学术视点

学术视点

日期:2025-03-24

【字号:

题目:Exploring Applications of Convolutional Neural Networks in Analyzing Multispectral Satellite Imagery: A Systematic Review(探索卷积神经网络在分析多光谱卫星图像中的应用:系统评价)

作者:A. Ivanda, L. Šerić and M. Braović

来源:Big Data Mining and Analytics(大数据挖掘与分析), vol. 8, no. 2, pp. 407-429.

摘要:Remote sensing is of great importance for analyzing and studying various phenomena occurrence and development on Earth. Today is possible to extract features specific to various fields of application with the application of modern machine learning techniques, such as Convolutional Neural Networks (CNN) on MultiSpectral Images (MSI). This systematic review examines the application of 1D-, 2D-, 3D-, and 4D-CNNs to MSI, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines. This review addresses three Research Questions (RQ): RQ1: In which application domains different CNN models have been successfully applied for processing MSI data?, RQ2: What are the commonly utilized MSI datasets for training CNN models in the context of processing multispectral satellite imagery?, and RQ3: How does the degree of CNN complexity impact the performance of classification, regression or segmentation tasks for multispectral satellite imagery?. Publications are selected from three databases, Web of Science, IEEE Xplore, and Scopus. Based on the obtained results, the main conclusions are: (1) The majority of studies are applied in the field of agriculture and are using Sentinel-2 satellite data; (2) Publications implementing 1D-, 2D-, and 3D-CNNs mostly utilize classification. For 4D-CNN, there are limited number of studies, and all of them use segmentation; (3) This study shows that 2D-CNNs prevail in all application domains, but 3D-CNNs prove to be better for spatio-temporal pattern recognition, more specifically in agricultural and environmental monitoring applications. 1D-CNNs are less common compared to 2D-CNNs and 3D-CNNs, but they show good performance in spectral analysis tasks. 4D-CNNs are more complex and still underutilized, but they have potential for complex data analysis. More details about metrics according to each CNN are provided in the text and supplementary files, offering a comprehensive overview of the evaluation metrics for each type of machine learning technique applied.

编者译:遥感对于分析和研究地球上各种现象的发生和发展具有重要意义。如今,通过应用现代机器学习技术,例如,多光谱图像(MSI)上的卷积神经网络(CNN),可以提取各个应用领域的特定特征。本文分别研究了1D2D3D以及4DCNNMSI分析中的应用,主要解决了三个研究问题:①哪些领域中已成功利用CNN模型来处理MSI数据?②在多光谱卫星图像分析方面,有哪些常用的MSI数据集可用于训练CNN模型?③CNN复杂性对多光谱卫星图像分类、回归以及分割任务性能的影响?研究主要结论为:(1)农业领域应用最多,主要使用的是Sentinel-2卫星的数据;(21D2D3DCNN数据集多应用于分类任务。目前研究中所有4D-CNN数据集都应用于分割任务;(3)在所有应用领域占主导地位的2D-CNN3D-CNN更适合于时空模式识别的场景,即应用到农业和环境监测中。1D-CNN的应用不常见,但其在光谱分析任务中表现出良好的性能。4D-CNN更复杂,具有进行复杂数据分析的潜力,但目前未得到充分利用。文本和补充文件中提供了有关每个CNN指标的具体信息,并对所应用的各种机器学习技术的评估指标进行了总结描述。

题目:Generative AI-Driven Semantic Communication Networks: Architecture, Technologies, and Applications(生成式AI驱动的语义通信网络:架构、技术和应用)

作者:C. Liang et al

来源:IEEE Transactions on Cognitive Communications and NetworkingIEEE认知通信与网络汇刊), vol. 11, no. 1, pp. 27-47.

摘要:Generative artificial intelligence (GAI) has emerged as a rapidly burgeoning field demonstrating significant potential in creating diverse content intelligently and automatically. To support such artificial intelligence-generated content (AIGC) services, future communication systems must fulfill stringent requirements, including high data rates, throughput, and low latency, while efficiently utilizing limited spectrum resources. Semantic communication (SemCom) has been deemed as a revolutionary communication scheme to tackle this challenge by conveying the meaning of messages instead of bit reproduction. GAI algorithms serve as the foundation for enabling intelligent and efficient SemCom systems in terms of model pre-training and fine-tuning, knowledge base construction, and resource allocation. Conversely, SemCom can provide AIGC services with low latency and high reliability due to its ability to perform semantic-aware encoding and compression of data, as well as knowledge- and context-based reasoning. In this survey, we break new ground by investigating the architecture, wireless communication schemes, and network management of GAI-driven SemCom networks. We first introduce a novel architecture for GAI-driven SemCom networks, comprising the data plane, physical infrastructure, and network control plane. In turn, we provide an in-depth analysis of the transceiver design and semantic effectiveness calculation of end-to-end GAI-driven SemCom systems. Subsequently, we present innovative generation level and knowledge management strategies in the proposed networks, including knowledge construction, update, and sharing, ensuring accurate and timely knowledge-based reasoning. Finally, we explore several promising use cases, i.e., autonomous driving, smart cities, and the Metaverse, to provide a comprehensive understanding and future direction of GAI-driven SemCom networks.

编者译:生成式人工智能(GAI)已成为一个迅速发展的领域,在智能和自动生成多样化内容方面显示出巨大潜力。为了支撑人工智能生成内容(AIGC)服务,未来通信系统需要在满足高数据速率、吞吐量和低延迟等严格要求的同时,实现对有限频谱资源的有效利用。语义通信(SemCom)被认为是应对这一挑战的有效方案。SemCom可进行数据语义感知编码和压缩,实现基于知识和上下文的推理,从而提供低延迟和高可靠性的AIGC服务。GAI算法在模型预训练和微调、知识库构建和资源分配方面,为实现智能高效的SemCom系统奠定基础。因此,本文引入为GAI驱动的SemCom网络引入了一种新的架构,包括数据、物理基础设施和网络控制。然后,在该架构中本文提出了创新的生成和知识管理策略,包括知识的构建、更新和共享,以确保基于知识的推理准确性和及时性。最后,本文通过自动驾驶、智慧成熟和元宇宙等应用案例,分析了GAI驱动的SemCom网络的实际应用情况和未来发展方向。

题目:基于多类特征的社交网络影响力预测研究综述

作者:水映懿、张琪、李根、张士豪、吴尚

来源:数据与计算发展前沿, 2025, 7(1): 2-18.

摘要:影响力预测作为社交网络分析的重要内容,对于舆情监控、网络营销、情报分析、个性化推荐、广告定位、传播预测等多个领域具有重要的社会价值和现实意义。早期基于特征工程的影响力预测方法,通过提取并构建关键特征,建立不同特征与流行度之间的关系模型。本文重点关注与社交网络影响力相关的多类特征,从多类特征提取、预测模型构建和预测评估方法等方面进行了研究和综述,旨在综合分析已有研究方法,为提高社交网络影响力预测精度提供借鉴和参考。本文立足于当前广泛采用的深度学习方法,通过查阅文献资料,对社交网络的视觉特征、文本特征、情感特征、时间特征和用户特征分别进行了总结和阐述,并对基于多类特征的社交网络影响力预测方法的研究现状和局限性进行了分析。随着深度学习理论的发展,深度特征提取和预测模型构建取得了突破性进展,但目前在社交网络影响力预测方面,基于多类特征的特征组合预测方法仍然存在不足,需要研究更有效的特征预提取模型来提升社交网络影响力预测精度。

题目:材料领域知识图谱构建与应用研究

作者:袁扬、刘祺霖、陈子逸、万萌、李凯、王彦棡、王婧、王宗国

来源:数据与计算发展前沿, 2025, 7(1): 152-162.

摘要:为方便处理大规模非结构化材料科学信息,深入挖掘材料信息数据之间的关系,知识图谱为材料的特性和结构的集成研究提供了技术支持。本研究提出了一种新的知识图谱构建方式,构建了基于知识图谱的材料知识智能问答及材料知识图谱云平台。知识智能问答系统通过BERT+CRF构建,图谱云平台基于B/S框架开发。本研究构建的系统框架可以实现材料的知识问答,并以云平台的形式为用户提供使用入口,并可直观展示材料知识图谱。以III-V族化合物计算数据集为例,构建了III-V族化合物材料知识图谱。本研究扩展了知识图谱在材料领域的应用,对加快新材料的发现和挖掘材料的潜在信息具有重要意义。

题目:新质生产力驱动数字经济高质量发展

作者:马费成、孙玉姣、熊思玥

来源:信息资源管理学报, 2025, 15(1): 4-12.

摘要:新质生产力是新一轮科技革命和产业变革背景下催生的全新生产力形态,符合中国国内战略布局和国际竞争形势的现实需求。当前,数字经济已成为引领全球经济社会变革的重要力量,推动数字经济高质量发展成为新时代中国经济发展的战略选择。在此背景下,探究新质生产力如何驱动数字经济高质量发展,对于深化生产力理论创新、把握数字时代发展规律、推动经济转型升级具有重要意义。本研究首先引入新质生产力的孕育背景,从“新”的三重维度、“质”的双重含义和“生产力”的本质特征系统梳理其理论内涵,分析新质生产力的实践特征;在此基础上,从理论逻辑角度揭示新质生产力如何驱动数字经济高质量发展,归纳出新技术激发新动能、新要素重塑生产关系、新产业重构竞争格局三重内在机理;最后,提出推动数字经济高质量发展的五大路径:完善制度供给、加强人才培养、释放要素价值、促进区域协调、深化对外开放,以期对推动数字经济高质量发展提供参考。