Advancing Methods for Estimating Cropland Area

LeeAnn King, University of Maryland College Park, mkinglee@umd.edu (Presenter)
Matthew Hansen, University of Maryland College Park, mhansen@umd.edu
Peter Potapov, University of Maryland, College Park, peter.potapov@hermes.geog.umd.edu
Alexander Krylov, University of Maryland, College Park, amkrylov@umd.edu
Xiao-Peng Song, University of Maryland, College Park, xpsongrs@gmail.com
Stephen Stehman, University of Maryland, College Park, svstehma@syr.edu
Bernard Adusei, University of Maryland, College Park, benadusei@yahoo.com

Remote sensing technology can provide timely, precise, and objective information about land dynamics to accurately quantify the area and extent of the land under cultivation for different crops and, in turn, advance methods for understanding crop production. This paper describes a new approach for crop-specific area estimation, relying on freely available remote sensing imagery and hierarchical, non-linear classification trees to describe soybean area and extent for a MODIS-stratified (Chang et al 2007) random-sample population. Each model has been applied to individual sample blocks and per total block populations, akin to local and national scale mapping in the United States. The primary objective of this study is to evaluate the agreement of Landsat derived classification and the CDL. The National Agricultural Statistics Service’s Cropland Data Layer (CDL) has been used for a reference data set to assess model performance for all datasets, estimating 28.4 Mha of soybean in the study area. Results of this analysis show that MODIS characterization is not appropriate for estimating the area of land under cultivation for soybean when using a sampling approach, underestimating soybean by nearly 50% at 18.4 Mha. Model results and final estimates were highly accurate when a combination of single date images were used for classification at the local scale, estimating 96.1% of the CDL at 27.3 Mha, and demonstrating that Landsat data is appropriate for estimating soybean extent in the United States using a sampling approach. Landsat metrics performed nearly perfectly when applied locally, describing 97% of the CDL at 27.5Mha. When national models were executed, employing Landsat derived metrics, results show a strong agreement with the CDL soybean area, capturing 90% when nationally applied, 98% of the CDL when applied to each stratum. This analysis proves that Landsat single-date time series and Landsat metrics can be used to accurately map soybean and estimate cultivated area in the United States, indicate the potential for testing metrics in other soybean-growing parts of the world, where robust reference data, such as the CDL, are lacking.

Presentation Type:  Poster

Session:  General Contributions   (Tue 4:35 PM)

Associated Project(s): 

• Hansen, Matt: Advancing methods for global crop area estimation ...details

Poster Location ID: 166