An Ensemble-Based Training Data Refinement for Automatic Crop Discrimination Using WorldView-2 Imagery

Abstract

This paper presents a new approach for refining and selecting training data for satellite imagery-based crop discrimination. The goal of this approach is to automate the pixel-based “multievidence crop classification approach,” proposed by the authors in their previous research. The present study is used to feed this classification approach with automatically selected training samples based on available vector data (agricultural parcels representing crop boundaries with crop codes). The vector data are created by farmers to support subsidy claims and are, therefore, prone to errors such as boundary digitization mismatch and mislabeling of crop codes. The proposed approach, ensemble-based cluster refinement approach (ECRA), refines the declared crop clusters in an iterative training-classification scheme and provides potential training samples that give correct class descriptions. ECRA operates based on two assumptions: 1) mislabels in each class will be far from their cluster centroid and 2) each crop class based on the available vector data has more correctly labeled samples than mislabeled samples. Three datasets, derived from bitemporal WorldView-2 multispectral imagery, are used in an ensemble framework to iteratively refine the samples from crop clusters declared by the farmers: spectral, texture, and vegetation indices. Once the refinement of clusters is complete, final training samples are identified. They are used for learning and the satellite imagery is classified using the multievidence classification approach. The study is implemented with WorldView-2 imagery acquired for a study area in Denmark containing 15 crop classes. The multievidence classification approach with ECRA-based refinement is compared with the classification based on common training sample selection methods (manual and random). It is also compared with the winner-takes-all-based classification approach. ECRA achieves an overall classification accuracy of 92.8%, which is higher than existing common approaches.