Combining experiments and satellite observations to measure yield benefits from crop rotation

StanfordFoodSecurity
6 min readJun 2, 2022

--

Population growth and the unprecedented rate of climate change create a need for rapid agricultural adaptation. Therefore, it is important to answer questions about which agricultural practices are most effective for growing crops under various weather conditions. There are two main ways that researchers try to answer such questions, but as we will discuss below, neither approach is perfect.

Two approaches: Randomized Field Experiments and Observational Datasets

Traditionally, agricultural scientists have used randomized field experiments to determine which farm management practices cause higher crop yields. Randomized field experiments are considered the “gold standard” approach because the random treatment assignment allows researchers to reliably assess whether the treatment caused the observed change in the crop yields. Randomized field experiments helped drive the increases in crop yield over the past 150 years, and one could even argue that they are partially responsible for feeding billions of people!

The downside to randomized field experiments is that they suffer from external validity issues, meaning that their conclusions may not apply to farms with different climate, soil, or management conditions than those of the experimental farms. For example, the conclusions from an experiment on a farm where the temperature never exceeded 30°C may not apply to a different farm in a warmer climate. The conclusions from such an experiment may not even apply to the same farm 10 years later, when the temperature is expected to exceed 30°C at least a few days per year.

Thanks to technological advances, a recent alternative to randomized field experiments is to use observational datasets generated by satellite-based remote sensing imagery (for example, this approach has recently been used to study tillage[1] and crop rotation [2]). Compared to randomized field experiments, satellite-based observational datasets have the advantage of including a wider array of growing conditions, such as extreme temperature or precipitation conditions that we might expect to see more frequently in the coming decades. Such datasets allow researchers to substantially reduce external validity issues.

Unfortunately, studies of agricultural practices using these large observational datasets suffer from internal validity issues, meaning that such studies cannot provide strong evidence for claims about the causal effect of an agricultural practice on yield. These issues arise because in such observational datasets, the agricultural practice is not randomly assigned, so the estimated effects of an agricultural practice on yield could be biased by unmeasured confounding variables. For example, if fertilizer use is not measured in a dataset, an observational study using such a dataset to estimate the effect of crop rotation on corn yield would underestimate the true effect of rotation on corn yield because farmers that rotate their crops are encouraged to use less fertilizer than farmers that grow corn every year.

Does combining data from randomized experiments with observational data help?

Randomized field experiments and observational studies have complementary strengths and weaknesses. Can combining data from randomized field experiments with observational data in the same analysis lead to estimates of the causal effect that are both internally and externally valid?

In our recent paper [3], published in Environmental Research Letters, we develop a hybrid method to combine randomized field experiments with observational data to assess the effects of farm management practices on crop yields. As a case study we focused on estimating the effects of the corn-soybean rotation on corn and soybean yields in the Midwestern United States from 2000–2018.

We used open-source data from 11 experiments in which crop rotations were randomized. We coupled this with a satellite-based observational dataset spanning the study region that included satellite-based estimates for crop yield and crop rotation. To implement our hybrid method, we first estimated the benefit of rotation on yield at each location and year in our study region and period using only the observational data. This included computing estimates of the rotation benefit at the locations of each experimental site, without using any of the experimental data (besides the latitude and longitude of the sites). We then computed the benefit of crop rotation at each site and year within the experimental data, without using any of the observational data. Below is a scatter plot of the estimates of the benefit of the soybean-corn rotation on corn yield (in metric tons per hectare) using only satellite-based observational data on the x-axis versus those using only the experimental data on the y-axis.

The experimental estimates and observational estimates have quite a different range, but they do have a statistically significant positive correlation! We used the pairs of estimates in the scatter plot to fit a calibration model that can be used to produce estimates of rotation benefit on corn yields at any location and year for which the observational estimates exist.

In the task of predicting the rotation benefit in held-out experiments, based on cross-validation, our hybrid method performed better than using only the experimental data or using only the observational data. Our hybrid method can also be used to study how the benefit of rotation varies based on weather and geography.

Does the benefit of crop rotation vary based on weather and geography?

Below is a visualization of how our estimates based on the hybrid method vary across geographical locations (the 2–3 character abbreviations on the maps indicates the locations of the experimental sites that were used for our calibration, and the units are in metric tons per hectare). The estimated effect of rotation on both corn and soybean yield is positive throughout the study region. This is no surprise given that it has been known for centuries that crop rotation tends to improve crop yields. We also see that the benefit of rotation on corn yield tends to be highest in the northwest parts of the study region while the benefit of rotation on soybean yields tends to be highest in the southern and central parts of the study region.

We also looked for associations between temperature and rotation benefits. A main finding of our work was that the estimated rotation benefits for corn yields were lower at high temperatures whereas the estimated rotation benefits for soybean were higher at high temperatures. It turns out this data-driven insight has biological and chemical explanations, relating to pest pressure and nitrogen availability in the soil (see the Discussion section of our paper [3] for more details). Our findings suggest two points: first, as the climate warms crop rotation may become a less beneficial management practice for corn yields, though crop rotation will remain a net beneficial practice. Second, crop rotation will become an increasingly important management practice for maximizing soybean yields.

References

[1]: Deines J.M., Wang S., and Lobell D.B., 2019. Satellites reveal a small positive yield effect from conservation tillage across the US Corn Belt. Environmental Research Letters. 14 124038.

[2]: Cohen A. A. B., Seifert C.A. , Azzari G., and Lobell D.B., 2019. Rotation effects on corn and soybean yield inferred from satellite and field-level data Agronomy Journal: 111 (6) 2940–2948

[3]: Kluger D.M., Owen A.B., and Lobell D.B., 2022. Combining randomized field experiments with observational satellite data to assess the benefits of crop rotations on yields. Environmental Research Letters. 17 044066.

About the Author: Dan Kluger is a Ph.D. candidate in the Department of Statistics at Stanford.

Dan Kluger

--

--

StanfordFoodSecurity
StanfordFoodSecurity

Written by StanfordFoodSecurity

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

Responses (1)