The global navigation satellite system reflectometry (GNSS-R) plays a crucial role in sea surface wind speed measurement, and convolutional neural networks (CNNs) have been a widely used method for wind speed retrieval from GNSS-R data. However, CNNs have limitations in global feature extraction due to the fixed convolutional kernels. Moreover, current studies on wind speed retrieval from GNSS-R data exhibit significant overfitting at low wind speed and underestimation at high wind speed due to the extremely imbalanced data distribution. To address these issues, a hybrid CNN-transformer network (CTN) with a weighted mean square error (mse) loss is proposed in this study. Specifically, the designed CTN incorporates CNN and transformer encoder blocks to capture local and global features, with a dedicated feature fusion module to integrate these features. In addition, a novel weighted mse loss function is designed to tackle the issue of imbalanced data distribution and enhance the estimation accuracy at high wind speeds. The proposed method is validated on Cyclone GNSS (CYGNSS) data and demonstrates improved performance compared with other machine learning algorithms. It achieves a root mean square difference (RMSD) of 1.417 m/s compared to the European Center for Medium-Range Weather Forecasts (ECMWF) reanalysis data and 1.620 m/s compared to the buoy data from the National Data Buoy Center (NDBC). Notably, the proposed model trained with the weighted mse loss function shows an improvement of 20.3% in RMSD for samples with wind speeds exceeding 15 m/s, highlighting the effectiveness of the designed loss function in handling high wind speed data.

Hyperspectral images (HSIs) provide hundreds of continuous spectral bands and have been widely used for the fine identification of targets with similar appearances. In earlier studies, convolutional neural networks (CNNs) have been an effective method for HSIs classification due to their powerful feature extraction capabilities. Recently, self-attention-based vision transformer (ViT) architecture has been widely explored to fully represent global information. However, most existing transformer-based models primarily focus on global relationships and lack the ability to capture the multi-scale features which are crucial for HSIs classification. This limitation results in inferior performance for transformer-based methods compared to state-of-the-art CNN-based models. To solve this problem, a novel network called multi-scale neighborhood attention transformer (MSNAT) is proposed in this paper. Unlike previous transformer-based models, MSNAT emphasizes the neighborhood pixels within a local window size and extracts multi-scale spatial information by using different local window sizes. In addition, a spatial transformation module is integrated to generate optimized spatial input. The effectiveness of the proposed MSNAT is verified on three real hyperspectral datasets including University of Pavia (UP), University of Houston (UH), and University of Trento (UT). Experimental results demonstrate that the proposed MSNAT method outperforms both CNNs and existing transformer-based models, achieving state-of-the-art classification performance with an overall accuracy of 93.34%, 86.26%, and 96.63% on UP, UH, and UT, respectively. The source code will be available at https://github.com/xinqiao123/MSNAT.

Hyperspectral images (HSIs) provide detailed spectral information of objects to be detected and play an important role in distinguishing targets with a similar appearance. However, the characteristics of high dimensionality and complexity impose significant challenges for realizing pixel-wise classification. Although existing convolutional neural networks (CNNs) and transformer-based models have presented promising performance for HSIs classification, they mainly extract features from spectral-spatial perspective and do not fully consider the information in the frequency domain. To address this issue, in this paper, we reconsider feature extraction and HSIs classification from the frequency domain. Specifically, inspired by the observation that high-frequency information contains detailed features within a local receptive field whereas low-frequency information provides global smooth variations, a frequency domain feature extraction (FDFE) block with dual branches is developed. In the FDFE block, an multi-head neighborhood attention (MSNA) block and a global filter block are designed to capture high- and low-frequency features, respectively. Besides, a pixel embedding module is constructed. Based on these, a novel hierarchical dual frequency transformer network (DFTN) is developed. Extensive experiments are performed on three open public hyperspectral datasets to evaluate the performance of our developed method. The experimental results demonstrate that our method is efficient and robust for HSIs classification, achieving overall accuracies of 94.14%, 86.92%, and 96.72% on the University of Pavia, University of Houston, and University of Trento datasets, respectively

Global navigation satellite system reflectometry (GNSS-R) is a novel remote sensing technique for global significant wave heights (SWHs) observation. Previous studies have illustrated the efficacy of deep learning methods in SWH retrieval from GNSS-R data. However, most of these methods rely on convolutional layers to extract features from Delay Doppler maps (DDMs), facing the limitations imposed by the fixed receptive field. To address this issue, in this study, a transformer-based network called WaveTransNet is proposed for SWH retrieval from GNSS-R data. Specifically, the transformer encoder block is exploited to capture long-range dependencies from DDMs. Besides, an attention mechanism-aided ancillary parameters feature extraction branch is devised to extract discriminative features from ancillary parameters, including geometry-related and map-related parameters. The developed model is evaluated on the Cyclone GNSS (CYGNSS) dataset and the experimental results demonstrate its improved performance. Compared with the European Center for Medium-Range Weather Forecasts (ECMWF) reanalysis data, it achieves a root mean square difference (RMSD) of 0.443 m and 0.444 m when National Data Buoy Center (NDBC) buoy data is used for evaluation.

Accurate global ocean surface wind speed estimation is crucial for weather forecasting and maritime transportation. Traditional retrieval methods based on Global Navigation Satellite System Reflectometry (GNSS-R) data and deep learning techniques often struggle to capture the complex, nonlinear relationships between GNSS-R observables and wind speed. This challenge is further exacerbated by imbalanced data distributions, which lead to significant underestimation under high wind conditions (15 - 30 m/s, here). To mitigate these issues, this study proposes a novel Physics-Informed Attention-Aided Convolutional Neural Network (PA-CNN). The proposed model incorporates an attention mechanism into the CNN architecture to adaptively focus on the most informative features. In addition, geophysical principles related to GNSS-R signal scattering are integrated with data-driven learning, improving both the interpretability and generalization capability of the network. Moreover, a spatial-temporal smoothing post-processing step is designed to enhance consistency in wind speed retrieval. Extensive experiments using Cyclone Global Navigation Satellite System (CYGNSS) datasets demonstrate that PA-CNN with smoothing outperforms existing deep learning approaches, achieving an overall root mean square difference (RMSD) of 1.38 m/s compared with ERA5 reanalysis data and 1.56 m/s against buoy measurements, highlighting the potential of combining physics-informed modeling with advanced deep-learning techniques for improved ocean surface wind speed retrieval.

Due to the advantage of high spectral resolution, hyperspectral imaging techniques have been extensively used in a variety of fields. Hyperspectral images (HSIs) classification is one of the fundamental tasks and attracts significant research interest. HSIs classification is pivotal as it facilitates precise identification of objects, providing invaluable insights for Earth observation tasks, such as resource management and land cover analysis. In existing studies, convolutional operations have been broadly applied for HSIs classification, especially 3-D convolution, which has shown its effectiveness in extracting spectral–spatial features from the raw HSIs. However, HSIs exhibit the characteristic of high dimensionality and pose challenges in extracting more discriminative features. In order to enhance the capability of capturing discriminative spectral–spatial features, in this article, a novel and effective 3-D sharpened cosine similarity (SCS) operation is proposed, serving as a replacement for conventional 3-D convolutional operation in HSIs classification and enhancing the classification accuracy. The 3-D SCS operation calculates and sharpens the cosine similarity between kernels and HSI input data. Based on the 3-D SCS operation, a 3-D SCS neural network is developed for HSIs classification tasks. To evaluate the effectiveness of 3-D SCS operation, experiments are conducted on three real-world HSIs datasets, including the University of Pavia, the University of Trento, and the University of Houston. Quantitative and qualitative experimental results illustrate that the SCS operation can effectively extract discriminative spectral-spatial features, achieving superior performance over the CNNs under the same model configuration.