Precipitation nowcasting is a highly challenging task in weather forecasting and plays a crucial role in protecting lives and property. However, autoregressive methods encounter training difficulties and are prone to error accumulation, whereas non-autoregressive models struggle to effectively utilize temporal information. To address these issues, this paper proposes a novel decoupled spatiotemporal network, TPDTC-Net, specifically for precipitation nowcasting. TPDTC-Net introduces an encoder–temporal predictor–decoder architecture based on a non-autoregressive model to prevent error accumulation, combining a time predictor and an adaptive dynamic weighting module that integrates the strengths of Transformer and convolutional neural networks, thereby improving the accuracy of precipitation nowcasting. From the spatial perspective, TPDTC-Net utilizes a multi-scale self-attention module in the encoder to extract global spatial features and employs a multiplicative convolution module to capture detailed features. In particular, the proposed adaptive dynamic weighting module effectively combines global and local features, allowing the encoder to dynamically adjust the weights based on different inputs and scenarios while learning feature fusion strategies. This enhances the supplementary role of convolution-extracted detail features to the global features extracted by the self-attention mechanism. From the temporal perspective, the time predictor adopts a Fourier self-attention mechanism and reshapes the feature maps passed from the encoder into abstract multivariate time series prediction tasks. By transforming unordered temporal information into ordered sequences, the time predictor effectively captures temporal dependencies, enabling the decoder to generate more accurate predictions. Extensive experiments on benchmark datasets show that TPDTC-Net significantly outperforms state-of-the-art networks in precipitation nowcasting. Specifically, on the KNMI dataset, compared to the second-best baseline model LPT-QPN (r≥ 10), the critical success index (CSI) and Heidke skill score (HSS) of TPDTC-Net increase by 10.35% and 9.37%, respectively. Besides, the balanced mean square error (BMSE) and balanced mean absolute error (BMAE) of TPDTC-Net decrease to 15.4787 and 1.1535. Similarly, on the CIKM AnalytiCup 2017 dataset, TPDTC-Net also delivers the best performance, with comparable performance trends observed. These results demonstrate the superior performance in terms of prediction accuracy of the proposed TPDTC-Net.

Precipitation nowcasting is a long-standing challenge due to the inherent unpredictability, which often lead to significant risks and damage. Traditional approaches that model nonlinear relationships between initial and future precipitation states often fail to accurately capture precipitation dynamics, including distribution and intensity patterns. Current data-driven methods are limited in their ability to represent the chaotic nature of precipitation without guidance from physical theory. To address this, we present Rainfusion, a generative model that integrates Prandtl’s mixing length theory from fluid dynamics with computer vision diffusion models. This integration accounts for nonlinear interactions between large-scale evolution and turbulent fluctuations in precipitation, generating physically plausible predictions. Rainfusion significantly improves forecasting skill on two benchmark dataset over the next 3 hours. Furthermore, we enhance Rainfusion with a control network trained on multi-source data, particularly lightning observations, enabling more accurate and controllable predictions of precipitation’s spatial-temporal patterns. Weather forecasters can utilize Rainfusion to guide predictions toward either growth or decay based on their domain expertise. Our approach advances precipitation nowcasting, offering a robust framework that bridges physical theory with modern deep learning techniques.

In the field of numerical weather prediction, fine-grained precipitation fields play a crucial role in forecasting and analyzing the spatial distribution and intensity of the precipitation. Historically, it is customary to employ the interpolation technique to downscale the low-resolution initial field output by assimilation systems, aligning with the requirements of a high-resolution forecasting model. Currently, data-driven deep learning methods offer novel solutions to address this challenge. In this letter, we propose a spatial downscaling algorithm for precipitation data generated from the North American Land Data Assimilation System (NLDAS), called Gaussian-perceptual-based generative adversarial network (GP-GAN). Specifically, the GP-GAN introduces a Siamese Gaussian-perceptual module (SGPM) which maps the data reconstructed from the generator and ground-truth to Gaussian latent space to learn the distribution of precipitation. Moreover, the adaptive weighted loss function (AWLF) is proposed to strengthen the emphasis and understanding of extreme precipitation events. Experimental results on the RainNet dataset comprising hourly precipitation over the USA demonstrate that GP-GAN provides better performance than other generative adversarial networks (GANs) and diffusion models in improving spatial resolution.

Accurate and timely estimation of tropical cyclone (TC) wind radii is significant for characterizing wind structure, disaster prevention, and mitigation. The existing methods have not sufficiently considered and utilized multimodal (i.e., multisource heterogeneous) data for wind radii estimation. Meanwhile, complete modalities (i.e., all used modalities) can hardly be available simultaneously, especially in real-time monitoring scenarios, which restricts the applicability of multimodal estimation models. It is challenging to maintain the accuracy of wind radii estimates when confronted with the issue of missing-modality. Therefore, to address these issues, this article aims to achieve robust TC wind radii estimation under both conditions with complete modalities and missing-modality. We first present a multimodal fusion network, MT-TCNet, for estimating TC wind radii under conditions with complete modalities. MT-TCNet benefits from multimodal data including satellite infrared (IR) images, reanalysis of wind fields, and the physical parameter maximum sustained wind (MSW) speed. MSW, which reflects TC intensity, is incorporated to embed the implicit relationship between TC intensity and wind radii. It is capable of providing superior and robust wind radii estimates in scenarios without time constraints, and can be used to generate long-term historical results. Furthermore, this article proposes MT-TCNet-Distill to alleviate the issue of missing-modality caused by delays in ERA5 reanalysis wind fields through generalized distillation and missing modality imputation. MT-TCNet-Distill broadens the applicability of MT-TCNet, which heavily relies on reanalysis data, enabling robust wind radii estimation in real-time scenarios. Comprehensive experiments demonstrate the superior performance of MT-TCNet and MT-TCNet-Distill compared to state-of-the-art methods.

Sea ice cannot only have a significant impact on hydrological changes, climate systems, and energy balance on Earth, but it can also directly interfere with maritime activities, posing serious obstacles to ocean economic development, polar scientific research, and other activities. In today’s world, where global warming is accelerating the melting of polar sea ice, the ability to accurately identify and classify sea ice becomes particularly important. Compared with traditional statistical methods, sea ice classification methods based on machine learning and deep learning have the advantages of fast computing speed and low resource consumption. These methods require a large amount of labeled data as a driver. However, the direct labeling cost of Synthetic Aperture Radar (SAR) sea ice image datasets is high, and previous sea ice classification methods struggle to balance data processing costs with classification accuracy. To address this problem, this paper proposes a sea ice classification method based on textural features and random forest, which is called SIC-TFRF. Specifically, we first extract the textural features based on the gray-level co-occurrence matrix (GLCM) and then use the wrapper method to filter and integrate them into the polarization features of the SAR images, guiding the training of the random forest model. Extensive experiments are conducted to verify the superiority of our proposals. Particularly, the accuracy of the binary classification of sea ice and seawater and the multi-classification of different types of sea ice and seawater can reach 96.51% and 92.78%, respectively.

Underwater acoustic signal denoising is a challenging task due to the complexity of the underwater environment. Most of the existing methods cannot effectively cope with the problem of underwater acoustic signal (UWAS) denoising at low signal-to-noise ratios (SNRs). According to the characteristics of UWAS, a novel idea is proposed to simultaneously model latent features from both the time and frequency dimensions of complex-valued spectrum in a dual-branch self-attention network, namely DBSA-Net. In this model, both magnitude and phase information in the complex spectrum are enhanced from different dimensions by two branches. Specifically, DBSA-Net is an encoder-decoder based network with several global-local-self-attention (GL-SA) blocks distributed on dual branches between encoder and decoder. Each GL-SA block incorporates global self-attention and local self-attention to capture distant context and fine-grained local dependencies along the temporal and frequency dimensions. Moreover, we also design an information interaction module between two branches to exchange complementary information. This interaction module together with a merge block fuse features extracted from different dimensions, thus enhancing the capability of our model to learn the target signal features. Extensive experiments are conducted to evaluate our model on a publicly available dataset. Results of the ablation experiments show that the different modules of DBSA-Net play their respective roles in improving denoising performance and are empirically valid. In both the seen ships and unseen ships scenarios, the proposed DBSA-Net outperforms existing approaches by a large margin on various evaluation metrics.

Quantitative precipitation nowcasting (QPN) is a highly challenging task in weather forecasting. The ability to provide precise, immediate, and detailed QPN products is necessary for a variety of situations, including storm warnings, air travel, and large gatherings. To address this challenge, this article proposes a new transformer lightweight physics-informed transformer (LPT)-QPN for QPN tasks, utilizing vertical cumulative liquid water content (VIL) products. This model adopts novel transformer modules to model the long-term evolution of precipitation and incorporates multihead squared attention (MHSA) to model its highly nonlinear relationships while reducing computational complexity. The results of experimental evaluations demonstrate the superiority of LPT-QPN when compared to existing state-of-the-art QPN models. In particular, the LPT-QPN model demonstrates greater accuracy for long lead time and in high-intensity areas, confirmed in both quantitative and qualitative evaluations. In addition, through three customized fine-tuning schemes, we are able to further improve the predictability of the LPT-QPN model for specific precipitation events. By incorporating the physical constraints of the convection-diffusion equation, our approach offers novel perspectives for future explorations that combine physical prior knowledge and deep-learning (DL) techniques.

Underwater acoustic signal recognition systems face challenges in achieving high accuracy when processing complex data with low signal-to-noise ratio (SNR) in underwater environments, leading to limited noise robustness. Conventional approaches typically employ pre-trained denoising models for preprocessing noisy signals. However, due to disparate optimization goals between denoising and recognition models, denoising methods might introduce signal distortion, hampering effective enhancement of system accuracy. To address this issue, this paper proposes a novel joint training framework with cross-attention fusion for robust underwater acoustic signal recognition (UASR), called CAF-JT. CAF-JT consists of a denoising module, a recognition module, and the CAF module. It addresses the mismatch problem arising from different optimization directions by jointly training the denoising frontend and the recognition backend. Additionally, inspired by the multi-condition training (MCT) method, the CAF module is designed to fuse characteristics from both denoised and noisy audio, thus incorporating noise information. This fusion mechanism enables the model to better adapt to the characteristics of the noisy environment and enhance its noise robustness. Furthermore, to improve the performance of UASR, TF-Transformer blocks are incorporated into both the denoising module and the recognition module to capture the spatio-temporal distribution of spectral features. The proposed approach is evaluated on two open-source underwater acoustic signal datasets, namely ShipsEar and DeepShip. Extensive experimental demonstrate the superiority of CAF-JT over conventional joint training approaches, showcasing its improved noise robustness. Particularly in low SNR conditions, CAF-JT achieves the best average recognition rates of 94.84% and 93.61% on the two datasets, respectively.

Dimensionality reduction is a fundamental technique to address the curse of dimensionality problem in real-world big datasets. However, most existing methods either only target raw datasets that contain explicit relationships between data points, or construct the complete neighborhood graph of the dataset by calculating pairwise similarities, and then generate contexts of data points by random walking to measure the structure of the dataset, which are computationally expensive. In this paper, we propose a fast nonlinear locality-preserving dimensionality reduction approach called FVec2vec, which extends the Skip-gram model to embedding representation of general numerical matrices. Specifically, instead of constructing neighborhood graph by calculating pairwise similarities between data points, we approximate the k-nearest neighbors (kNN) of each data point in matrices by exploring its neighbors’ neighbors first. Then, we design a novel sampling algorithm to randomly sample on the kNN to depict the structure of the dataset. Experimental results show that FVec2vec is faster than most existing methods while achieving acceptable accuracy, and the accuracy is even higher than the state-of-the-art method under certain similarity metrics.

Cloud computing has emerged as a promising computational infrastructure for cost-efficient workflow execution by provisioning on-demand resources in a pay-as-you-go manner. While scientific workflows require accessing community-wide resources, they usually need to be performed in collaborative cloud environments composed of multiple datacenters. Although such environments facilitate scientific collaboration, the movements of input and intermediate datasets across geographically distributed datacenters may cause intolerable latency that would hinder efficient execution of large-scale data-intensive scientific workflows. To address the problem, in this article we propose a novel multi-level K-cut graph partitioning algorithm to minimize the volume of data transfer across datacenters while satisfying load balancing and fixed data constraints. The algorithm first contracts the fixed input datasets in the same datacenter and their consuming tasks, and coarsens the contracted graph to a predefined scale in a level-by-level manner. Then, a K-cut algorithm is used to partition the resulted graph into K parts such that the cut size is minimized. After that, the partitioned graph is projected back to the original workflow graph, during which the load balancing constraint is maintained. We evaluate our algorithm using three real-world workflow applications and the results demonstrate that the proposed algorithm outperforms other state-of-the-art algorithms.

Efficient scheduling algorithm is critical for DAG-based applications to obtain high-performance in heterogeneous computing systems. In comparison with heuristic-based algorithms, meta-heuristic based scheduling algorithms can produce better results by searching in a guided manner. Biogeography-based optimization (BBO) is a recently proposed optimization technique which has shown less parameters, faster convergency, and superior performance than existing meta-heuristics. In this article, we introduce this novel optimization technique into the field of DAG scheduling. To reduce scheduling overhead, the proposed algorithm only encodes task mapping while using a heuristic strategy to determine task ordering. Moreover, it uses heuristic-based algorithms as baseline algorithms to obtain better results. We evaluate the BBO-based scheduling algorithm using three real world DAG-based applications under various parameter settings. The results show that the BBO-based scheduling algorithm outperforms the state-of-the-art meta-heuristic based algorithms.

Long-term execution of scientific applications often leads to dynamic workloads and varying application requirements. When the execution uses resources provisioned from IaaS clouds, and thus consumption-related payment, efficient and online scheduling algorithms must be found. Portfolio scheduling, which selects dynamically a suitable policy from a broad portfolio, may provide a solution to this problem. However, selecting online the right policy from possibly tens of alternatives remains challenging. In this work, we introduce an abstract model to explore this selection problem. Based on the model, we present a comprehensive portfolio scheduler that includes tens of provisioning and allocation policies. We propose an algorithm that can enlarge the chance of selecting the best policy in limited time, possibly online. Through trace-based simulation, we evaluate various aspects of our portfolio scheduler, and find performance improvements from 7% to 100% in comparison with the best constituent policies and high improvement for bursty workloads.

SUMMARY Due to its advantages of cost‐effectiveness, on‐demand provisioning and easy for sharing, cloud computing has grown in popularity with the research community for deploying scientific applications such as workflows. Although such interests continue growing and scientific workflows are widely deployed in collaborative cloud environments that consist of a number of data centers, there is an urgent need for exploiting strategies which can place application datasets across globally distributed data centers and schedule tasks according to the data layout to reduce both latency and makespan for workflow execution. In this paper, by utilizing dependencies among datasets and tasks, we propose an efficient data and task coscheduling strategy that can place input datasets in a load balance way and meanwhile, group the mostly related datasets and tasks together. Moreover, data staging is used to overlap task execution with data transmission in order to shorten the start time of tasks. We build a simulation environment on Tianhe supercomputer for evaluating the proposed strategy and run simulations by random and realistic workflows. The results demonstrate that the proposed strategy can effectively improve scheduling performance while reducing the total volume of data transfer across data centers. Concurrency and Computation: Practice and Experience, 2013.© 2013 Wiley Periodicals, Inc.

Due to the advantages of cost-effectiveness, on-demand resource provision and easy for sharing, cloud computing has grown in popularity with research community for deploying scientific applications such as workflows. When such interest continues growing and workflows are widely performed in collaborative cloud environments that consist of a number of data centers, there is an urgent need for exploiting strategies which can place the application data across globally distributed data centers and schedule tasks according to the data layout to reduce both the latency and make span for workflow execution. In this paper, by utilising dependencies among datasets and tasks, we propose an efficient data and task co scheduling strategy that can place input datasets in a load balance way and meanwhile group the mostly related datasets and tasks together. We build a simulation environment on Tianhe supercomputer to evaluate the proposed strategy and run simulations by random and realistic workflows. The results demonstrate that the proposed strategy can effectively improve workflows performance while reducing the total volume of data transfer across data centers.

Incorporating features extracted from clickthrough data (called clickthrough features) has been demonstrated to significantly improve the performance of ranking models for Web search applications. Such benefits, however, are severely limited by the data sparseness problem, i.e., many queries and documents have no or very few clicks. The ranker thus cannot rely strongly on clickthrough features for document ranking. This paper presents two smoothing methods to expand clickthrough data: query clustering via Random Walk on click graphs and a discounting method inspired by the Good-Turing estimator. Both methods are evaluated on real-world data in three Web search domains. Experimental results show that the ranking models trained on smoothed clickthrough features consistently outperform those trained on unsmoothed features. This study demonstrates both the importance and the benefits of dealing with the sparseness problem in clickthrough data.

Vertical search is a promising direction as it leverages domain-specific knowledge and can provide more precise information for users. In this paper, we study the Web object-ranking problem, one of the key issues in building a vertical search engine. More specifically, we focus on this problem in cases when objects lack relationships between different Web communities, and take high-quality photo search as the test bed for this investigation. We proposed two score fusion methods that can automatically integrate as many Web communities (Web forums) with rating information as possible. The proposed fusion methods leverage the hidden links discovered by a duplicate photo detection algorithm, and aims at minimizing score differences of duplicate photos in different forums. Both intermediate results and user studies show the proposed fusion methods are practical and efficient solutions to Web object ranking in cases we have described. Though the experiments were conducted on high-quality photo ranking, the proposed algorithms are also applicable to other ranking problems, such as movie ranking and music ranking.

In this paper, we propose, IGroup, an efficient and effective algorithm that organizes Web image search results into clusters. IGroup is different from all existing Web image search results clustering algorithms that only cluster the top few images using visual or textual features. Our proposed algorithm first identifies several query-related semantic clusters based on a key phrases extraction algorithm originally proposed for clustering general Web search results. Then, all the resulting images are separated and assigned to corresponding clusters. As a result, all the resulting images are organized into a clustering structure with semantic level. To make the best use of the clustering results, a new user interface (UI) is proposed. Different from existing Web image search interfaces, which show only a limited number of suggested query terms or representative image thumbnails of some clusters, the proposed interface displays both representative thumbnails and appropriate titles of semantically coherent image clusters. Comprehensive user studies have been completed to evaluate both the clustering algorithm and the new UI.

In this demo, we present IGroup, a Web image search engine that organizes the search results into semantic clusters. Different from all existing Web image search results clustering algorithms that only cluster the top few images using visual or textual features, IGroup first identifies several query-related semantic clusters based on a key phrases extraction algorithm originally proposed for clustering general Web search results. Then, all the resulting images are separated and assigned to corresponding clusters. To make the best use of the clustering results, a new user interface is proposed. Please go to http://igroup.msra.cn for real experience.

In this paper, we propose building a vertical image search engine called EnjoyPhoto that leverages rich metadata from various photo forum web sites to meet users' requirements for enjoying high-quality photos, which is virtually impossible in traditional image search engines. To solve the ranking problem when aggregating multiple photo forums, we propose a novel rank fusion algorithm that uses duplicate photos to normalize rating scores. To further improve user experiences in enjoying photos, we design an in-place image browsing interface, and compare it with several other interfaces in a user study. With rich metadata and rating information, more attractive user interfaces are enabled, including slideshow authoring and photo recommendations. We conducted experiments and user studies on a 2.5-million image database to evaluate the proposed rank fusion algorithm, investigate the rationale behind building a vertical image search engine, and study user interfaces and preferences for the purpose of enjoying high-quality photos. The experimental results demonstrate the effectiveness of the proposed ranking algorithm. The results also show that the 2.5-million high-quality image database in EnjoyPhoto performs comparably with Google's 1- billion image database for queries related to location, nature, and daily life categories. Finally, our results show that the in-place browsing interface-called Force-Transfer view-is much more convenient for users than traditional interfaces.