Integrating cloud-based workflows in continental-scale cropland extent classification.

Abstract

Accurate information on cropland spatial distribution is required for global-scale assessments and agricultural land use policies. Cloud computing platforms such as Google Earth Engine (GEE) provide unprecedented opportunities for large-scale classifications of Landsat data. We developed a novel method to fuse pixel-based random forest classification of continental-scale Landsat data on GEE and an object-based segmentation approach known as recursive hierarchical segmentation (RHSeg). Using our fusion method, we produced a continental-scale cropland extent map for North America at 30 m spatial resolution for the nominal year 2010. The total cropland area for North America was estimated at 275.18 million hectares (Mha). The overall accuracies of the map are >90% across the continent. This map also compares well with the United States Department of Agriculture (USDA) cropland data layer (CDL), Agriculture and Agri-food Canada (AAFC) annual crop inventory (ACI), and the Mexican government agency Servicio de Información Agroalimentaria y Pesquera (SIAP)'s agricultural boundaries. Furthermore, our map compared well with sub-country statistics including state-wise and county-wise cropland statistics in regression models resulting in R2 > 0.84. This key contribution paves the way for more detailed products such as crop intensity, crop type, and crop irrigation, and provides a method for creating high-resolution cropland extent maps for other countries where spatial information about croplands are not as prevalent.

Department

Natural Resources and the Environment

Publication Date

6-15-2018

Journal Title

Remote Sensing of Environment

Publisher

Elsevier Inc.

Digital Object Identifier (DOI)

10.1016/j.rse.2018.10.013

Document Type

Article

Rights

© 2018 Elsevier Inc.

Link to Full Text

Share

COinS