[{"data":1,"prerenderedAt":277},["ShallowReactive",2],{"content-query-RzhOku4CLw":3},{"_path":4,"_dir":5,"_draft":6,"_partial":6,"_locale":7,"title":8,"description":9,"date":10,"cover":11,"type":12,"body":13,"_type":271,"_id":272,"_source":273,"_file":274,"_stem":275,"_extension":276},"/news/zh/2859","zh",false,"","从感知识别到认知预测：空天•灵眸时空预测基础模型","随着对地观测手段的不断进步，每天可获取海量多模态、多分辨率、多时相的遥感数据。","2023-10-31","https://obs-mindspore-file.obs.cn-north-4.myhuaweicloud.com/file/2023/11/10/a2d84dcc41bb4e7fb24b88d468476d10.png","news",{"type":14,"children":15,"toc":268},"root",[16,24,38,54,85,90,99,104,118,126,131,136,143,148,156,161,166,173,178,186,191,196,203,208,216,221,233,241,246,254,263],{"type":17,"tag":18,"props":19,"children":21},"element","h1",{"id":20},"从感知识别到认知预测空天灵眸时空预测基础模型",[22],{"type":23,"value":8},"text",{"type":17,"tag":25,"props":26,"children":27},"p",{},[28,30,36],{"type":23,"value":29},"随着对地观测手段的不断进步，每天可获取海量多模态、多分辨率、多时相的遥感数据。智能解译是遥感数据处理分析的核心手段之一，但是目前侧重于感知目标要素位置、类别等信息，难以反映目标要素的未来动向。因此，如何通过掌握目标要素时空规律，预测演变趋势，成为遥感领域研究热点问题。近日，中国科学院空天信息创新研究院（空天院）牵头研制出",{"type":17,"tag":31,"props":32,"children":33},"strong",{},[34],{"type":23,"value":35},"遥感领域通用预测基础模型“空天•灵眸 时空预测”（RingMo-Sense）",{"type":23,"value":37},"。该模型通过共享学习遥感异构多时相数据，支持多空间、多时间尺度的预测任务。",{"type":17,"tag":25,"props":39,"children":40},{},[41,43,52],{"type":23,"value":42},"研制团队从2021年致力遥感基础模型研究，于2022年发布",{"type":17,"tag":44,"props":45,"children":49},"a",{"href":46,"rel":47},"http://mp.weixin.qq.com/s?__biz=MzkxMTM2MjMzNg==&mid=2247583205&idx=1&sn=60819a40a897e46b36feb321d57aa509&chksm=c11e95aaf6691cbcc9569c9ca6e761d6a1132ecbf593e1f5314b2b9ff11a9dfcae0c7595baa2&scene=21#wechat_redirect",[48],"nofollow",[50],{"type":23,"value":51},"多模态遥感基础模型“空天•灵眸”（RingMo）",{"type":23,"value":53},"，在场景分类、检测定位、要素提取及变化检测等典型感知识别下游任务上展现出其有效性。现在推出的遥感时空预测基础模型RingMo-Sense的核心突破，是在RingMo对目标要素分类定位的“感知识别”基础上，扩展至对目标要素演化规律的“认知预测”。",{"type":17,"tag":25,"props":55,"children":56},{},[57],{"type":17,"tag":31,"props":58,"children":59},{},[60,68,70],{"type":17,"tag":31,"props":61,"children":62},{},[63],{"type":17,"tag":31,"props":64,"children":65},{},[66],{"type":23,"value":67},"空天",{"type":23,"value":69},"•",{"type":17,"tag":31,"props":71,"children":72},{},[73,78,80],{"type":17,"tag":31,"props":74,"children":75},{},[76],{"type":23,"value":77},"灵眸",{"type":23,"value":79}," ",{"type":17,"tag":31,"props":81,"children":82},{},[83],{"type":23,"value":84},"时空预测",{"type":17,"tag":25,"props":86,"children":87},{},[88],{"type":23,"value":89},"RingMo-Sense结合遥感异构多时相数据与多空间、多时间尺度预测任务的特点，创新提出基于时空演化解耦的三支路网络架构，学习空天多时相数据的变化情况，获取稳定泛化的时空通用特征，并在要素分割预测、场景演化预测、数值趋势预测等典型应用任务中进行验证，探索时空预测基础模型在遥感专业领域应用的可能。同时，空天院与华为公司进行深度技术合作，基于昇思MindSpore AI框架与昇腾AI澎湃算力进行RingMo-Sense基础模型设计训练，实现自主创新。",{"type":17,"tag":25,"props":91,"children":92},{},[93],{"type":17,"tag":94,"props":95,"children":98},"img",{"alt":96,"src":97},"image.png","https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20231110062343.11579271464183342679706930390012:50541109072707:2400:1A056F18A1F5A8193823D05DD5C8E7FBE41BD83CCDCA82D8EF67244858B2BE25.png",[],{"type":17,"tag":25,"props":100,"children":101},{},[102],{"type":23,"value":103},"图1空天•灵眸 时空预测",{"type":17,"tag":25,"props":105,"children":106},{},[107],{"type":17,"tag":31,"props":108,"children":109},{},[110],{"type":17,"tag":31,"props":111,"children":112},{},[113],{"type":17,"tag":31,"props":114,"children":115},{},[116],{"type":23,"value":117},"核心能力 创新创造",{"type":17,"tag":25,"props":119,"children":120},{},[121],{"type":17,"tag":31,"props":122,"children":123},{},[124],{"type":23,"value":125},"1、架构创新设计",{"type":17,"tag":25,"props":127,"children":128},{},[129],{"type":23,"value":130},"RingMo-Sense的核心目标是打通遥感多域异构多时相数据输入及多时间、多空间尺度任务输出的信息通路，通过提取稳定泛化特征实现对未来状态的精准预测。",{"type":17,"tag":25,"props":132,"children":133},{},[134],{"type":23,"value":135},"因此，团队提出基于时空演化解耦的三支路网络架构。通过构建时间-空间-时空的三支路网络结合块状（Block-wise）、管状（Tube-wise）、帧状（Frame-wise）的掩码方式，将时空关联解耦，分别学习遥感多时相数据中的空间位置关系、时序动态关系和时空交互关系，并利用渐进式训练策略实现三支路的权重共享和联合优化，以适应不同时间尺度、不同空间尺度的任务特点，达到稳定预测的目的。",{"type":17,"tag":25,"props":137,"children":138},{},[139],{"type":17,"tag":94,"props":140,"children":142},{"alt":96,"src":141},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20231110062510.03384164069537153987333050638765:50541109072707:2400:5AA4A84C17EF5AA1FC0548AB889E9F33D7CE4DB071D2AF8224B91C7F496CEDAB.png",[],{"type":17,"tag":25,"props":144,"children":145},{},[146],{"type":23,"value":147},"图2 基于时空演化解耦的三支路网络架构",{"type":17,"tag":25,"props":149,"children":150},{},[151],{"type":17,"tag":31,"props":152,"children":153},{},[154],{"type":23,"value":155},"2、模型训练优化",{"type":17,"tag":25,"props":157,"children":158},{},[159],{"type":23,"value":160},"遥感预测所涉及的场景分布广泛，为提升模型的鲁棒性和泛化性，RingMo-Sense的训练优化需要规模庞大、丰富度高的遥感多时相数据。",{"type":17,"tag":25,"props":162,"children":163},{},[164],{"type":23,"value":165},"团队收集了来自卫星、航空飞机、无人机等多平台的多时相数据，如高分系列卫星和高空遥感飞机的时序图像、吉林卫星和无人机的视频数据，涵盖了厘米级、米级、数十米级等多种分辨率，数据总量达四百万帧，用以支持RingMo-Sense的训练优化。但不同遥感数据的时间间隔不统一，存在时空演化规律难捕捉的问题。因此团队提出时序变化幅度引导的多尺度数据自适应掩码机制，将时序变化的一阶导数作为超参数，通过掩码自主调节上下帧间的信息流通量，为时序冗余度高或低的数据选取更大或小的掩码比率，使得模型能够处理不同时间间隔的数据，实现对不同遥感数据自适应提取时序特征的能力。",{"type":17,"tag":25,"props":167,"children":168},{},[169],{"type":17,"tag":94,"props":170,"children":172},{"alt":96,"src":171},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20231110062547.95899051329933730216179322981808:50541109072707:2400:C0D55EB6F69E219CEDD70C583540F8B5DF0C4E8F9036D0F79264330EA6543030.png",[],{"type":17,"tag":25,"props":174,"children":175},{},[176],{"type":23,"value":177},"图3 时序变化幅度引导的多尺度数据自适应掩码机制",{"type":17,"tag":25,"props":179,"children":180},{},[181],{"type":17,"tag":31,"props":182,"children":183},{},[184],{"type":23,"value":185},"3、应用任务泛化",{"type":17,"tag":25,"props":187,"children":188},{},[189],{"type":23,"value":190},"时空预测的应用涵盖地物要素分割预测、遥感场景预测、无人机场景预测、数值趋势预测等多种任务，各种应用间差异大，具有时间尺度、空间尺度不一的特点。",{"type":17,"tag":25,"props":192,"children":193},{},[194],{"type":23,"value":195},"RingMo-Sense因其自身架构特点，使得应用任务间的转换非常灵活，基础模型训练结束后，可以将其快速应用到时空预测下游任务中。RingMo-Sense会丢弃掉基础模型训练时采用的解码器，然后将其替换为对应下游预测任务特定的解码器，进而利用下游任务的少量数据对模型进行简单微调。RingMo-Sense在多个时空预测任务上取得了有竞争力的效果。",{"type":17,"tag":25,"props":197,"children":198},{},[199],{"type":17,"tag":94,"props":200,"children":202},{"alt":96,"src":201},"https://fileserver.developer.huaweicloud.com/FileServer/getFile/cmtybbs/e64/154/b38/90a1d5d431e64154b387b3660e356ff5.20231110062620.48865566907306364745336345103593:50541109072707:2400:B9E1926316185B1889398AB7F6A90D6E4705F3A061A3AE1972DD41A452CCA8D2.png",[],{"type":17,"tag":25,"props":204,"children":205},{},[206],{"type":23,"value":207},"图4 空天•灵眸 时空预测下游应用",{"type":17,"tag":25,"props":209,"children":210},{},[211],{"type":17,"tag":31,"props":212,"children":213},{},[214],{"type":23,"value":215},"4、昇腾AI基础软硬件适配",{"type":17,"tag":25,"props":217,"children":218},{},[219],{"type":23,"value":220},"空天院与华为深度合作，由北京昇腾人工智能生态创新中心提供技术保障，在昇腾AI澎湃算力上，利用昇思MindSporeAI框架对RingMo-Sense基础模型和下游任务完成了昇腾AI基础软硬件适配，并针对时序数据训练方面进行性能优化，为各行各业研究者基于昇腾AI基础软硬件平台进行遥感时序模型训练与开发提供有力支撑。",{"type":17,"tag":25,"props":222,"children":223},{},[224,226,231],{"type":23,"value":225},"目前，“空天•灵眸”时空预测基础模型的",{"type":17,"tag":31,"props":227,"children":228},{},[229],{"type":23,"value":230},"核心成果已在遥感领域权威期刊IEEE Transactions on Geoscience and Remote Sensing公开发表",{"type":23,"value":232},"。同时，该模型已在住建交通、国土资源等行业开展试用，支持对重点目标、地物要素的时空规律分析，实现精准认知预测。",{"type":17,"tag":25,"props":234,"children":235},{},[236],{"type":17,"tag":31,"props":237,"children":238},{},[239],{"type":23,"value":240},"相关论文信息",{"type":17,"tag":25,"props":242,"children":243},{},[244],{"type":23,"value":245},"F. Yao et al., \"RingMo-Sense: Remote Sensing Foundation Model for Spatiotemporal Prediction via Spatiotemporal Evolution Disentangling,\" in IEEE Transactions on Geoscience and Remote Sensing, vol. 61, pp. 1-21, 2023, Art no. 5620821, doi: 10.1109/TGRS.2023.3316166.",{"type":17,"tag":25,"props":247,"children":248},{},[249],{"type":17,"tag":31,"props":250,"children":251},{},[252],{"type":23,"value":253},"论文链接",{"type":17,"tag":25,"props":255,"children":256},{},[257],{"type":17,"tag":44,"props":258,"children":261},{"href":259,"rel":260},"https://ieeexplore.ieee.org/document/10254320",[48],[262],{"type":23,"value":259},{"type":17,"tag":25,"props":264,"children":265},{},[266],{"type":23,"value":267},"以上内容由地理与赛博空间信息技术研究部赛博智能团队提供。",{"title":7,"searchDepth":269,"depth":269,"links":270},4,[],"markdown","content:news:zh:2859.md","content","news/zh/2859.md","news/zh/2859","md",1776506074300]