一支研究团队在两个在轨平台上成功验证了 NASA 与 IBM 联合开发的开源地理空间人工智能基础模型 Prithvi。
A team of researchers demonstrated NASA and IBM’s open-source Prithvi Geospatial artificial intelligence foundation model aboard two in-orbit platforms.
研究团队选择Prithvi开展其研究,一方面因其在地球观测任务中展现出强大的泛化能力,另一方面则因其作为开源模型的可获取性。基础模型(foundation model)是一种在海量无标签数据上训练而成的人工智能模型,使其能够初步识别出人类难以自行察觉的数据模式;随后,该模型仅需少量标注数据即可针对特定应用场景进行微调。地球观测卫星持续采集关于地球的海量数据。若能在卫星将数据传回地面之前,在轨完成数据处理与分析,将有助于研究人员更快获得洞见。然而,由于带宽限制,现役卫星通常无法接收大型软件更新,因此其搭载的数据分析AI模型往往轻量化且高度专用。研究人员可借助基础模型的灵活性,在单一软件架构内支持广泛多样的地球观测任务。一旦卫星入轨后需新增任务,研究人员仅需上传一个体积较小的附加解码器包——所需带宽远低于向卫星上传一整套新模型。将Prithvi送入轨道,是基础模型有望变革地球观测领域的早期示范之一。除数据处理分析外,基础模型未来还可能助力科学家与数据采集仪器直接交互。“大语言模型(large language model)同样属于基础模型的一种,”杜博士指出,“未来,这或将使操作人员得以使用自然语言与卫星交互,就星载数据或系统状态提出问题,并以对话形式获得回应。”
The team chose Prithvi for their research because of its strong generalization across Earth observation tasks, and because of its availability as an open-source model. A foundation model is an AI model trained on an enormous amount of unlabeled data, which allows the model to begin detecting patterns in the data that humans wouldn’t notice on their own. The model can then be fine-tuned for specific applications using much smaller amounts of labeled data. Earth-observing satellites collect enormous amounts of data about our planet. Processing and analyzing the data in orbit before the satellite sends it back to Earth can help researchers gain insights more quickly. However, active satellites often can’t accept large software updates because of bandwidth limits, so the AI models they carry for data analysis tend to be lightweight and highly specialized. Researchers can use the flexibility of a foundation model to facilitate a wide range of Earth observation tasks in one software architecture. If they want the model to take on a new task once the satellite is in orbit, they only need to upload a small extra decoder package – using far less bandwidth than uploading a whole new model to the satellite. Sending Prithvi to orbit is an early demonstration of how foundation models could transform Earth observation. In addition to data analysis, foundation models could eventually help scientists interact with the instruments collecting the data. “A large language model is also a type of foundation model,” Du said. “In the future, this could allow operators to interact with satellites in natural language, asking questions about onboard data or system status and receiving responses in a conversational way.”