Search

Remote sensing image scene classification is a challenging task that involves automatically assigning labels to remote sensing images based on predefined categories. The inherent intra-class diversity and inter-class similarity of remote sensing images make it difficult for classification models to capture the discriminative key information necessary for accurate labeling, resulting in classification confusion. This paper proposes a novel method called Multidimensional Attention and Feature Enhancement (MA-FE) to address this issue. The proposed MA-FE method comprehensively captures essential features in different dimensions of channel and position through the Multidimensional Attention (MA) module, which integrates and combines the captured features. The Feature Enhancement (FE) module then amplifies the discriminative features to suppress the interference of useless information, thus improving the representation ability of the model. We conducted detailed experiments on three public remote sensing datasets and performed a comparative evaluation with multiple remote sensing scene classification methods proposed in recent years. The overall accuracies of the proposed MA-FE method on these datasets were 99.66%, 95.68%, and 93.21%, respectively. Our experimental results demonstrate that the proposed MA-FE method is more effective in extracting complex features in remote sensing images than other methods, thereby proving its effectiveness. [ABSTRACT FROM AUTHOR]

Time series forecasting (TSF) is a prevalent research task in various fields such as medicine, transportation, environment, network detection, finance, and others. The TSF task aims to identify underlying patterns in data and make relatively accurate estimates of future data based on known values. In recent years, deep learning models have gained popularity for TSF tasks due to their capability to capture internal information effectively. However, traditional deep-learning models encounter difficulties when parallelizing data calculations, leading to error accumulation and reduced forecasting accuracy. Additionally, when dealing with excessively long input data, traditional deep learning models may experience performance degradation despite providing sufficient information and making it arduous to predict future data. Transformer-based models, with Self-Attention as the core, have shown the ability to facilitate global information interaction and enhance prediction efficiency. Nonetheless, they may encounter problems with significant and redundant parameters, causing unnecessary time overhead. To overcome these challenges, we propose a novel model called VarSeg-Trans, which incorporates three key optimizations: the cut-up mechanism, the variables-isolating mechanism, and an improved attention calculation method to enhance the transformer model's performance. Specifically, the cut-up mechanism enables the model to process longer input sequences, the variables-isolating mechanism mitigates overfitting, and the improved attention method leverages sequence information more effectively. Compared to other baseline TSF models and previous Transformer-based models, VarSeg-Trans has achieved an average reduction of 9% in MSE and MAE, along with a 3% increase in the coefficient of determination R². This trend is substantiated by consistent results across multiple experimental trials. [ABSTRACT FROM AUTHOR]

With the continuous development of 3D data acquisition technology, point cloud data, an essential form of 3D data representation is widely used in fields such as autonomous driving, indoor navigation, virtual reality, and augmented reality. As an essential branch of point cloud data processing, cloud classification has important research significance and value. However, this task poses challenges due to the sparsity, irregularity, and unordered nature of point cloud data. Among most of the methods dealing with this problem, there are problems of inadequate extraction of local features, low accuracy of classification, and poor generalization of point cloud data. Therefore, this paper addresses the above issue. This paper presents an improved 3D point cloud classification network by enhancing the Dynamic Graph Convolutional Neural Network (DGCNN). Firstly, this paper combines K-Nearest Neighbors (KNN) and ball radius query for feature extraction to better capture the local structural information in cloud classification data. Secondly, a conventional convolution layer is incorporated between the second and third graph convolution layers to enhance the feature representation in the proposed approach. Finally, this paper uses a global pooling method that combines maximum pooling and average pooling to construct the global structure of the point cloud while preserving the most critical information. Realize 3D point cloud classification. The experimental results demonstrate a 0.9% improvement in accuracy on the ModelNet40 dataset using the proposed improved method. This validates the effectiveness of the enhancement process described in this paper and offers valuable insights for enhancing the performance and application of point cloud classification. [ABSTRACT FROM AUTHOR]

Traffic flow prediction holds significant practical value in enabling efficient traffic management, enhancing transportation planning, and optimizing resource allocation. However, accurately forecasting traffic flow within a spatiotemporal context remains challenging due to complex dependencies within the traffic network. To model the spatiotemporal dependence of traffic sequence data, the existing traffic flow prediction models usually use the adjacency graph based on the prior spatial distribution as the input of shallow graph neural network (GNNs), but the over-expression of spatial adjacency relationship will affect the extraction effect of spatio-temporal dependence. Moreover, the spatio-temporal feature extraction ability of traditional GNNs is limited by the number of levels and the diffusion speed of node features. To address these issues, we propose a novel traffic flow prediction model with a unique historical time perspective: Temporal Branch Convolution Graph Neural ODE (TBC-GNODE), where Temporal Branch Convolution (TBC) is specifically designed to capture complex temporal dependencies in the data, and the Graph Neural ODE (GNODE) module uses adjacency graph with continuous correlation values to extract spatial dependencies, and applies optimized graph neural differential equations to represent the dynamic system of traffic flow variation. These modules are combined in a parallel structure and stacked in a residual manner to extract long-range spatiotemporal dependencies synchronously. Our experiments on four real datasets demonstrate that our proposed model improves prediction performance by over 4% on average compared to the baseline model. This showcases its practical application potential in the field. [ABSTRACT FROM AUTHOR]