Land cover classification using convolutional neural network with remote sensing data and digital surface model

Abstract

Land cover map is widely used in urban planning, environmental monitoring and monitoring of the changing world. This paper proposes a framework with convolutional neural network (CNN), object-based voting and conditional random field (CRF) for land cover classification. Both very-high-resolution (VHR) remote sensing images and digital surface model (DSM) are inputs of this CNN model. To solve the "salt and pepper" effect caused by pixel-based classification, an object-based voting classification is performed. And to capture accurate boundary of ground objects, a CRF optimization using spectral information, DSM and deep features extracted through CNN is applied. Area one of Vaihingen datasets is used for experiment. The experimental results show that method proposed in this paper achieve an overall accuracy of 95.57%, which demonstrate the effectiveness of proposed method.