Learning an urban grammar from satellite data through AI

该项目将提出一种描述城市形态的城市语法，并将开发人工智能（AI）技术，从卫星图像中学习这种语法。城市形态对经济生产力、社会（不）平等以及地方财政和环境的可持续性具有重要影响。然而，目前衡量城市形态的方法是分散和粗糙的，阻碍了它们在决策和规划中的适当使用。该项目旨在：1）概念化城市语法，将城市形态描述为“空间特征”的组合，可计算的类别描述了城市发展的独特空间模式（例如“碎片化低密度”，“紧凑有机”，“规则密集”）; 2）开发数据驱动的空间特征类型学作为构建模块; 3）创建可以从卫星图像中学习特征的人工智能技术; 4）从空间特征的高分辨率轨迹中构建英国的可计算城市语法，帮助我们了解其未来的演变。该项目建议通过利用卫星数据源和最先进的机器学习和人工智能技术，使概念城市语法变得可计算。卫星技术正在经历一场革命，为研究社会挑战提供更多更好的数据。然而，卫星数据的潜力只能通过应用精细的机器学习和人工智能算法来释放。在此背景下，我们将结合联合收割机地理人口统计学、深度学习、迁移学习、序列分析和递归神经网络。这些方法扩展和补充了社会科学中使用的传统技术，允许从高度非结构化的数据（如图像）中提取洞察力。在此过程中，该项目的方法学方面将开发出为社会科学领域的其他应用奠定基础的方法。该项目的框架分为四个主要阶段，或工作包（WP）：1）数据获取-两大数据集将被汇集在一起，并在空间上对齐在一个一致的数据库中：城市形态属性和卫星图像。2）空间特征类型学的发展--使用城市形态属性，地理人口统计学将用于为英国建立高空间分辨率的空间特征类型。3）卫星图像+人工智能-该类型将用于训练深度学习和迁移学习算法，以自动识别空间特征，并以可扩展的方式从中等分辨率卫星图像中识别空间特征，这将使我们能够将这种方法追溯到过去三十年的图像。4）轨迹分析-使用前一个包中生成的空间签名序列，我们将使用机器学习通过研究过去三十年英国城市形态的演变来识别城市语法。学术成果包括期刊文章，开源软件，和开放数据产品，以努力接触尽可能广泛的学术受众，并使交付渠道多样化，以便产出在各种情况下提供价值。影响力战略围绕两个主要领域构建：通过双向传播与利益攸关方建立持续沟通;传播数据见解，确保所产生的数据和证据到达预期用户手中。

This project will propose an urban grammar to describe urban form and will develop artificial intelligence (AI) techniques to learn such a grammar from satellite imagery. Urban form has critical implications for economic productivity, social (in)equality, and the sustainability of both local finances and the environment. Yet, current approaches to measuring the morphology of cities are fragmented and coarse, impeding their appropriate use in decision making and planning. This project will aim to: 1) conceptualise an urban grammar to describe urban form as a combination of "spatial signatures", computable classes describing a unique spatial pattern of urban development (e.g. "fragmented low density", "compact organic", "regular dense"); 2) develop a data-driven typology of spatial signatures as building blocks; 3) create AI techniques that can learn signatures from satellite imagery; and 4) build a computable urban grammar of the UK from high-resolution trajectories of spatial signatures that helps us understand its future evolution.This project proposes to make the conceptual urban grammar computable by leveraging satellite data sources and state-of-the-art machine learning and AI techniques. Satellite technology is undergoing a revolution that is making more and better data available to study societal challenges. However, the potential of satellite data can only be unlocked through the application of refined machine learning and AI algorithms. In this context, we will combine geodemographics, deep learning, transfer learning, sequence analysis, and recurrent neural networks. These approaches expand and complement traditional techniques used in the social sciences by allowing to extract insight from highly unstructured data such as images. In doing so, the methodological aspect of the project will develop methods that will set the foundations of other applications in the social sciences.The framework of the project unfolds in four main stages, or work packages (WPs):1) Data acquisition - two large sets of data will be brought together and spatially aligned in a consistent database: attributes of urban form, and satellite imagery.2) Development of a typology of spatial signatures - Using the urban form attributes, geodemographics will be used to build a typology of spatial signatures for the UK at high spatial resolution.3) Satellite imagery + AI - The typology will be used to train deep learning and transfer learning algorithms to identify spatial signatures automatically and in a scalable way from medium resolution satellite imagery, which will allow us to back cast this approach to imagery from the last three decades.4) Trajectory analysis - Using sequences of spatial signatures generated in the previous package, we will use machine learning to identify an urban grammar by studying the evolution of urban form in the UK over the last three decades.Academic outputs include journal articles, open source software, and open data products in an effort to reach as wide of an academic audience as possible, and to diversify the delivery channel so that outputs provide value in a range of contexts. The impact strategy is structured around two main areas: establishing constant communication with stakeholders through bi-directional dissemination; and data insights broadcast, which will ensure the data and evidence generated reach their intended users.