GeoAI for Terrain Analysis: A Deep-Learning Approach for Landform Feature Detection

地理空间人工智能(GeoAI)代表着一个令人兴奋的新研究领域，它结合了机器智能和数据驱动的方法来解决地理空间问题。人工智能方法的快速进步、空间大数据的激增和巨大的计算能力正在改变研究的进行方式，并促使新的发现。该项目开发了一种新的地理空间人工智能(GeoAI)解决方案，以实现大规模、自动化、智能化和高精度的地貌特征检测和地形分析。传统的地形分析方法仅限于基于像素/对象的图像分析和浅层机器学习，在处理复杂分类任务中的大数据时，这些方法面临着巨大的性能挑战。这项研究通过结合空间原理和空间关系数据来空间化深度(机器)学习，代表了GeoAI和更广泛的空间数据科学在方法论上的突破。利用GeoAI进行地貌特征识别丰富了空间知识，增强了对地球和其他星球陆面过程的理解。它还有利于具有社会效益的地理空间应用，包括可用于搜索和救援行动的异常检测，以及通过识别指示环境变化的地形特征。研究人员打算使用过去用于建立地理信息科学(GISci)社区的模式来发展GeoAI社区。将组织与GeoAI相关的研讨会，作为来自不同学科和组织的研究人员讨论GeoAI的新进展和开放挑战的重要场所。该项目包括一名博士后学者，并将本科生培养为跨学科科学家。这两名研究人员都是女性科学家，他们积极指导代表人数不足的群体中的科学家。在这个项目中开发的所有数据和代码都将是开源的，以造福于更广泛的地理空间社区。成功地将GeoAI应用于地形分析存在重大挑战，包括地形特征的复杂性和多样性、训练数据的缺乏、模型设计中空间知识的缺乏以及对机器推理过程的有限理解。这项研究将通过开发(1)全面的地形数据集GeoNAT来支持地形分析和各种机器学习任务来应对这些挑战；(2)机器学习模型Terrain AI，它注入了关键的空间原理(空间自相关性和空间异质性)，以实现从多源、地理参考数据的跨尺度深度学习；以及(3)交互式可视化工具，它打开了机器学习和决策过程的“黑匣子”。依靠这些工具，将回答三个研究问题：(I)机器学习以区分地貌特征的独特空间结构、模式和空间尺度是什么？(2)人类识别过程与机器识别过程相比如何？以及(Iii)如何预测在不同景观中产生某种形式特征的潜在地貌过程？确定过程-形式关系对地貌学等地理、空间科学和相关科学的议程产生重大影响，以促进建立社区共识的地貌分类系统。Terrain AI模型不仅限于地貌研究，而且可以推广并适用于检测任何地理对象，包括自然和人造对象。该奖项反映了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。

Geospatial Artificial Intelligence (GeoAI) represents an exciting new research area that incorporates machine intelligence and data-driven approaches for geospatial problem solving. Rapid advances in AI methods, the proliferation of spatial big data, and immense computing power are transforming the way research is conducted and prompts new discoveries. This project develops a new Geospatial Artificial Intelligence (GeoAI) solution to enable large-scale, automated, intelligent, and highly accurate landform feature detection and terrain analysis. Conventional approaches to terrain analysis have been limited to the use of pixel/object-based image analysis and shallow machine learning, which suffer from significant performance challenges when dealing with big data in complex classification tasks. This research, which spatializes deep (machine) learning by incorporating spatial principles and spatial relational data, represents a methodological breakthrough in GeoAI and spatial data science more broadly. Leveraging GeoAI for landform feature recognition enriches spatial knowledge and enhances the understanding of land-surface processes on Earth and other planets. It also benefits geospatial applications that have societal benefit, including anomaly detection that can be used in search and rescue operations and by recognizing landform features indicative of environmental change. The investigators intend on developing the GeoAI community using the model used to establish the Geographic Information Science (GISci) community in the past. GeoAI-related symposia will be organized to serve as an important venue for researchers from diverse disciplines and organizations to discuss new advances and open challenges in GeoAI. The project includes a postdoctoral scholar and trains undergraduate students as interdisciplinary scientists. Both investigators are female scientists and they actively mentor scientists from underrepresented groups. All data and code developed during this project will be open-sourced to benefit the broader geospatial community.Significant challenges exist in successfully applying GeoAI to terrain analysis, including the complexity and diversity in landform features, the dearth of training data, the lack of spatial knowledge in model design, and the limited understanding of machine inferential processes. This research will tackle such challenges by developing (1) a comprehensive terrain dataset GeoNat to support terrain analysis and various machine learning tasks; (2) a machine learning model TerrainAI that injects key spatial principles (spatial autocorrelation and spatial heterogeneity) to enable cross-scale deep learning from multi-source, georeferenced data; and (3) an interactive visualization tool that opens up the "black box" of the machine's learning and decision process. Relying on these tools, three research questions will be answered: (I) What are the unique spatial structures, patterns, and spatial scale that a machine learns to differentiate landform features? (II) How do human and machine recognition processes compare? And (III) How can the underlying geomorphological processes that yield certain forms of a feature in different landscapes be predicted? The identification of process-form relationships significantly impacts the agenda of geographical, spatial scientific and related sciences such as geomorphology in fostering the creation of a community-consensus landform classification system. The TerrainAI model is not limited to the study of landforms but is generalizable and applicable to detect any geographical objects, both natural and human-made.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.