27 · Satellite II: Earth Engine

Thursday, Dec 3, 2026

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

Objectives

By the end of this session you can:

• Say when to send the computation to the data instead of downloading the data to your machine.
• Find a dataset in the catalog and read its documentation.
• Filter an image collection by date, place and cloud cover.
• Reduce a collection over your own polygons and load the table in R.
• Estimate golf-course water use from evapotranspiration data and compare it with the water budgets an agency publishes.

What we cover

• Satellite archives in economics:
  • Burgess et al. (QJE, 2012) built yearly deforestation data for Indonesia from satellite imagery and found that splitting districts raised deforestation.
  • Assunção, Gandour & Rocha (AEJ: Applied, 2023) used Brazil’s satellite deforestation alerts, with clouds that block detection as a source of variation in enforcement, to show that monitoring curbs clearing in the Amazon.
  • Garg, Jagnani & Pullabhotla (AEJ: Policy, 2024) counted satellite fire detections within 10 km of about 11,000 Indian villages and found that new rural roads raised crop burning and downwind infant mortality.
• Sending the computation to the data. Some questions need more pixels than your disk holds, such as a decade of images over a continent. Google Earth Engine, a Google service that holds satellite archives, runs your code on the computers that store the data (server-side computation), and only the result comes back.
• The data catalog, the searchable list of datasets the service holds: Landsat, Sentinel, MODIS and derived products. Reading a documentation page for units, cloud masks and the scale factor, the number you multiply stored values by to get real units.
• Filtering an image collection (a stack of images of one kind, such as every Landsat image of a region) by date range, bounding box and cloud cover. Print the number of images left and compare it with the count you expect from how often the satellite passes over.
• Server-side thinking. A loop that works in R hangs in Earth Engine, because the work runs on remote machines. You map a function over the collection (apply it to every image) and then apply a reducer, which collapses many values into one, such as a mean. Each call takes a scale, the pixel size in metres at which the calculation runs.
• reduceRegions(). Zonal statistics over your own polygons (closed shapes on a map, such as fields or counties) summarise the pixel values inside each one, such as mean rainfall per county or greenness per field. Export the table as a file, load it in R as a data frame, and verify it by reducing one small polygon with terra and comparing the two numbers.
• Derived products and where each is weakest: land-cover maps, and evapotranspiration (water leaving the land as vapour, from soil evaporation plus water released by plants) from OpenET, a public dataset estimated from satellite images at 30-metre pixels across the contiguous US.
• Application: golf-course water use under drought restrictions. When a drought brings limits on outdoor watering, do golf courses cut back? Golf-course polygons combined with OpenET evapotranspiration. Compare each course with the land around it to see whether it is watered, then convert the difference to consumptive use (water that evaporates or is taken up by plants and does not return to rivers or aquifers) in acre-feet per year (an acre-foot covers one acre one foot deep). Reconcile each estimate with the water budget, the yearly amount a regional agency says the course needs or may use.
• When to stay local. Small areas, one scene, or work you must rerun exactly. Cache the exported table and commit it.
• Access: register for noncommercial Earth Engine use at least a week before class, since approval can take a few days. A setup guide with a first test script comes with the session materials.