Accurate weather forecasting plays a vital role in safeguarding human activities, mitigating the risks of extreme climate events, and supporting environmental policy and disaster preparedness. However, existing data-driven approaches often struggle to effectively model the complex spatiotemporal dynamics and multivariate dependencies inherent in meteorological systems, limiting their reliability and scalability. To address these challenges, we propose a novel dual-gated spatiotemporal attention network (DSANet) for multivariate weather prediction. DSANet integrates a convolutional self-attention hybrid module to jointly capture local and global spatial features, and a dual-gated channel-time module to model temporal patterns and inter-variable relationships. A wavelet-guided composite loss function is introduced to enhance prediction accuracy in fluctuating and dynamic weather regions. Extensive experiments on both global and regional datasets demonstrate that DSANet outperforms baseline models in terms of accuracy, with a mean absolute error of 1.78 K in 3-day lead-time global temperature forecasting. In addition, DSANet exhibits strong generalization and fast inference, making it well-suited for real-time and off-site forecasting. By significantly improving the accuracy, efficiency, and transferability of multivariate weather forecasting, DSANet provides a scalable and effective tool for next-generation climate intelligence and decision-making support systems.

Accurate prediction of sea surface temperature (SST) is practically important to ocean-related fields. However, due to the nonlinear temporal characteristics and complex spatial correlation of SST, fully extracting its spatiotemporal features for dependable short-term SST prediction remains challenging. In this work, we presented a multilayer channel spatial attention convolutional LSTM (CSACL) network to better capture the spatiotemporal information of SST recordings. Using multiscale spatial attention to obtain global–local spatial information and dynamic channel attention for interchannel interaction, the CSA module achieves comprehensive extraction of spatial features in SST data. Moreover, the multilayer long short-term memory (LSTM) structure ensures that the network can capture the temporal dependency of short-term SST very well. Finally, we proposed a novel two-stage adaptive loss function training method to solve the accuracy degradation problem for longer prediction cycles. We have evaluated CSACL on three public datasets, and the results demonstrate its superior performance over other classical methods in short-term SST prediction.

Real-time and accurate modeling of environmental variables in a specific region is of great significance to human production activities. In the era of intelligence, there is an increasing demand for accuracy and real-time performance of modeling environmental variables. However, due to the lack of a priori knowledge of spatiotemporal features and the high computational complexity, the existing deep learning methods have poor accuracy, and the real-time performance needs to be improved. To solve the above problems, this article proposes a novel SpatioTemporal MultiWaveNet (STMWNet) based on 3-D convolution for end-to-end environmental spatiotemporal variables generation. The Spatiotemporal MultiWave Layer (STMW Layer) is designed based on the principle of wavelet transform to extract multifrequency spatiotemporal features from the series. The multiscale embedding module (MSEM), time infofusion module (TIFM), and embedding integration module (EIM) are proposed to further extract and integrate the spatiotemporal features to improve the accuracy of the model. Considering the short-term and long-term temporal characteristics of spatiotemporal series, the date encoding strategy and the cross-attention head are proposed to integrate relative and absolute date information. The improved loss function is proposed to optimize the model learning. Comparative experiments on three datasets with different physical fields, different spatial scales, and resolutions show that the proposed method can accurately generate the mesoscale environmental spatiotemporal variables in real time for different tasks with the generation error lower than other methods by 1.02%~83.52% on root mean square error (RMSE), and 1.67%~83.68% on mean absolute error (MAE). Further experiments show that the proposed method can maintain its performance in practical applications.

For hyperspectral cross-domain recognition applications, the unseen target domain is inevitable, and the model can only be trained on the source domain but directly applied to unknown domains. A major challenge of this domain generalization problem comes from the domain shift caused by differences in environments, devices, etc. One feasible strategy is performing domain expansion with latent variables and learning domain-invariant representation. Inspired by this framework, the study proposes a generation network for extension, which consists of symmetric encoder-decoder to implicitly build local joint feature under style randomization. Moreover, supervised contrastive learning is employed to avoid duplicate augmentation. Besides, considering the trade-off between domain-specific and domain-invariant, an adversarial penalty term is formed by inter-class and intra-class contrastive regularization in the discriminator. Multiple evaluations on three public HSI datasets indicate that proposed method outperforms state-of-the-art approaches. The codes is available from the website: https://github.com/HUOWUMO/IEEE_HSIC_LLURnet.