Crop classification using crop knowledge of the previous-year: Case study in Southwest Kansas, USA

Abstract

Crop-type distribution products of the previous-year were used to generate training samples in the classification year. For each pixel, if the frequency of one crop was higher than 50%, the pixel was assumed to be a “possible training sample” of the high-frequency crop. Next, features of the “possible samples” were compared with reference crop features, and matching “possible samples” were confirmed as training samples. The Crop Data Layer (CDL) in Southwest Kansas during 2006-2013 was used as the crop products and MODIS EVI time series were crop features; training samples in 2014 were then acquired. Most of these training samples had the same crop label as the 2014 CDL data, and the training samples achieved good classification accuracies.