University of Twente Proceedings

Login

Dropband: a convolutional neural network with data augmentation for scene classification of VHR satellite images

Share/Save/Bookmark

Yang, Naisen and Tang, Hong and Sun, Hongquan and Yang, Xin (2016) Dropband: a convolutional neural network with data augmentation for scene classification of VHR satellite images. In: GEOBIA 2016 : Solutions and Synergies., 14 September 2016 - 16 September 2016, University of Twente Faculty of Geo-Information and Earth Observation (ITC) .

[img]
Preview
PDF
498kB
Event: GEOBIA 2016 : Solutions and Synergies., 14 September 2016 - 16 September 2016, University of Twente Faculty of Geo-Information and Earth Observation (ITC)
Abstract:Data augmentation is a common method that can prevent the overfitting of classification tasks in deep neural networks. This paper presents another kind of data augmentation method called DropBand that is useful for remote sensing image classification. Data augmentation is usually used along two dimensions of the image plane. This method executes this operation in the third dimension formed by all the spectral bands of an input image. With dropping a band of images out, the error rate of deep neural networks can be reduced. This method can also be viewed as a peculiar version of deterministic Dropout. The normal Dropout does not work well when it is applied to input channels of neural networks. To release this issue, dropping a band of input by schedule is employed. Moreover, model synthesis plays a key role in this procedure. To exclude the influence of increasing parameters, extra comparison groups are set up. The final experimental result shows that deep neural networks indeed benefit from the method of DropBand. This method improves the state-of-the-art on the latest SAT-4 and SAT-6 benchmarks.
Item Type:Conference or Workshop Item (Paper)
Link to this item:https://doi.org/10.3990/2.403
Conference URL:https://www.geobia2016.com/
Export this item as:BibTeX
EndNote
HTML Citation
Reference Manager

 

Repository Staff Only: item control page