Land cover classification from very high spatial resolution images via multiscale object-driven CNNs and automatic annotation

Abstract

Very high spatial resolution (VHR) images are accessible for land-cover classification. Because of the complex information and deep features associated with increased spatial resolution, it is difficult to find an effective method for improving the classification accuracy of finer-resolution remote sensing images. We propose a scheme for sample selection and sample autoannotation to establish a new labeled benchmark using vector maps. In addition, a suitable multiscale object-driven convolutional neural networks (multi-OCNNs) were developed for the new benchmark. We trained multi-OCNNs to capture the deep and contextual information contained in samples of the VM-RSI benchmark to predict the land cover categories from target region images with no labels. Furthermore, an object-driven voting strategy applied in multi-OCNNs provides clear boundary information of ground objects and less noise in CNN predictions. Experiments on multisource VHR images, including SPOT-6, Gaofen-2, and ALOS satellite images, show that the proposed approach outperforms other object-based algorithms by achieving good performance results, verifying its feasibility, and effectiveness for land cover classification.