Improving Land Cover Classification With a Shift-Invariant Center-Focusing Convolutional Neural Network

Abstract

Convolutional neural networks (CNNs) are widely employedin remote sensing community. The CNN-based, also knownas patch-based land cover classification method has gained in-creasing attention. However, this method very often requiresthe aid of post-processing, otherwise it is difficult to obtainaccurate boundaries separating different land cover classes.In this paper, we discuss the reason of this phenomenon andpropose a shift-invariant center-focusing (SICF) network todeliver more accurate boundaries to improve the patch-basedland cover classification. The principle of SICF is calculat-ing the class score from a center-focusing area based on ashift-invariant feature extraction module to calibrate predic-tion. We employ three modern CNNs to build correspond-ing SICF networks, the evaluation results indicate that com-pared with the conventional CNNs, the improvements madeby SICF for delivering accurate boundaries in land cover clas-sification are significant.