内容来源：https://www.geekwire.com/2026/terrabyte-ai-stealth-geospatial-data/

内容总结：

西雅图初创公司TerraByte AI发布“地球搜索引擎”，用AI实时解析卫星数据

一家名为TerraByte AI的低调西雅图初创公司日前推出了一款人工智能软件平台，能够实时筛选卫星数据，从中挖掘具有地理空间价值的信息。该公司的“地球搜索引擎”可分析连续的卫星影像，识别感兴趣的目标，并通过自然语言查询将信息串联起来。该平台的核心优势在于，其数据集无需经过耗时且昂贵的人工标注过程。

“我们采用自监督学习技术，在不进行人工标注的情况下理解像素信息，”公司联合创始人兼首席执行官Rishi Madhok向科技媒体GeekWire表示。他举例说明了平台的多项应用场景：识别电力线分段、查找高速公路附近没有电动汽车充电桩的停车场、监控进入港口的集装箱船。“如果你想监测霍尔木兹海峡，可以用我们的模型来实现。亚利桑那州的森林砍伐区域、露天采矿——这些都是截然不同的概念，但我们的模型能够理解它们，因为它是一个基础模型。”

本周，TerraByte正在阿拉巴马州亨茨维尔举行的欧空局-美国宇航局地球观测AI模型研讨会上展示其技术方案。一场实操环节定于周三举行，公司联合创始人兼首席技术官Fuxun Yu将于周四发表口头演讲。

Madhok与Yu于去年共同创立TerraByte，业务横跨西雅图和旧金山。Madhok此前曾在微软行星计算机项目领导地理空间AI计划，而Yu则担任微软首席研究经理，并主导了公司的地理空间基础模型项目。

在融资方面，Madhok透露公司已获得Ascend、PSL Ventures及部分天使投资者的种子轮前期投资，但未透露具体金额。Ascend创始管理合伙人Kirby Winfield在邮件声明中表示，TerraByte“正在构建卫星智能领域的基础模型层……他们打造的API和AI基础设施，将支撑下一代位置智能应用。”

Pioneer Square Labs 董事总经理、PSL Ventures 普通合伙人 Vivek Ladsariya 表示，之所以投资 TerraByte，“是因为地理空间数据是企业技术领域最具影响力但服务最不充分的类别之一。地理空间数据量呈指数级增长，但大多数组织仍然无法有效获取或利用这些数据。Rishi和Fuxun是罕见的实干家——他们亲身经历过这一难题，知道如何构建解锁数据价值的基础设施层。TerraByte正是这一层，我们认为它将变得具有基础性意义。”

Madhok指出，虽然目前已有单传感器基础模型用于分析卫星公司采集的广谱传感器数据，但“TerraByte是首个原生融合光学、合成孔径雷达、热红外和高光谱数据的基础模型层”。这种多传感器方法能支持对自然灾害等快速变化场景的迅速响应。

“例如，我们可以与公用事业公司或应急救援人员合作。当野火发生时，他们可以提问：‘向我展示加州或华盛顿州活跃野火一英里范围内的居民区和电力线分段。’”Madhok解释说，“因为灾害发生时，必须能实时即时监控，而目前能提供这种洞察的技术并不多。”

风险评估是另一个潜在应用方向，例如帮助保险公司评估野火风险。“如果能通过识别植被侵占等高风险因素来准确评估风险，就能在这些方面制定更合理的定价。”他说。

Madhok表示，该软件还可安装在轨道卫星上，在下传数据前进行过滤。这能最大限度地减少从地球观测卫星或轨道数据中心下传海量数据所造成的时间延迟和成本。“我们的目标是构建一个既能在地面站运行、也能在边缘端运行的模型，从而支持所有对时间敏感的应用进行监控。”他透露公司已在筹备一次在轨演示。

在商业策略方面，TerraByte仍在完善中，当前重点面向企业客户。“我们初创阶段更聚焦B2B模式，正在与大型企业洽谈许可协议，但最终会转向订阅制。”Madhok说。

当前市场上并非只有TerraByte一家提供地理空间数据分析服务，其他参与者包括Google Earth Engine和BlackSky Spectra。此外，Starcloud、Sophia Space等公司也在计划将计算能力部署到太空边缘。面对竞争，Madhok表示：“我认为Sophia Space等许多公司是合作伙伴。那些正在建造卫星或采集影像的公司，各自擅长自己的领域——建造卫星或将计算机送入太空。而我们的DNA是打造最优秀的地理空间模型，这就是我们的目标。”

中文翻译：

西雅图一家名为TerraByte AI的初创公司正低调亮相，推出了一款软件平台，该平台利用人工智能从实时卫星数据中筛选出地理空间领域的宝贵信息。
TerraByte的“地球搜索引擎”分析卫星图像流，识别感兴趣的特征，并通过自然语言查询将各信息点串联起来。该平台的关键优势在于，其数据集无需经过耗时且昂贵的人工标注过程。
公司联合创始人兼首席执行官里希·马多克向GeekWire表示：“我们采用自监督学习技术，本质上是理解像素含义而不必手动标注。”
他解释道：“这项技术应用广泛，例如识别输电线段、寻找高速公路旁没有电动汽车充电桩的停车场、监控进入港口的集装箱船等。如果你想监测霍尔木兹海峡，也可以使用我们的模型。森林砍伐区域、亚利桑那州的露天采矿——这些都是截然不同的概念，但我们的模型能够理解它们，因为它是一个基础模型。”
本周，TerraByte在阿拉巴马州亨茨维尔举行的欧空局-美国国家航空航天局地球观测人工智能模型研讨会上展示了其技术方案。周三安排了实践操作环节，公司联合创始人兼首席技术官余福讯定于周四进行口头报告。
马多克和余福讯于去年创立了TerraByte，业务运营分布在西雅图和旧金山两地。马多克此前在微软行星计算机项目领导地理空间人工智能计划，而余福讯曾任微软首席研究经理，并主导了公司的地理空间基础模型项目。
在财务方面，马多克透露公司已获得Ascend、PSL Ventures及天使投资者的种子轮前融资，但未说明具体投资金额。
Ascend创始管理合伙人柯比·温菲尔德在一封邮件声明中表示，TerraByte“正在构建卫星智能的基础模型层……他们创建的API和人工智能基础设施，将为下一代位置智能应用提供支撑。”
Pioneer Square Labs董事总经理兼PSL Ventures普通合伙人维韦克·拉达萨里亚表示，其基金投资TerraByte“是因为地理空间数据是企业技术领域中最具影响力且服务最不足的类别之一。”
拉达萨里亚在邮件中说：“地理空间数据量呈指数级增长，但大多数组织仍无法有效获取或利用这些数据。里什和余福讯是罕见的实干家——他们亲身经历过这一难题，并清楚如何构建解锁数据价值的基础设施层。TerraByte就是这一层，我们相信它将成为行业基石。”
尽管目前已有针对卫星公司采集的宽光谱传感器数据的单传感器基础模型，但马多克指出，“TerraByte是首个将光学、合成孔径雷达、热红外和高光谱数据原生融合于一个基础模型的模型层。”这种多传感器方法可支持对自然灾害等快速变化场景的快速响应。
马多克说：“例如，我们可以与电力公司甚至一线救援人员合作。发生野火时，他们可以查询‘显示加州或华盛顿州活跃野火周边一英里内的社区和输电线段’。因为灾害发生时，需要实时监控，而现有技术很少能提供这种洞察力。”
风险评估是另一个潜在应用场景——例如解决保险公司评估野火风险时面临的挑战。马多克表示：“如果能通过分析植被侵占等高危因素准确评估风险，就能为相关保险产品制定更合理的定价。”
此外，该软件还可安装在轨道卫星上，在数据下行前进行筛选。这能最大限度减少地球观测卫星或轨道数据中心下行海量数据所需的时间延迟和成本。
马多克说：“我们的目标是构建一个既能在地面站运行，也能在边缘端运行的模型，从而支持所有时间敏感型应用的实时监控。”他表示公司已在筹备在轨演示计划。
尽管TerraByte仍在完善商业战略，但当前重点是服务企业客户。马多克说：“我们初创阶段更聚焦B2B模式，因此正与企业客户洽谈大规模许可证，但后续将转向订阅制。”
TerraByte并非唯一提供地理空间数据分析的公司。市场中的其他参与者包括Google Earth Engine和BlackSky Spectra。还有Starcloud、Sophia Space等公司正计划在太空边缘部署计算能力。TerraByte能否与之竞争？
马多克表示：“我将Sophia Space及类似拥有太空计算能力的公司视为合作伙伴。许多公司——无论是制造卫星还是采集图像的——都精于本职：建造卫星或将计算机送入太空。而我们的基因是打造最出色的地理空间模型，这正是我们的目标。”

英文来源：

A stealthy Seattle startup called TerraByte AI is unveiling a software platform that uses artificial intelligence to sift through real-time satellite data for geospatial gems.
TerraByte’s “Earth Search Engine” analyzes streams of satellite imagery, recognizes features of interest and connects the dots through natural-language queries. The platform’s key advantage is that its data set doesn’t have to go through the laborious, expensive process of manual annotation.
“We’re just using self-supervised learning techniques to essentially understand the pixels without having to manually annotate it,” CEO and co-founder Rishi Madhok told GeekWire.
“There are many applications that you can do, like identifying power-line segments, finding parking lots near highways without EV charging stalls, watching container ships entering port,” he explained. “If you want to monitor the Strait of Hormuz, you can use our models to do that. Deforestation areas, open-pit mining in Arizona — all of these are very different concepts, but our model is able to understand them because it’s a foundational model.”
TerraByte is laying out its approach this week in Huntsville, Ala., at an ESA-NASA workshop on AI models for Earth observation. A hands-on session is scheduled on Wednesday, and the company’s co-founder and chief technology officer, Fuxun Yu, is due to make an oral presentation on Thursday.
Madhok and Yu founded TerraByte last year, with operations split between Seattle and San Francisco. Madhok previously led geospatial AI initiatives at Microsoft Planetary Computer, while Yu worked as a principal research manager at Microsoft and led the company’s Geospatial Foundational Model project.
On the financial front, Madhok said the venture has received pre-seed funding from Ascend, PSL Ventures and angel investors, though he declined to specify the size of the investment.
Kirby Winfield, founding managing director of Ascend, said in an emailed statement that TerraByte “is building the foundation model layer for satellite intelligence. … They’re creating the foundational API and AI infrastructure that will power the next generation of location intelligence applications.”
Vivek Ladsariya, managing director at Pioneer Square Labs and general partner of PSL Ventures, said his fund invested in TerraByte “because geospatial data is one of the most consequential and underserved categories in enterprise tech.”
“The volume of geospatial data is growing exponentially, yet most organizations still can’t access or act on it effectively,” Ladsariya said via email. “Rishi and Fuxun are rare operators — they’ve lived this problem firsthand and know exactly what it takes to build the infrastructure layer that unlocks it. TerraByte is that layer, and we think it becomes foundational.”
While single-sensor foundation models already exist to analyze the wide spectrum of sensor data captured by satellite companies, Madhok noted that “TerraByte is the first model layer to natively fuse optical, synthetic aperture radar, thermal and hyperspectral in one foundation model.” That multi-sensor approach could support a rapid response to fast-changing situations such as natural disasters.
“For example, we can work with utility companies or even with first responders, and when there’s a wildfire, they can ask for things like, ‘Show me neighborhoods and power-line segments within one mile of an active wildfire in California or Washington,'” Madhok said. “Because when catastrophes happen, you need to be able to monitor them instantly in real time, and there aren’t a lot of technologies out there which can actually give them the insights.”
Risk assessment is another potential application — for example, addressing the challenges that insurance companies face in evaluating wildfire risk. “If they’re able to characterize the risk well by understanding if there are high-risk things like vegetation encroachment, you’re able to have better pricing on some of those things,” he said.
Madhok said the software could also be installed on orbiting satellites to filter data before it’s downlinked. That could minimize the time delay and costs that might otherwise be associated with downlinking massive amounts of data from an Earth observation satellite or an orbital data center.
“Our goal is to build a model which not only works on the ground station, but also works on the edge, so that it helps all these time-sensitive applications to monitor things,” Madhok said. He said the company is already working on arrangements for an on-orbit demonstration.
While TerraByte is still refining its commercial strategy, the immediate focus is on enterprise clients. “Our business startup use case is more focused on B2B, so we’re working with large enterprise licenses, but eventually it will be subscription-based,” Madhok said.
TerraByte isn’t the only company offering geospatial data analysis. Other players in the market include Google Earth Engine and BlackSky Spectra. Still more companies, such as Starcloud and Sophia Space, are working on plans to put computing power on the edge in space. Will TerraByte be able to compete?
“I see Sophia Space and a lot of similar companies who have compute up in space as partners,” Madhok said. “A lot of the companies out there, whether they’re building satellites or collecting imagery, are really good at doing what they do — which is building satellites or sending computers up into space. Our DNA is to build the best geospatial models out there, and that’s what our goal is.”