III: Medium: Collaborative Research: Exploiting Context in Cartographic Evolutionary Documents to Extract and Build Linked Spatial-Temporal Datasets

如今，数以百万计的历史地图保存在数字档案中。例如，美国地质调查局已经创建并扫描了超过20万张地形图，覆盖了125年的时间。地图是一种“进化的视觉文件”，因为它们显示了长时间和大面积的景观变化。这些文件具有巨大的价值，因为它们提供了一个高分辨率的窗口，可以在大陆范围内了解过去。不幸的是，没有时间密集的人工数字化扫描地图是无法用于研究目的的。地图特征，如湿地和道路，虽然人类可以阅读，但只能以图像的形式提供。这个跨学科的合作项目涉及南加州大学和科罗拉多大学博尔德分校的研究人员和他们的学生，他们将开发一套开源技术和工具，允许用户从大量的地图册中提取地图特征，并在地理信息系统中跟踪地图版本之间特征的变化。由此产生的开源工具将使历史、人口统计学、经济学、社会学、生态学和其他学科的研究和学习成为令人兴奋的新形式。该项目产生的数据将通过案例研究与其他历史档案相结合，向公众提供。125多年来，涵盖人造和自然特征的时空关联知识为物理科学和社会科学提供了巨大的潜力。这些地图中包含的丰富信息是独一无二的，特别是在航空摄影广泛使用之前。将存储在大型档案中的扫描纸质地图自动转换为时空关联知识的能力，将为社会和自然科学家研究全球变化和其他在大面积和长时间内发挥作用的社会地理过程创造重要资源。该项目的研究目标是开发一个识别和数据集成框架，该框架可以提取、组织和链接随着时间的推移而演变的视觉文档（如地图系列）中的知识。虽然过去的工作主要集中在从单个条件良好的地图图像中提取特征，但该框架将处理大量历史地图档案，以高效、可靠地提取人造和自然特征，并将这些特征跨时间（地图版本）、空间（地图页）和比例联系起来。该框架将通过利用链接知识形式的上下文信息，对图形质量较差的地图进行识别。这些上下文信息来自现有的空间数据源或从最近的高质量地图版本中提取，可用于改进和完善自动处理存档地图的训练步骤。该框架还利用特征之间语义关系的知识来提高所开发方法的鲁棒性、效率和自动化程度，并表征提取数据中的不确定性，以及在空间、时间和规模上提取数据之间的链接。该研究项目将通过案例研究来验证这些方法，这些案例研究将评估从美国地质调查局和地形测量局地图中提取的主要特征类型（建成区、基础设施、水文和植被）的完整关联数据集。研究人员将使用多个研究区域，这些区域代表了城市化及其对农村和野生景观的影响等过程所驱动的景观演变和转变的不同历史（例如I-95大都市城市走廊）。该项目的出版物、软件和数据集将在项目网站（http://spatial-computing.github.io/unlocking-spatiotemporal-map-data）上提供。

Millions of historical maps are in digital archives today. For example, the U.S. Geological Survey has created and scanned over 200,000 topographic maps covering a 125-year period. Maps are a form of "evolutionary visual documents" because they display landscape changes over long periods of time and across large areas. Such documents are of tremendous value because they provide a high-resolution window into the past at a continental scale. Unfortunately, without time-intensive manual digitization scanned maps are unusable for research purposes. Map features, such as wetlands and roads, while readable by humans, are only available as images. This interdisciplinary collaborative project involving researchers and their students at University of Southern California and University of Colorado, Boulder will develop a set of open-source technologies and tools that allow users to extract map features from a large number of map sheets and track changes of features between map editions in a Geographical Information System. The resulting open-source tools will enable exciting new forms of research and learning in history, demography, economics, sociology, ecology, and other disciplines. The data produced by this project will be made publically available and through case studies integrated with other historical archives. Spatially and temporally linked knowledge covering man-made and natural features over more than 125 years holds enormous potential for the physical and social sciences. The wealth of information contained in these maps is unique, especially for the time before the widespread use of aerial photography. The ability to automatically transform the scanned paper maps stored in large archives into spatio-temporally linked knowledge will create an important resource for social and natural scientists studying global change and other socio-geographic processes that play out over large areas and long periods of time. The research goal of this project is to develop a recognition and data integration framework that extracts, organizes, and links the knowledge found in visual documents that evolve over time, such as a map series. While past work has focused on feature extraction from single well-conditioned map images, this framework will handle large volume historical map archives for efficient, robust extraction of man-made and natural features and link the features across time (map editions), space (map sheets), and scale. The framework will perform recognition in maps with poor graphical quality by exploiting contextual information in the form of linked knowledge. This contextual information comes from existing spatial data sources or has been extracted from more recent high-quality map editions, which can be used to improve and refine the training steps for automatically processing maps in an archive. The framework also exploits knowledge of the semantic relationships between features to increase robustness, efficiency, and the degree of automation of the methods developed and characterize uncertainty in the extracted data as well as in linking between extracted data across space, time, and scale. This research project will validate the methods by using case studies that evaluate the extracted, fully linked data collections for major feature types (built-up area, infrastructure, hydrography and vegetation) from both the USGS and Ordnance Survey maps. The researchers will use multiple study regions that represent different histories in landscape evolution and transitions driven by processes such as urbanization and its effects on rural and wild landscapes (e.g., the I-95 megapolitan urban corridor). Publications, software, and datasets for this project will be made available on the project website (http://spatial-computing.github.io/unlocking-spatiotemporal-map-data).