor sends its own signal and records the echo, which shows surface roughness and moisture through cloud and at night.
  • Night lights: artificial light after dark, a stand-in for local economic activity.
  • Solar-induced fluorescence: the faint glow plants give off while using sunlight to grow, which tracks plant activity more closely than greenness does.
• Three resolutions decide what a sensor can answer. Spatial resolution is the ground size of one pixel, such as 10 by 10 metres. Temporal resolution, or revisit time, is how often the satellite images the same place. Spectral resolution is how many bands it records and how narrow each one is. A sharper pixel usually comes with a longer wait between images, so confirm the pixel is small enough to resolve the unit you study, such as a field or a village.
• Band algebra: NDVI (near-infrared minus red, divided by their sum, with values near 1 for dense green plants), and related indices for water and burned area. Read the index at a forest, a parking lot, a lake and an irrigated field, where you know roughly what to expect.
• Land-cover products (ready-made maps labelling each pixel as forest, water, cropland, built-up land and so on), burned area and crop type, against supervised classification (training a model on pixels you labelled by hand so it labels all the others). NAIP’s 1-metre aerial imagery for the US helps with the labelling. Hold back part of your ground truth (labels collected on the ground or read from sharper imagery) to verify the map: measure its accuracy on those pixels, and compare your classes with a ready-made land-cover map over the same area.
• One scene (one image of a fixed patch of ground on one date) end to end: find it, download it, read the bands with terra, apply the cloud mask (a layer marking pixels hidden by cloud or shadow), crop to your area and compute the index.
• Big downloads: read the file listing, loop through download, extract and delete so the disk survives, cache the large intermediate with saveRDS(), and have the script beep when an hour-long job finishes.