Superpixel based land cover classification of VHR satellite image combining multi-scale CNN and scale parameter estimation

Abstract

Traditional classification methods, which use low-level features, have failed to gain satisfactory classification results of very high spatial resolution (VHR) remote sensing images. Even though per-pixel classification method based on convolutional neural network (CNN) (Per-pixel CNN) achieved higher accuracy with the help of high-level features, this method still has limitations. Per-superpixel classification method based on CNN (Per-superpixel CNN) overcomes the limitations of per-pixel CNN, however, there are still some scale related issues in per-superpixel CNN needed to be explored and addressed. Firstly, in order to avoid the misclassification of complex land cover objects caused by scale effect, the per-superpixel classification method combining multi-scale CNN (Per-superpixel MCNN) is proposed. Secondly, this paper analyzes how scale parameter of CNN impacts the classification accuracy and involves spatial statistics to pre-estimate scale parameter in per-superpixel CNN. This paper takes two VHR remote sensing images as experimental data, and employs two superpixel segmentation algorithms to classify urban and suburban land covers. The experimental results show that per-superpixel MCNN can effectively avoid misclassification in complex urban area compared with per-superpixel classification method combining single-scale CNN (Per-superpixel SCNN). Series of classification results also show that using the pre-estimated scale parameter can guarantee high classification accuracy, thus arbitrary nature of scale estimation can be avoided to some extent. Additionally, through discussion of the influence of accuracy evaluation method in CNN classification, it is stressed that random selection of ground truth validation points from study area is recommended and more responsibly other than using part of a reference dataset.