Hyperspectral Image Classification Based on Multilayer Perceptron Trained with Eigenvalue Decay


Lokman G., ÇELİK H. H., TOPUZ V.

CANADIAN JOURNAL OF REMOTE SENSING, cilt.46, sa.3, ss.253-271, 2020 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 46 Sayı: 3
  • Basım Tarihi: 2020
  • Doi Numarası: 10.1080/07038992.2020.1780572
  • Dergi Adı: CANADIAN JOURNAL OF REMOTE SENSING
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Applied Science & Technology Source, Aquatic Science & Fisheries Abstracts (ASFA), CAB Abstracts, Compendex, Computer & Applied Sciences, Environment Index, INSPEC, Pollution Abstracts, Civil Engineering Abstracts
  • Sayfa Sayıları: ss.253-271
  • Marmara Üniversitesi Adresli: Evet

Özet

Hyperspectral Images (HSI) require sufficient labeled samples and a complex classifier to identify an area. Support Vector Machine (SVM) is one of the most competent algorithms in this field. Neural Networks (NN) is another approach used for classification problems, and both have been widely proposed in the literature. The Convolutional Neural Network (CNN) method has also received significant attention in the deep learning field recently. Nevertheless, during NN training, the overfitting problem may cause continuous dragging of the algorithm toward larger error. In this case, a regularization technique is needed to constitute the most useful decision boundary. The Eigenvalue Decay method is one of the regularization techniques that may be applied for HSI. This study investigates the performance of Multilayer Perceptron trained with an Eigenvalue Decay (MLP-ED) algorithm for HSI classification. The SVM, CNN with Pixel-Pair and CNN-Ensemble methods are used as comparison algorithms for MLP-ED performance assessment. All methods were tested with 3 different high-resolution HSI datasets. While SVM is one of the classic classifiers, and the 2 new CNN algorithms show high performance, the proposed MLP-ED method has more computational efficiency and achieves higher success than the others do.