Uncertainty in Cropland Data layer derived land-use change estimates: puting corn and soy expansion estimates in context

Abstract

Increased demand for corn for ethanol and the subsequent record high commodity prices has resulted in rapid expansion of corn and soy in the Midwestern U.S. Whether or not this expansion is replacing existing agricultural production or is expanding onto previously uncultivated grass or pasture land has profound implications for ecosystem services such as soil carbon storage, soil erosion prevention, and water quality. Several studies have used the Cropland Data Layer (CDL) to track fine scale land-use changes driven by corn and soy. However, these studies rarely account for the variability in data quality throughout the CDL's history. Here I compare established techniques as well as the application of USDA's Common Land Unit (CLU) data for removing `noise' from change rasters and quantifying the land covers lost to corn and soy expansion. I compare these estimates to equivalent measures from the National Agricultural Statistics Services (NASS) and use the discrepancy between them to identify spatial and temporal variability in CDL data that could influence land-use change study results. The CLU results differed little from established techniques, but both improved over direct comparisons of CDLs. Comparison to NASS data revealed pre-2010 versions of the CDL underestimate corn and soy area much more than later versions, leading to the detection of illusory land-use change when they are compared to post-2010 versions. Spatial and temporal variability resulted in errors that were several times larger than the trends the data are being used to detect. According to the CDL, approximately five million hectares of corn and soy expanded onto the grass, pasture, hay, and wheat between 2007 and 2012. However, over the same time period the CDL overestimated the amount of corn and soy expansion by 3.5 million hectares in the unmodified treatment and 1.5 million with cleaning methods applied. This work suggests that studies that use the CDL should test for and report variability and uncertainty in their results.