EAGER: IMPRESS-U: Exploratory Research on Generative Compression for Compressive Lidar

这个IMPRESS-U项目由NSF、波兰国家科学中心、美国国家科学院和全球海军研究办公室(DoD)联合资助。这项研究将在一个多边国际伙伴关系中进行，该伙伴关系由美国特拉华大学、乌克兰哈尔科夫国立航空航天大学和波兰什切钦西博美拉尼亚理工大学联合开展。合作努力的美国部分将由国际科学和工程办公室(OISE)、激励竞争研究的既定计划(EPSCoR)以及通信和信息基础计划(CCF)共同资助。拟议的工作旨在：(A)在乌克兰、波兰和美国的学术研究团队之间建立合作伙伴关系；(B)在乌克兰建立一个在卫星激光雷达地球遥感方面具有弹性和协作性的卓越研究和教育计划；以及(C)探索用于存储或传输压缩激光雷达测量数据的生成性压缩的全新概念。这些方法将应用于从新一代卫星激光雷达--压缩激光雷达(CS激光雷达)--获得的数据，该雷达将用于揭示对生态系统过程具有深远影响的地球表面及其森林的拓扑结构。今天的星载激光雷达在空间分辨率和覆盖范围上是有限的，因为激光反射只能沿着一维足迹线扫描进行测量。压缩激光雷达从地球上空数百公里处对地球进行稀疏测量，然后通过计算重建具有分辨率和覆盖范围的3D图像，就好像数据是从数百米高的地方收集的一样。到目前为止，卫星激光雷达不使用数据压缩来避免信息丢失。然而，CS激光雷达依赖于深度学习重建算法，覆盖了地球上更大范围内自然出现数据压缩机会的数量级区域。因此，该项目将使用生成性模型从根本上探索数据压缩的新方法，这些模型显示了在更深的压缩级别下产生更准确的图像重建的潜力。该项目将在乌克兰和波兰招募一批新的人才，他们将受到激励，开始围绕机器学习和地球遥感的交叉领域从事持久的职业生涯。星载激光雷达是一项重要的成像技术，用于揭示地球表面及其森林的拓扑结构，这些拓扑结构对生态系统过程产生深远影响，这些过程决定了地球上的养分、水和碳循环。今天的星载激光雷达在空间分辨率和覆盖范围上是有限的，因为激光反射只能沿着一维足迹线扫描进行测量。在这两者之间，大量的景观仍然没有采样照明。为了克服这一限制，NASA正在开发新一代卫星激光雷达，采用一种全新的传感模式，传统的一维线扫描被稀疏和大视场照明的压缩卫星(CS)激光雷达所取代。目标是从地球上空数百公里处压缩感知地球，然后通过计算重建具有分辨率和覆盖范围的3D图像，就好像数据是从数百米高的地方收集的一样。NASA的卫星激光雷达迄今没有使用测量数据压缩，以避免信息丢失。然而，CS激光雷达依赖于深度学习重建算法，覆盖了地球上更大范围内自然出现数据压缩机会的数量级区域。NASA的CS激光雷达团队目前没有探索这个问题，因此拟议的探索性研究工作是有价值的、补充的和及时的。虽然传统的图像压缩依赖于手工制作的编码器/解码器对，但研究团队将使用生成性模型探索全新的数据压缩方法，这些模型显示出在更深的压缩级别下产生更准确的图像重建的潜力。CS激光雷达承诺显著提高卫星测高的视场和成像分辨率，其中数据压缩变得越来越重要。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

This IMPRESS-U project is jointly funded by NSF, National Science Center of Poland, US National Academy of Sciences, and Office of Naval Research Global (DoD). The research will be performed in a multilateral international partnership that unites the University of Delaware, US, the National Aerospace University in Kharkiv, Ukraine, and the West Pomeranian University of Technology in Szczecin, Poland. US portion of the collaborative effort will be co-funded by Office of International Science and Engineering (OISE), Established Program to Stimulate Competitive Research (EPSCoR), and Communications and Information Foundations Program (CCF). The proposed effort aims at: (a) establishing a partnership among academic research teams from Ukraine, Poland, and the US; (b) building a resilient and collaborative research and education program of excellence in Ukraine in machine learning (ML) for satellite lidar sensing of Earth; and (c) exploring radically new concepts in generative compression for the storage or communication of compressive lidar measurements. The methods will be applied to data obtained from a new generation of satellite lidars, coined compressive lidars (CS lidar), that will be used to unravel the topological structure of the Earth’s surface and its forests, which have a profound effect on ecosystem processes. Spaceborne lidars today are limited in spatial resolution and coverage since laser reflections are only measured along 1D footprint line scans. Compressive lidars take sparse measurements of Earth from hundreds of km above Earth to then computationally reconstruct the 3D imagery with resolution and coverage, as if the data was collected from just hundreds of meters in height. To date, satellite lidars do not use data compression to avoid loss of information. CS lidars, however, rely on deep learning reconstruction algorithms covering orders of magnitude larger areas of Earth where the opportunity of data compression arises naturally. This project will thus explore radically new approaches to data compression using generative models which have shown the potential to produce more accurate image reconstructions at much deeper compression levels. The project will recruit a fresh cohort of talent in Ukraine and Poland who will be galvanized to embark on enduring careers that revolve around the intersection of machine learning and Earth remote sensing. Spaceborne lidar is an important imaging technology that is used to unravel the topological structure of the Earth’s surface and its forests which have a profound effect on ecosystem processes that determine nutrient, water, and carbon cycles on Earth. Spaceborne lidars today are limited in spatial resolution and coverage since laser reflections are only measured along 1D footprint line scans. In between these, vast amounts of landscape remain without sampling illumination. To overcome this limitation, a new generation of satellite lidars are being developed at NASA taking on a radically new sensing paradigm where the traditional 1D line scanning is abandoned and replaced by sparse and wide-field-of view lidar illumination, coined compressive satellite (CS) lidars. The objective is to compressively sense Earth from hundreds of km above Earth to then computationally reconstruct the 3D imagery with resolution and coverage, as if the data was collected from just hundreds of meters in height. NASA’s satellite lidars to date do not use data compression of the measurements to avoid loss of information. CS lidar, however, relies on deep learning reconstruction algorithms covering orders of magnitude larger areas of Earth where the opportunity of data compression arises naturally. NASA’s CS lidar team is currently not exploring this problem and thus the proposed exploratory research effort is valuable, complementary, and timely. While traditional image compression relies on hand-crafted encoder/decoder pairs, the research team will explore radically new approaches to data compression using generative models which have shown the potential to produce more accurate image reconstructions at much deeper compression levels. CS lidars promise to significantly enhance both the field-of-view and imaging resolution of satellite altimetry where data compression becomes increasingly important.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.