The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLII-2/W13
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W13, 681–685, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W13-681-2019
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLII-2/W13, 681–685, 2019
https://doi.org/10.5194/isprs-archives-XLII-2-W13-681-2019

  04 Jun 2019

04 Jun 2019

SUPERPIXEL CLASSIFICATION OF HIGH SPATIAL RESOLUTION REMOTE SENSING IMAGE BASED ON MULTI-SCALE CNN AND SCALE PARAMETER ESTIMATION

Y. Chen and D. Ming Y. Chen and D. Ming
  • School of Information Engineering, China University of Geosciences (Beijing), 100083, Beijing, China

Keywords: Deep Learning, Spatial Statistics, High Spatial Resolution Remote Sensing Image, Image Segmentation, OBIA

Abstract. In recent years, considerable attention has been paid to integrate convolutional neural network (CNN) with land cover classification of high spatial resolution remote sensing image. Per-pixel classification method based on CNN (Per-pixel CNN) achieved higher accuracy with the help of high-level features, however, this method still has limitations. Even though per-superpixel classification method based on CNN (Per-superpixel CNN) overcome the limitations of per-pixel CNN, classification accuracy of complex urban is easily influenced by scale effect. To solve this issue, superpixel classification method combining multi-scale CNN (Per-superpixel MCNN) method is proposed. Besides, this paper proposes a novel spatial statistics based method to estimate applicable scale parameter of per-superpixel CNN. Experiments using proposed method were performed on Digital Orthophoto Quarer Quad (DOQQ) images in urban and suburban area. Classification 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.