Abstract: To extract ground content is the basis for a large number of geoscientific applications. Existing pixel-based classification methods do not fully exploit the contextual associations in multispectral remote sensing images, and fragmented labels are observed everywhere in classified images. In order to improve the classification accuracy of high-resolution multispectral images, this paper proposes a new method which is based on context-aware networks and super-pixel post-processing. The method designs a new convolutional neural network to learn the spatial contextual information in multispectral images. Super-pixel post-processing uses a strategy of small region segmentation and voting to merge structurally associated regions, which can eliminate fragmented labels. The new method is tested on the Gaofen-1 satellite data and compared with six classification algorithms. The experimental results show that the new method outperforms the competing algorithms in terms of accuracy and visual effect. Among them, the super-pixel post-processing can reduce the fragmentation of classification results as well as improve the classification accuracy.

Key words: image classification, land use, Gaofen-1