Adaptive Residual Convolutional Neural Network for Hyperspectral Image Classification

Abstract

In this article, we designed an adaptive residual convolutional neural network (ARCNN) that takes raw hyperspectral image (HSI) cubes as input data for land-cover classification. In this network, spectral and spatial feature extraction blocks are explored to learn discriminative features from abundant spectral information and spatial contexts in HSIs. The proposed ARCNN is an end-to-end deep learning framework that alleviates the decliningaccuracy phenomenon of deep learning models, and it also ranks the correlation and importance of each band in HSIs. Furthermore, the residual blocks connect every other 3-D convolutional layer by using an identity mapping, which facilitates backpropagation of gradients. In order to address the common issue of imbalance between high dimensionality and limited availability of training samples for HSI classification, an attention mechanism and a feature fusion block are investigated to improve the performance of the ARCNN. Finally, some strategies, batch normalization and dropout, are imposed on every convolutional layer to regularize the learning process. Therefore, the ARCNN method brings benefits to extract discriminative features, and it is easier to avoid overfitting. Experimental results on three public HSI datasets demonstrate the effectiveness of the ARCNN over some state-of-the-art methods.

Spectral–Spatial Self-Attention Networks for Hyperspectral Image Classification

“…Subsequently, many attention mechanisms proposed for scene segmentation of generic natural images in [1], [30], and [49]- [55] have been directly applied into the patch-based CNN for HSI classification. The squeeze-and-excitation (SE) block [55], which uses global pooling to generate the channel attention matrix, was applied to a patch-based CNN to recalibrate their channel-wise feature responses [51], [56]. Subsequently, many similar spectral attention modules [57], [58] were proposed for HSI classification to selectively excite informative channels and suppress useless ones.…”

Fully Group Convolutional Neural Networks for Robust Spectral–Spatial Feature Learning

“…Recent studies demonstrate the success of deep learning in HSI feature extraction [4][5][6]. Diverse deep learning models have been applied to spectral feature extraction, including stacked auto-encoders [7], deep belief networks [8], recurrent neural networks [9], one dimensional convolutional neural networks (1D-CNN) [10], and graph convolutional networks [11].…”

A Pixel Cluster CNN and Spectral-Spatial Fusion Algorithm for Hyperspectral Image Classification With Small-Size Training Samples

“…Numerous algorithms have been proposed for HSIC [10]- [21] and they can be mainly classified into unsupervised classification, semi-supervised classification and supervised classification based on whether the prior information is required in the classification process. Iterative self-organizing data analysis techniques algorithm (ISODATA) [10] and k-nearest neighbor (k-NN) [11] are typical unsupervised classification algorithms.…”

“…Then the trained classifier is used as decision rules to discriminate and classify the testing samples. Numerous supervised classification algorithms have been proposed and discussed, including support vector machine (SVM) [14], random forest (RF) [15], [16], decision tree [17], neural network [18]- [21], etc. All of those algorithms provide good ways for the classification of HSIs.…”