Graph convolutional network (GCN) is an efficient network for learning graph representations. However, it costs expensive to learn the high-order interaction relationships of the node neighbor. In this paper, we propose a novel graph convolutional model to learn and fuse multihop neighbor information relationships. We adopt the weight-sharing mechanism to design different order graph convolutions for avoiding the potential concerns of overfitting. Moreover, we design a new multihop neighbor information fusion (MIF) operator which mixes different neighbor features from 1-hop to k-hops. We theoretically analyse the computational complexity and the number of trainable parameters of our models. Experiment on text networks shows that the proposed models achieve state-of-the-art performance than the text GCN.

Graph convolutional network (GCN) has attracted increasing attention in hyperspectral image (HSI) classification due to its capability to capture the long-range correlations between adjacent land covers. Most GCN-based HSI classification methods have been proposed to address the four limitations (shape-fixed kernel, massive calculations and parameters, limited classification ability with limited labelled samples, and difficult to capture the long-term relationships of land covers) of Convolutional Neural Networks (CNNs) by operating on superpixel-based nodes. However, the pixels in each superpixel share the spectral-spatial features, overlooking the unique characteristics of individual pixels. To address these limitations of GCN and CNN and fully exploit their advantages, we propose a novel Multi-Branch Attention Framework (MFAF), in which the specially designed GCN and Convolutional Neural Network (CNN) branches learn the complementary spectral-spatial features. Specifically, we develop a multi-scale attentional graph convolutional network to enhance the ability to understand the long-range correlations between land covers, accomplished by constructing the multi-scale attentional adjacency matrix. Then, based on the two designs of the dual branch depthwise separable convolution and the attention-based residual block, we present a new complementary dual convolutional attention network that extracts more discriminative spectral-spatial features of pixels. Finally, we introduce an Attention Fusion Pooling (AFP) mechanism to combine the features generated by different network branches. Extensive experimental evaluations on four public HSI datasets demonstrate that the proposed MFAF achieves better performance than several state-of-the-art methods, delivering superior and consistent results in terms of overall accuracy (OA), average accuracy (AA), and kappa coefficient (KAPPA).

Deep learning-based hyperspectral image (HSI) classification methods, such as Transformers and Mambas, have attracted considerable attention. However, several challenges persist, e.g., (1) Transformers suffer from quadratic computational complexity due to the self-attention mechanism; and (2) both the local and global feature extraction capabilities of large kernel convolutional neural networks (LKCNNs) need to be enhanced. To address these limitations, we introduce a multi-scale large kernel asymmetric CNN (MSLKACNN) with the large kernel sizes as large as $1\times 17$ and $17\times 1$ for HSI classification. MSLKACNN comprises a spectral feature extraction module (SFEM) and a multi-scale large kernel asymmetric convolution (MSLKAC). Specifically, the SFEM is first utilized to suppress noise, reduce spectral bands, and capture spectral features. Then, MSLKAC, with a large receptive field, joins two parallel multi-scale asymmetric convolution components to extract both local and global spatial features: (C1) a multi-scale large kernel asymmetric depthwise convolution (MLKADC) is designed to capture short-range, middle-range, and long-range spatial features; and (C2) a multi-scale asymmetric dilated depthwise convolution (MADDC) is proposed to aggregate the spatial features between pixels across diverse distances. Extensive experimental results on four widely used HSI datasets show that the proposed MSLKACNN significantly outperforms ten state-of-the-art methods, with overall accuracy (OA) gains ranging from 4.93% to 17.80% on Indian Pines, 2.09% to 15.86% on Botswana, 0.67% to 13.33% on Houston 2013, and 2.20% to 24.33% on LongKou. These results validate the effectiveness of the proposed MSLKACNN.

Existing popular methods for semi-supervised node classification with high-order convolution improve the learning ability of graph convolutional networks (GCNs) by capturing the feature information from high-order neighborhoods. However, these methods with high-order convolution usually require many parameters and high computational complexity. To address these limitations, we propose HCNP, a new higher-order GCN for semi-supervised node learning tasks, which can simultaneously aggregate information of various neighborhoods by constructing high-order convolution. In HCNP, we reduce the number of parameters using a weight sharing mechanism and combine the neighborhood information via multi-scale neighborhood pooling. Further, HCNP does not require a large number of hidden units, and it fits a few parameters and exhibits low complexity. We show that HCNP matches GCNs in terms of complexity and parameters. Comprehensive evaluations on publication citation datasets (Citeseer, Pubmed, and Cora) demonstrate that the proposed methods outperform MixHop in most cases while maintaining lower complexity and fewer parameters and achieve state-of-the-art performance in terms of accuracy and parameters compared to other baselines.

Deep learning-based hyperspectral image (HSI) classification models typically utilize multiple feature extraction layers to learn the features of land covers. Nevertheless, they encounter challenges, e.g., 1) Transformers require substantial computational resources, and 2) these layers are carefully assembled and designed. Recently, large kernel convolutional neural networks (LKCNNs) show excellent performance in natural visual tasks. To tackle these limitations and explore the capability of LKCNNs for HSI classification, we present a novel simple and powerful multi-scale large kernel convolutional neural network architecture (MSLKCNN) with the largest kernel size as large as 15 × 15, in contrast to commonly used 3 × 3, for HSI classification. MSLKCNN avoids these specialized designs, comprising a noise suppression module (NSM) and a multi-scale large kernel convolution (MSLKC). Specifically, NSM is first used to suppress the noise and reduce the number of the bands before extracting the features. Then, MSLKC, as the only feature extraction layer of MSLKCNN, joints three parallel convolutions to capture the features of various types (i.e. spectral, spectral-spatial) and ranges (i.e., small local, larger local, and global) from the dimension of scale: (C1) convolution with a kernel size of 1 × 1 is used to extract spectral features; (C2) multi-scale large kernel depthwise separable convolution (MLKDC) is proposed to learn the spectral-spatial features of different ranges including short-range, middle-range, and long-range; and (C3) multi-scale dilated depthwise separable convolution (MDDC) is designed to aggregate the spectral-spatial features between land covers at various distances. Extensive experimental results on three public HSI datasets demonstrate the competitiveness of the proposed MSLKCNN compared with several state-of-the-art methods.