To improve the performance of the sea ice model in the Community Earth System Model (CESM) under a heterogeneous computing environment, this work conducts an in-depth study on the memory access optimization of the Elastic-Viscous-Plastic (EVP) dynamics model in Community Ice Code(CICE) on the SW39000 heterogeneous many-core processor, which is deployed in the new-generation Sunway supercomputer. The processor’s complex on-chip heterogeneous architecture, multi-level memory hierarchy, and unique inter-core communication mechanism present significant challenges for the parallel optimization of the sea ice dynamics simulation. To address the inefficiencies caused by diverse data access patterns, a differentiated processing strategy based on data read/write characteristics is proposed to reduce unnecessary data transfers. In addition, to alleviate load imbalance arising from the sparsity of sea ice boundary update data, a local dynamic compression method incorporating the probability density of data sparsity is designed. This method dynamically compresses data according to the probability density of Direct Memory Access (DMA) data transfers, thereby reducing communication volume and balancing the workload across slave cores. Finally, to enhance the computational intensity of the slave cores and reduce data dependencies between master and slave cores, an operator fusion algorithm based on Remote Memory Access (RMA) communication is introduced to achieve efficient data caching and transmission between operators. Experimental results demonstrate that, under the standard gx3 grid configuration, the optimized EVP model achieves a 27.54×speedup over the serial version running on a single master core when executed with a single core group. Multi-core group parallel tests validate the excellent scalability of the proposed optimization strategies, achieving up to a 123.93×speedup with a 10-core group, while also exhibiting effective load balancing in terms of both clock cycles and instruction counts across the slave-core array.

Monitoring glacier changes is essential for understanding global climate dynamics and assessing their environmental impacts. However, accurate detection remains challenging due to seasonal variations, illumination differences, and heterogeneous textures in remote sensing imagery. To address these issues, we propose VPGCD-Net, a Transformer-based dual-branch network that achieves robust glacier change detection through visual prompt engineering. The visual prompting branch integrates threshold segmentation and difference calculation, leveraging a visual prompt transformer (VPT) to encode regions of significant change and generate high-level semantic prompts. Meanwhile, the change detection branch adopts ResNet18 as the backbone to extract dual-temporal features, followed by a Transformer module for modeling global spatiotemporal dependencies and a FiLM module for adaptive feature modulation to emphasize real change regions. Complementing the method, we introduce the first polar glacier-focused dataset specifically designed for deep learning-based glacier change detection in remote sensing. Experimental results demonstrate that VPGCD-Net outperforms existing state-of-the-art methods, achieving superior accuracy even under complex conditions such as shadow interference. The dataset is publicly available at https://huggingface.co/datasets/cuibinge/Glacier-Dataset.

Domain adaptation methods enable model migration and adaptation across different domain data distributions. However, the source and target domains of hyperspectral images (HSIs) have large spectral offsets and spatial distribution differences, making the extraction of high-quality domain-invariant features between different domains is essential for classification. To achieve a more consistent feature representation for each category between the source and target domains, we propose a category semantic guided unsupervised domain adaptation network (CSGNet) for HSIs classification. CSGNet is designed to learn cross-domain invariant representation from category semantic information. First, to embed category semantic prior knowledge during feature learning, we extracted textual semantic features from the textual descriptions for each category and projected visual features into the semantic space via visual-linguistic alignment. A category representation memory pool is then introduced to store the visual-linguistic representations of different categories. Second, we propose a bi-classifier adversarial learning method designed to generate inconsistent category predictions in the unlabeled target domain, thereby enhancing the classifier’s discriminative capability regarding those hard-to-transfer features. Finally, to utilize the domain-invariant features stored in the category memory pool, a category attention module is proposed to guide the model’s adaptation to the data from different domains, mitigating the impact of the differences in the domain data distributions. Extensive experimental results validated on three cross-domain datasets demonstrate that the proposed method outperforms other state-of-the-art methods. The source code is available at http://github.com/cuibinge/CSGNet.

Radiology report generation that aims to accurately describe medical findings for given images, is pivotal in contemporary computer-aided diagnosis. Recently, despite considerable progress, current radiology report generation models still struggled to achieve consistent quality across difficult and easy samples, which dramatically impacts their clinical value. To solve this problem, we explore the difficult samples mining in radiology report generation and propose the Linear Hybrid-Reward based Reinforced Focal Learning (LHR-RFL) to effectively guide the model to allocate more attention towards some difficult samples, thereby enhancing its overall performance in both general and intricate scenarios. In implementation, we first propose the Linear Hybrid-Reward (LHR) module to better quantify the learning difficulty, which employs a linear weighting scheme that assigns varying weights to three representative Natural Language Generation (NLG) evaluation metrics. Then, we propose the Reinforced Focal Learning (RFL) to adaptively adjust the contributions of difficult samples during training, thereby augmenting their impact on model optimization. The experimental results demonstrate that our proposed LHR-RFL improves the performance of the base model across all NLG evaluation metrics, achieving an average performance improvement of 20.9% and 13.2% on IU X-ray and MIMIC-CXR datasets, respectively. Further analysis also proves that our LHR-RFL can dramatically improve the quality of reports for difficult samples. The source code will be available at https://github.com/ SKD-HPC/LHR-RFL.

Significant progress has been made in modern remote sensing (RS) image change detection (CD) by leveraging the powerful feature learning capabilities of convolutional neural networks (CNNs) and transformers. However, current popular change detection techniques primarily focus on extracting deep semantic features and pixel-level interactions while overlooking the potential benefits of cluster-level semantic interaction in bitemporal images. In this letter, we propose a novel approach called the Bitemporal Graph Semantic Interaction Network for Remote Sensing Images Change Detection (BGSINet-CD). Specifically, the land cover types in bitemporal images are clustered by employing soft clustering for each pixel, and then each cluster is separately projected to a vertex in graph space. Additionally, we introduce a graph semantic interaction module (GSIM) that enhances the interactions between bitemporal features at the semantic level. GSIM effectively improves the information coupling between bitemporal features, thereby suppressing task-irrelevant information. In comparison to other competing methods, our approach demonstrates a significant improvement in F1 scores, achieving 88.25% and 91.02% on the GZ-CD and WHU-CD datasets, respectively. Furthermore, our method employs a reduced number of network parameters and exhibits lower complexity, striking a superior balance between accuracy and computational efficiency.

In this paper, we investigate the hierarchical clustering for dense small cells and devise non-cooperative game-theoretic scheme, with aim at increasing the network throughput and minimizing both cross- and co-tier interference. By studying the distances between the small cells and the requirements of cluster formation, we propose a novel hierarchical clustering scheme for the densely deployed small cells, which consists of two phases, i) an absorption mechanism is designed to form the small cells into clusters; ii) to balance cluster populations, a segmentation algorithm is proposed for the clusters that contains excessive small cells. Within each small cell cluster, the spectrum is split into central frequency and marginal frequency which can be reused in a FFR manner. Then, we formulate the spectrum allocation and interference mitigation issues as a non-cooperative game, in which a game theoretical strategy optimization algorithm based on regret-matching is proposed to reach the correlated equilibrium. Numerical results reveal that our approach can achieve the correlated equilibrium with fast convergence and it is effective in offloading traffic and increasing the system throughput in dense small cell networks.