Space lasers to the rescue: using lidar to map crop types

5 min readNov 8, 2021

Chances are if you’ve heard about space lasers in the past year, it’s related to conspiracy theories about what was causing so many fires in California. While that’s obviously crazy (to put it kindly), the truth is that there actually are space lasers, and they actually have been shaking things up lately. In a good way.

A couple of years ago, a laser instrument called the Global Ecosystem Dynamics Investigation (GEDI) was installed on the International Space Station (ISS) and began taking measurements of the Earth’s land surface. Like some other lidar (Light Detection and Ranging) systems, GEDI was designed to improve understanding of forest systems. Lidar works by measuring the time it takes photons beamed out from the instrument to return to the sensor, which relates to the height distribution of the biomass on the ground, as well as the height of the ground itself. The image below, taken from the GEDI website, shows an example of how the GEDI returns vary along a forest transect, due to changes in the vertical distribution of vegetation..


Seeing the rapid progress being made for forest systems made us wonder: could GEDI also help us solve a longstanding challenge in agricultural mapping? Specifically, in many regions we lack geo-located ground data on where different crop types are located. This has limited the ability to map crop types with satellite data, because methods typically need a lot of local ground training data to work well. And the lack of good crop type maps then limits our ability to do lots of other related stuff, like forecast total crop production for a region, or map and investigate yield variation for specific crops.

How would GEDI help? Although all crops are short when compared to the forests that GEDI was meant to study, one crop in particular is typically much taller than the others: maize (also known as corn in the U.S.). No matter where you go in the world, a fully developed maize crop is typically much taller than a person, while other common crops like wheat or rice or beans will be about waist high. That’s why crop mazes are almost always made from corn, and not other crops — although apparently some soy mazes do exist, for those who scare easily and want to see over the top.

An example of corn (tall) next to soybean (short) (Photo from

While lidar sensors like GEDI offer some unique measurements, they typically only measure a small fraction of the land surface. In the case of GEDI, it acquires data spaced roughly 60m along a track, with 600m between tracks. The figure below shows tracks for one area in Iowa, along with the returns for each shot within that track (in (b), colored by crop type).

In a recent paper, we set out to answer a couple of questions. First, in the agricultural fields where GEDI observations have been made, do the GEDI data easily distinguish between corn and other crops? And if so, can we use GEDI measurements at these locations as a substitute for ground data when we train models to estimate crop type? In answering these, we wanted to be sure not to just look at one area (like in the figure above) but to test in a variety of systems. We settled on regions within three countries that each had a reliable independent map of crop type for the 2019 growing season: China, France, and the United States.

The paper provides the detailed analysis, but here is the brief summary. First, GEDI seems to have little trouble separating maize from other common crops in the area. The figure below shows the average profiles of GEDI by crop type in each region, with the shading showing the 25th-75th percentile. Notice how clear the separation is at the low and high ends.

Second, using GEDI predictions of whether a pixel is a tall or short crop is a pretty useful way of training a model using Sentinel-2 to map these classes for the entire region. The figure below compares accuracy in each region using the GEDI data to train vs. using the ground truth to train, with both models tested on the same test dataset. In all cases, GEDI-trained models (blue bars) do much better than models trained on ground labels in other regions (gray bars). The GEDI models don’t quite reach the accuracy of models trained on local ground truth (shown as dashed horizontal line), but they get pretty close.

We can then use these models to map crop class (“maize” (tall) vs. “other” (short)) for the whole region, which compare pretty well to the ground truth. The figure below compares the local “ground truth” (on the top row) row to maps we obtain using GEDI shots to train a model that predicts crop class based on Sentinel 2 (on the bottom row).

Based on these results, we are excited to see whether this works in other regions, with the hope being that we can begin mapping maize more accurately even in areas or years where we don’t have any field data. GEDI also will likely be useful for lots of other agricultural applications, some of which we discuss in the paper. Hopefully space lasers will be here to stay.

About the Authors:

David Lobell

David Lobell, Professor of Earth System Science, William Wrigley Senior Fellow at the Freeman Spogli Institute, Senior Fellow at the Woods Institute for the Environment and at the Stanford Institute for Economic Policy Research

Sherrie Wang

Sherrie Wang, postdoc at University of California, Berkeley

Stefania Di Tommaso

Stefania Di Tommaso is a Data Research Analyst at Center on Food Security and the Environment (FSE)




Stanford's Center on Food Security and the Environment (FSE) leads cutting-edge research on global issues of food, hunger, poverty and the environment.