The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Download
Publications Copernicus
Download
Citation
Articles | Volume XLIII-B2-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 1217–1224, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-1217-2020
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLIII-B2-2020, 1217–1224, 2020
https://doi.org/10.5194/isprs-archives-XLIII-B2-2020-1217-2020

  14 Aug 2020

14 Aug 2020

AUTOMATED AND LIGHTWEIGHT NETWORK DESIGN VIA RANDOM SEARCH FOR REMOTE SENSING IMAGE SCENE CLASSIFICATION

J. Li1,2,3, W. Diao1,2, X. Sun1,2,3, Y. Feng1,2,3, W. Zhang1,2, Z. Chang1,2,3, and K. Fu1,2,3 J. Li et al.
  • 1Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
  • 2Key Laboratory of Network Information System Technology, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing, China
  • 3University of Chinese Academy of Sciences, Beijing, China

Keywords: Scene Classification, Random Search, Neural Architecture Search, Remote Sensing Image, Deep Learning, Semantic Segmentation

Abstract. With the development of deep learning, remote sensing image scene classification technology has been greatly improved. However, current deep networks used for scene classification usually introduce ingenious extra modules to fit the characteristics of remote sensing images. It causes a high labor cost and brings more parameters, which makes the network more complicated and poses new intractable problems. In this paper, we rethink this popular “add module” pattern and propose a more lightweight model, called ProbDenseNet (PDN). PDN is obtained via a random search strategy in Neural Architecture Search (NAS) which is an automated network design manner. In our method, all topological connections are assigned importance degrees which subject to a uniform distribution. And we set a regulator to adjust the sparsity of the network. By this way, the design procedure is more automated and the network structure becomes more lightweight. Experimental results on AID benchmark demonstrate that the proposed PDN model can achieve competitive performance even with much fewer parameters. And we also find that excessive connections do not always improve the network’s performance while they can drag down the network’s behavior as well. Furthermore, we conduct experiments on Vaihingen dataset with classical Fully Convolutional Network (FCN) framework. Quantitative and qualitative results both indicate that the features learned by PDN can also transfer in semantic segmentation task.