Front Inner Page - Volume 3 No.5 October 2016

  • » Back to Index

  • Title

    :

    Deep Feature Extraction for Image Classification Using Remote Sensing

    Authors

    :

    Dr. A.R. Mohamed Shanavas1, M.A. Aysha Siddiqua2

    Keywords

    :

    Deep convolutional networks, deep learning, sparse features learning, feature extraction, aerial image classification, very high resolution (VHR), multispectral images, hyper-spectral image, classification, segmentation.

    Issue Date

    :

    October 2016

    Abstract

    :

    This paper presents the usage of particular layer and deep convolutional networks which are meant for remote sensing data analysis. For the given high input data dimensionality and the comparatively negligible amount of accessible labeled data, Direct use to multi- and hyper-spectral imagery of supervised (shallow or deep) convolutional networks remains significant. Hence, we put forward the technique of desirous layer-wise unsupervised pre-training united with an effective algorithm for sparse features in unsupervised learning. The algorithm is entrenched on sparse representations and implements lifetime sparsity and population of the extracted features, instantaneously. We effectively illustrate the expressive power of the mined representations in several scenarios: classification of aerial scenes, along with classification based on land very high resolution (VHR), or land-cover classification from multi- and hyper-spectral images. The proposed algorithm shows better performance than the kernel counterpart (kPCA) and Principal Component Analysis (PCA) .Results determine that the single layer convolutional networks can abstract influential discriminative structures only when the receptive field interpreted for adjacent pixels, and requires high resolution and detailed results. Though, deep architectures significantly outperforms single layers variants, taking cumulative levels of abstraction and complexity throughout the feature hierarchy.

    Page(s)

    :

    1-4

    ISSN

    :

    2347- 4734

    Source

    :

    Vol. 3, No.5, October 2016

    Download

    :


  • » Back index