Hyperspectral image (HSI) classification is commonly influenced by convolution neural networks (CNNs). However, the large number of parameters and computational complexity associated with CNNs can limit their practical application, particularly when computing and storage resources are limited. To address this challenge, we propose a channel-layer-oriented lightweight network for HSI classification. Motivated by existing structures that typically set large channels and stack multiple layers, we give more optimal solutions strategically to further compress the model. For intralayer feature extraction, we develop a channel-oriented spectral–spatial module (COS2M), which introduces a dual-single-channel (DSC) 3-D convolution that works in conjunction with depthwise convolution to fully extract spectral–spatial information. For interlayer information transmission, we propose a novel neighbor-pixel-aware activation function (NPAF), where the activation of a single pixel is determined by the learnable interaction with its neighbor range that enhances information transmission and improves the network’s fitting ability through the single activation layer. By implementing these strategies, we aim to overcome the limitations of traditional CNNs and enable efficient HSI classification within resource-constrained environments. The whole network is designed to be a compact end-to-end structure. It achieves better classification performance than other deep learning methods and lightweight models, even with limited training samples. The network parameters, model complexity, and inference time also demonstrate significant superiority, as confirmed by experiments on three benchmark datasets. The source codes are available publicly at: https://github.com/AchunLee/CLOLN_TGRS

In practical hyperspectral image cross-scene classification (HSICC) tasks, the arduous work of obtaining labels and the distribution inconsistency caused by spectral shift leave deep learning methods to face great challenges. Unsupervised domain adaptation aims to exploit knowledge from the annotated source domain and transfer it to the unlabeled target domain, thereby boosting the performance of unsupervised classification. Nevertheless, existing HSICC approaches cannot effectively exploit class structure information from target data. In particular, this article proposes an unsupervised joint adversarial domain adaptation (UJADA) architecture for HSICC to further narrow the distribution gap between distinct domains. The proposed method contains two modules: domain adversarial module that learns domain-invariant features, biclassifier adversarial module that explores task-specific decision boundaries between classes, and both share a feature generator consisting of the dense-based spectral–spatial convolution network. The UJADA simultaneously considers domain- and class-level feature alignment between source and target hyperspectral images in unified adversarial learning processing. Furthermore, the classifier determinacy disparity metric is introduced to fine-grained measure the output probabilistic discrepancy between two task-specific label predictors on target data, thus ensuring the discriminability of transferable features. Comprehensive experiments and ablation studies conducted on two public cross-scene data pairs and our newly acquired ultralow-altitude hyperspectral images under different illumination conditions demonstrate the superior performance of the proposed algorithm, which will greatly promote the practical application of hyperspectral intelligent perception technology.

The use of Transformer-based networks has been proposed for the processing of general point clouds. However, there has been little research related to multispectral LiDAR point clouds that contain both spatial coordinate information and multi-wavelength intensity information. In this paper, we propose networks for multispectral LiDAR point cloud point-by-point classification based on an improved Transformer. Specifically, considering the sparseness of different regions of multispectral LiDAR point clouds, we add a bias to the Transformer to improve its ability to capture local information and construct an easy-to-implement multispectral LiDAR point cloud Transformer (MPT) classification network. The MPT network achieves 78.49% $mIoU$, 94.55% $OA$, 84.46% $F_1$, and 0.92 $Kappa$ on the multispectral LiDAR point cloud testing dataset. To further extract the topological relationships between points, we present a standardization set abstraction (SSA) module, which includes the global point information while considering the relationships among the local points. Based on the SSA module, we propose an advanced version called MPT+ for the point-by-point classification of multispectral LiDAR point clouds. The MPT+ network achieves 82.94% $mIoU$, 95.62% $OA$, 88.42% $F_1$, and 0.94 $Kappa$ on the same testing dataset. Compared with seven point-based deep learning algorithms, our proposed MPT+ achieves state-of-the-art results for several evaluation metrics.

Convolutional neural networks (CNNs) have been prominent in most hyperspectral image (HSI) processing applications due to their advantages in extracting local information. Despite their success, the locality of the convolutional layers within CNNs results in heavyweight models and time-consuming defects. In this study, inspired by the excellent performance of transformers that are used for long-range representation learning in computer vision tasks, we built a lightweight vision transformer for HSI classification that can extract local and global information simultaneously, thereby facilitating accurate classification. Moreover, as traditional dimensionality reduction methods are limited in their linear representation ability, a three-dimensional convolutional autoencoder was adopted to capture the nonlinear characteristics between spectral bands. Based on the aforementioned three-dimensional convolutional autoencoder and lightweight vision transformer, we designed an HSI classification network, namely the “convolutional autoencoder meets lightweight vision transformer” (CAEVT). Finally, we validated the performance of the proposed CAEVT network using four widely used hyperspectral datasets. Our approach showed superiority, especially in the absence of sufficient labeled samples, which demonstrates the effectiveness and efficiency of the CAEVT network.

With the increasing demand for convolutional neural networks (CNNs) in many edge computing scenarios and resource-limited settings, researchers have made efforts to apply lightweight neural networks on hardware platforms. While binarized neural networks (BNNs) perform excellently in such tasks, many implementations still face challenges such as an imbalance between accuracy and computational complexity, as well as the requirement for low power and storage consumption. This paper first proposes a novel binary convolution structure based on the time domain to reduce resource and power consumption for the convolution process. Furthermore, through the joint design of binary convolution, batch normalization, and activation function in the time domain, we propose a full-BNN model and hardware architecture (Model I), which keeps the values of all intermediate results as binary (1 bit) to reduce storage requirements by 75%. At the same time, we propose a mixed-precision BNN structure (model II) based on the sensitivity of different layers of the network to the calculation accuracy; that is, the layer sensitive to the classification result uses fixed-point data, and the other layers use binary data in the time domain. This can achieve a balance between accuracy and computing resources. Lastly, we take the MNIST dataset as an example to test the above two models on the field-programmable gate array (FPGA) platform. The results show that the two models can be used as neural network acceleration units with low storage requirements and low power consumption for classification tasks under the condition that the accuracy decline is small. The joint design method in the time domain may further inspire other computing architectures. In addition, the design of Model II has certain reference significance for the design of more complex classification tasks.

Effective feature extraction (FE) has always been the focus of hyperspectral images (HSIs). For aerial remote-sensing HSIs processing and its land cover classification, in this article, an efficient two-staged hyperspectral FE method based on total variation (TV) is proposed. In the first stage, the average fusion method was used to reduce the spectral dimension. Then, the anisotropic TV model with different regularization parameters was utilized to obtain featured blocks of different smoothness, each containing multi-scale structure information, and we stacked them as the next stage’s input. In the second stage, equipped with singular value transformation to reduce the dimension again, we followed an isotropic TV model based on split Bregman algorithm for further detail smoothing. Finally, the feature-extracted block was fed to the support vector machine for classification experiments. The results, with three hyperspectral datasets, demonstrate that our proposed method can competitively outperform state-of-the-art methods in terms of its classification accuracy and computing time. Also, our proposed method delivers robustness and stability by comprehensive parameter analysis.