Wind forecasting is a typical and fundamental problem in efficient operation of wind power generation systems. Despite of the developed brilliant techniques, current methods for wind forecasting still rely heavily on numerical weather prediction. To investigate the huge amount of meteorological data and to improve the forecasting ability, this paper proposes a temporal dynamic graph convolutional network with learnable coupled adjacent matrix (TCNet), which takes observations of multiple meteorological elements as predictors for wind forecasting. Specifically, for better illustrating the inner property of wind components, a learnable coupled adjacent matrix (CAM) module is introduced. A spatially stable adjacent matrix is built to model the low-frequency part while a temporally dynamic adjacent matrix is developed to extract the high-frequency one. Casting on this mechanism, the CAM can be embedded to vanilla graph convolution network. Apart from that, a simple but effective temporal weights allocating strategy, named temporal dynamic (TED) module, is proposed to depict the cyclicity. By integrating TED with CAM, TCNet can effectively extract the spatio-temporal feature of wind speed and its correlation with other meteorological elements. Using observed datasets of meteorological elements in Denmark and Netherland, we conduct experiments to validate the performance and efficiency of our proposed model. The results indicate the proposed TCNet outperforms the state-of-the-art Graph Convolutional Networks methods and wind forecasting methods.

High temperature is one of the most severe disasters in the world, which causes the death of millions of people each year. Accurate temperature forecasting, as a key member of weather prediction, is of great application value. Many recent contributions treat weather forecasting as a spatio-temporal learning task, in which these methods mainly capture the motion of a rigid body while neglecting the generation and dispersion of fluid elements. However, typical spatio-temporal prediction is quite different from meteorological forecasting. To resolve this key issue, this letter proposes MetPGNet, a meteorological prior guided network for hourly temperature forecasting. Specifically, under the framework of atmospheric theory, three simple but effective multidimensional attention branches, i.e., the advection branch, the vertical branch, and the temporal branch, are elaborately designed to depict the temperature variation of spatial atmospheric advection, vertical atmospheric movement, and temporal heat exchange, respectively. Experiments compared with state-of-the-art spatio-temporal learning and weather prediction methods demonstrate the superiority of the proposed MetPGNet. Specifically, MetPGNet gets an improvement of 0.52 on mean absolute error (MAE) compared with vanilla ConvGRU.

Weather forecasting, which is challenging due to the complex atmospheric correlation, focuses on providing explicit meteorological estimations as accurate as possible. Recently, techniques based upon convolutional-recurrent networks have shown dramatic performance in domains including radar echo prediction and precipitation forecasting, indicating that deep learning models have great potential in this area. However, existing methods concentrate on adding extra paralleled memory cells to the inner recurrent unit, where the information is mutually independent. To extract spatial–temporal features with stronger correlation, this letter introduced an axial attention memory module with quasi state-in-state cascaded manner. Benefitting from this unit, the spatial–temporal features can be aggregated and embedded into standard ConvGRU formulating Axial aTtention Memory Cascaded ConvGRU (ATMConvGRU) for sequential weather prediction. Experimental results compared with many state-of-the-art methods on four types of weather forecasting related datasets, including temperature, relative humidity, wind, and radar echo, demonstrate the effectiveness and superiority of the proposed ATMConvGRU.

Precipitation downscaling, which is similar to the mechanism of single-image super-resolution (SR), aims to improve the spatial resolution of rain maps. It is of great practical value and theoretical significance. This letter presents a new deep precipitation downscaling (DPD) method, named auxiliary guided spatial distortion (AGSD) network, motivated by SR techniques. Specifically, an auxiliary guided module (AGM), which takes multiple meteorological elements (e.g., temperature, relative humidity, and wind) as input, is proposed for getting more accurate rain map features. Meanwhile, a simple but effective spatial distortion module (SDM) is proposed. Benefitting from SDM, the DPD model can rectify the rain map via terrain correlation. Furthermore, to improve the model performance among various rain intensity (including small rain, moderate rain, heavy rain, and storm), a threat score-driven pseudo threat-score (PTS) loss is presented. Experimental results compared with state-of-the-art methods demonstrate the superiority of the proposed method.

Action recognition plays a fundamental role in computer vision and video analysis. Nevertheless, extracting effective spatial-temporal features remains a challenging task. This paper proposes three simple but effective weakly semantic guided modules (SGMs) for both environment-constrained and cross-domain action recognition. The SGMs are composed of total 3-D convolution and element-wise gated operations; thus, they are efficient and easy to implement. The semantic guidance is obtained in a weakly supervised manner, in which each video clip is labeled with only an action class instead of pixel-level semantics. Benefitting from the semantic guidance, the network [called semantic guided network (SGN)] can focus on the salient parts of the video clips. Consequently, the redundant information can be reduced and the model is more robust to noise. Besides, benefitting from the intrinsic property of SGMs, SGN is totally end-to-end trainable. Quantities of experiments on both environment-constrained (e.g., Penn, HMDB-51, and UCF101) and cross-domain (e.g., ODAR) action recognition datasets demonstrate its effectiveness. Specifically, SGN gets improvements of 3.7%, 2.1%, and 5.2% for Penn, HMDB-51, and UCF-101 than the baseline ResNet3D, respectively, and SGN ranked third place in the ODAR 2017 challenge.

Current state-of-the-art action recognition approaches rely on optical flow to extract the local motion information and ignore the importance of global description of the videos. In this paper, we present a novel architecture, named Multi-Glimpse Network (MGN), to boost the performance of action recognition by combining the local and global information of the videos. Specifically, MGN makes predictions through two important modules, Local Glimpse and Global Glimpse. Local Glimpse extracts the local spatiotemporal features of different periods using temporal sampling method. Global Glimpse aggregates the extracted local features to develop global description of the videos. These two modules are complementary and indispensable. Our MGN achieves competitive results on four video action benchmarks of UCF10l, HMDB51, ODAR and Penn.

Autoencoder, which learns latent representations of samples in an unsupervised manner, has great potential in computer vision and signal processing. However, the diversity of samples makes learning a component autoencoder remaining a challenging task. This letter proposes a novel Self-Paced AutoEncoder (SPAE) for unsupervised feature extraction. The motivation behind this letter is to take samples gradually from simple to complex into consideration during training, which is similar to the mechanism of knowledge acquisition for humans. Under the unsupervised learning framework constructed on the autoencoder infrastructure, our SPAE first learns a weak autoencoder via samples with small losses and, then, elevates itself to a relatively strong autoencoder through samples with large losses. Then, the SPAE is generalized to a temporal domain, resulting to temporal SPAE (TSPAE), where the temporal information is explored and exploited to improve the performance. Typically, a TSPAE is capable of compressing temporal sequences into temporal-independent data. Experiments on the image classification and action recognition demonstrate the effectiveness of SPAE and TSPAE.

Extracting spatial-temporal descriptors is a challenging task for video-based human action recognition. We decouple the 3D volume of video frames directly into a cascaded temporal spatial domain via a new convolutional architecture. The motivation behind this design is to achieve deep nonlinear feature representations with reduced network parameters. First, a 1D temporal network with shared parameters is first constructed to map the video sequences along the time axis into feature maps in temporal domain. These feature maps are then organized into channels like those of RGB image (named as Motion Image here for abbreviation), which is desired to preserve both temporal and spatial information. Second, the Motion Image is regarded as the input of the latter cascaded 2D spatial network. With the combination of the 1D temporal network and the 2D spatial network together, the size of whole network parameters is largely reduced. Benefiting from the Motion Image, our network is an end-to-end system for the task of action recognition, which can be trained with the classical algorithm of back propagation. Quantities of comparative experiments on two benchmark datasets demonstrate the effectiveness of our new architecture.

Mutlilabel classification arises in various domains including computer vision and machine learning. Given a single instance, multilabel classification aims to learn a set of labels simultaneously. However, existing methods fail to address two key problems: 1) exploiting correlations among instances and 2) reducing computational complexity. In this letter, we propose a new semisupervised multilabel classification algorithm with joint dimensionality reduction. First, an elaborate matrix is designed for evaluating instance similarity; thus, it can take both labeled and unlabeled instances into consideration. Second, a linear dimensionality reduction matrix is added into the framework of multilabel classification. Besides, the dimensionality reduction matrix and the objective function can be optimized simultaneously. Finally, we design an efficient algorithm to solve the dual problem of the proposed model. Experiment results demonstrate that the proposed method is effective and promising.