Date Published: October 11, 2018
Publisher: Public Library of Science
Author(s): Jingyu Song, Michael S. Delgado, Paul V. Preckel, Nelson B. Villoria, Nathan J. Moore.
Despite substantial research and policy interest in pixel level cropland allocation data, few sources are available that span a large geographic area. The data used for much of this research are derived from complex modeling techniques that may include model simulation and other data processing. We develop a transparent econometric framework that uses pixel level biophysical measurements and aggregate cropland statistics to develop pixel level cropland allocation predictions. Such pixel level land use data can be used to investigate the impact of human activities on the environment. Validation exercises show that our approach is effective at downscaling cropland allocation to multiple levels of resolution.
Agricultural productivity and environmental sustainability are central focuses for policymakers and academics. Understanding the interaction between agricultural systems and economic and environmental systems is critical for enhancing public policies related to economic development, food security, and human well-being. One important element to understand such interactions is the distribution of cropland and crop types within a particular area, particularly when the goal is to design policies that focus on promoting agricultural productivity while ensuring environmental sustainability [1–3].
We develop two empirical applications: a single-crop model for maize, representing the case in which one is interested only in whether land is dedicated to a crop or some other use, and a multi-crop model of maize, soybeans, and wheat, for the case in which one is interested in multiple crops. For both models, we focus on harvested crop area spanning North, Central and South America. The 5 arc-minute resolution is a commonly used pixel measurement [30, 31], and yields pixels of about 100 square kilometers at the equator, 60 square kilometers in Minnesota, and nearly zero square kilometers near the north/south pole. S1 Table in the supplementary section provides a comprehensive list of all data we employ, including the units of measurement and source.
We develop a statistical method for predicting pixel level cropland allocation across a (large) geographic area in which pixel-level measurements are not available. Specifically, we develop a fractional response model that combines measurements of pixel level land attributes with observable aggregate land use patterns to predict the share of cropland allocated to a certain crop at the pixel level. We formulate the likelihood function and demonstrate application to a single-crop model for maize and a multi-crop model for maize, soybeans, and wheat. We validate our estimates against other available spatially explicit cropland datasets and show that both the single-crop and multi-crop models at the Administrative Unit Levels 1 and 2 are reasonably precise in predicting cropland allocation at the pixel level.