Hyperspectral Classification Based on Lightweight 3-D-CNN With Transfer Learning

Awduron Sefydliadau
  • Haokui Zhang(Awdur)
    Northwestern Polytechnical University
  • Ying Li(Awdur)
    Northwestern Polytechnical University
  • Yenan Jiang(Awdur)
    Northwestern Polytechnical University
  • Peng Wang(Awdur)
    Northwestern Polytechnical University
  • Qiang Shen(Awdur)
  • Chunhua Shen(Awdur)
    University of Adelaide
Math Erthygl
Iaith wreiddiolSaesneg
Tudalennau (o-i)5813-5828
Nifer y tudalennau16
CyfnodolynIEEE Transactions on Geoscience and Remote Sensing
Rhif y cyfnodolyn8
Dyddiad ar-lein cynnar11 Ebr 2019
Dangosyddion eitem ddigidol (DOIs)
StatwsCyhoeddwyd - 31 Awst 2019
Cysylltiad parhaol
Arddangos ystadegau lawrlwytho
Gweld graff cysylltiadau
Fformatau enwi


Recently, hyperspectral image (HSI) classification approaches based on deep learning (DL) models have been proposed and shown promising performance. However, due to very limited available training samples and massive model parameters, deep learning methods may suffer from over-fitting. In this
paper, we propose an end-to-end 3D lightweight convolutional neural network (CNN) (abbreviated as 3D-LWNet) for limited samples based HSI classification. Compared with conventional 3D-CNN models, the proposed 3D-LWNet has deeper network structure, less parameters and lower computation cost, while it
results in better classification performance. To further alleviate the small sample problem, we also propose two transfer learning strategies: 1) cross-sensor strategy. We pretrain a 3D model in the source HSI datasets containing a greater number of labeled samples and then transfer it to the target HSI datasets; and
2) cross-modal strategy. We pretrain a 3D model in the 2D RGB image datasets containing a large number of samples and then transfer it to the target HSI datasets. In contrast to previous approaches, we do not impose restrictions over the source datasets in that they do not have to be collected by the same senors as the target datasets. Experiments on three public HSI datasets captured by different sensors demonstrate that our model achieves competitive performance for HSI classification compared to several state-of-the-art methods