Farmland parcel boundary extraction based on local feature extraction and sparse representation from remote sensing images

Abstract

ABSTRACT Accurate information on agricultural field boundaries is important for precision agriculture. Contour detection combining local cues presents a high performance on nature images. Image sparse representation describes an image is reconstructed by using as few basic functions as possible. The number of farmland parcel boundaries is small and unbalanced for the whole agricultural fields. It fits the application category of sparse representation. In this research, we investigate an approach based on contour detection and sparse representation for the extraction of farmland parcel boundaries. First, field boundaries have an obvious brightness contrast with the internal parts of the farmland parcels. We capture the cue to describe per pixel. Then, we use efficient sparse coding algorithm to represent every pixel for boundary determination. Experimental results show that the proposed method achieves a sensitivity, specificity, accuracy, F 1 and AUC of 0.6089, 0.9055, 0.8865, 0.4073 and 0.7552, respectively. The purpose of this paper is to demonstrate the potential of combining local features with sparse representation for a fast and accurate farmland parcel boundary extraction approach from remote sensing images. Comparison results with existing methods on two datasets demonstrate that the proposed method is able for accurate discrimination of the farmland parcel boundaries.