Automated crop field extraction from multi-temporal Web Enabled Landsat Data

摘要

An automated computational methodology to extract agricultural crop fields from 30 m Web Enabled Landsat data (WELD) time series is presented. The results for three 150 × 150 kmWELD tiles encompassing rectangular, circular (center-pivot irrigation) and irregularly shaped fields in Texas, California and South Dakota are presented and compared to independent United States Department of Agriculture (USDA) National Agricultural Statistics Service (NASS) cropland data layer (CDL) classifications. Coherent fields that are visually apparent were extracted with relatively limited apparent errors of omission or commission compared to the CDL classifications. This is due to several factors. First, the use ofmulti-temporal Landsat data, as opposed to single Landsat acquisitions, that enables crop rotations and inter-annual variability in the state of the vegetation to be accommodated for and provides more opportunities for cloud-free, non-missing and atmospherically uncontaminated surface observations. Second, the adoption of an object-based approach, namely the variational region-based geometric active contour method that enables robust segmentation with only a small number of parameters and that requires no training data. Third, the use of a watershed algorithmto decompose connected segments belonging to multiple fields into coherent isolated field segments and a geometry-based algorithm to detect and associate parts of circular fields together. A preliminary validation is presented to gain quantitative insights into the field extraction accuracy and to prototype a validation protocol including new geometric measures that quantify the accuracy of individual field objects. Implications and recommendations for future research and large-area applications are discussed.