Convolutional neural network extreme learning machine for effective classification of hyperspectral images

Abstract

Due to its excellent performance in terms of fast implementation, strong generalization capability and straightforward solution, extreme learning machine (ELM) has attracted increasingly attentions in pattern recognition such as face recognition and hyperspectral image (HSI) classification. However, the performance of ELM for HSI classification remains a challenging problem especially in effective extraction of the featured information from the massive volume of data. To this end, we propose in this paper a new method to combine convolutional neural network (CNN) with ELM (CNN-ELM) for HSI classification. As CNN has been successfully applied for feature extraction in different applications, the combined CNN-ELM approach aims to take advantages of these two techniques for improved classification of HSI. By preserving the spatial features whilst reconstructing the spectral features of HSI, the proposed CNN-ELM method can significantly improve the accuracy of HSI classification without increasing the computational complexity. Comprehensive experiments using three publicly available HSI datasets, Pavia University, Pavia center, and Salinas have fully validated the improved performance of the proposed method when benchmarking with several state-of-the-art approaches.

Convolutional Extreme Learning Machines: A Systematic Review

“…Different from the previous work, other authors propose their architecture instead of using a known network for the transfer learning, such as [49], [46], [64], [102], [103], [116] that propose CELM architectures with a different number of convolutional and pooling layers. The authors use CNN architectures for training the data with the fully connected layers.…”

“…All these architectures presented better training results than classic machine learning models. [51], [48], [54], [50], [49], [46], and [47] used the Pavia dataset for remote sensing classification. Note that remote sensing approaches use another evaluation metrics such as average accuracy (AA), overall accuracy (OA), and Kappa, as shown in Table 14.…”

“…The most common approach used to this end is a CNN previously pre-trained in the Pavia dataset used for feature extraction and an ELM for classification task [50], [49], [46]. However, the ELM-HLRF proposed in [51] achieved the best AA and OA results, respectively 98.25% and 98.36%.…”

Local Block Multilayer Sparse Extreme Learning Machine for Effective Feature Extraction and Classification of Hyperspectral Images

“…6.1 Set t=0. Although LBMSELM can improve the classification accuracy of conventional ELM, the classification results can be further refined by utilizing the spectral and spatial information [45] of HSIs. Given the output of the proposed LBMSELM, we transform it to the following equation:…”

Land-use classification via constrained extreme learning classifier based on cascaded deep convolutional neural networks

“…Weng (Weng et al, 2017) took a deep convolutional neural network (CNN) similar to AlexNet as a feature extractor, and replaced the fully connected layers of CNN with an ELM classifier to obtain higher classification accuracy than other methods. Cao (F. Cao et al, 2018) applied the CNN-ELM architecture to hyperspectral image classification, which can preserve the spatial feature and reconstruct the spectral features of hyperspectral image, and improve the classification accuracy greatly. Kannojia (Kannojia & Jaiswal, 2018) proposed an ensemble of hybrid CNN-ELM model for image classification, which took advantage of three parallel hybrid CNN-ELMs model for feature extraction and classifier and obtained the final classification result by majority voting ensemble of the three outputs.…”