地球信息科学学报 ›› 2023, Vol. 25 ›› Issue (7): 1297-1311.doi: 10.12082/dqxxkx.2023.220795

• • 上一篇    下一篇

感知物理先验的机器学习及其在地理空间智能中的研究前景

张彤1,*(), 刘仁宇1, 王培晓1, 高楚林1, 刘杰1, 王望舒2   

  1. 1.武汉大学测绘遥感信息工程国家重点实验室,武汉 430079
    2.奥地利维也纳工业大学测量与地理信息系,维也纳 A-1040
  • 收稿日期:2022-10-17 修回日期:2022-11-04 出版日期:2023-07-25 发布日期:2023-06-30
  • 作者简介:张 彤(1979— ),男,福建武夷山人,博士,教授,研究方向为时空机器学习、高分遥感解译、交通地理信息系统。E-mail: zhangt@whu.edu.cn
  • 基金资助:
    国家重点研发计划“政府间国际科技创新合作/港澳台科技创新合作”重点专项(2019YFE0106500);国家自然科学基金项目(41871308)

Physics-informed Machine Learning and Its Research Prospects in GeoAI

ZHANG Tong1,*(), LIU Renyu1, WANG Peixiao1, GAO Chulin1, LIU Jie1, WANG Wangshu2   

  1. 1. State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing Science, Wuhan University, Wuhan 430079, China
    2. Department of Geodesy and Geoinformation, TU Wien, Vienna A-1040, Austria
  • Received:2022-10-17 Revised:2022-11-04 Online:2023-07-25 Published:2023-06-30
  • Contact: *ZHANG Tong, E-mail: zhangt@whu.edu.cn
  • Supported by:
    National Key R&D Program of China (International Scientific & Technological Cooperation Program)(2019YFE0106500);National Natural Science Foundation of China under Grant(41871308)

摘要:

许多复杂的物理现象和动态系统尚未为科学家所完全理解和解释,因此不能构建确定的数学方程来描述,不能直接使用紧凑的物理机理模型来进行分析和预测。随着观测数据的日益丰富,数据驱动的机器学习方法可以较好地描述复杂非线性现象,但是纯数据驱动模型在表征能力、可解释、泛化能力、样本利用效率方面还存在很多不足。常规机器学习方法在地学领域的应用还面临时空异质性、样本稀疏等带来的挑战。近年来感知物理先验的机器学习方法可以在物理原理不太明确的情况下更有效地利用观测数据描述和分析动态系统,受到了广泛关注,并在物理、计算机、生物、医学和地学等领域得到了一定的应用。近年来人工智能和机器学习技术已经大量应用于地理学尤其是地理信息和遥感领域,受到地理学者广泛重视,被称为地理空间智能,并已成为一个热门的研究方向。感知物理先验的机器学习方法融合了模型驱动和数据驱动思想,为地理空间智能研究带来新的研究范式,促进各种复杂地理现象的精细描述和预测。本文首先分别从物理先验的表达形式和如何在机器学习中集成物理先验两方面概述了该领域的进展。物理先验是在机器学习建模之前已经存在、独立于机器学习方法的知识。本文从增广的数据和定制特征、物理定律和约束规则、支配方程、几何特性等方面总结物理先验知识的表达形式。主要从机器模型约束建模、辅助任务设计和模型的训练推理角度总结如何在机器学习模型中有效集成各种物理先验。本文基于以上的综述框架,结合地学机器学习和地理空间智能的发展和前沿问题,探讨了地理时空先验与其他物理先验的关系,简要总结和分析了目前感知时空先验的地理空间智能方法研究案例,探讨了时空先验表征以及集成时空先验地理空间智能的未来研究规划和应用前景。随着感知物理先验的机器学习方法研究的快速发展,我们相信感知时空先验的地理空间智能研究未来将逐步构建起跨多时空尺度的通用地理表征、分析、预测和解释框架,不仅能更好地解决地理信息科学的传统问题,还将鼓励地理学者与其他相关学科一起建立交叉研究的前沿机会,探索解决人类未来面临的共同挑战。

关键词: 物理感知的机器学习, 物理先验, 机器学习, 深度学习, 时空先验, 时空表征, 地理空间智能, 地理信息科学

Abstract:

Scientists still cannot fully understand and explain many complex physical phenomena and dynamic systems, which cannot be described by deterministic mathematic equations and be analyzed and predicted through compact physical mechanistic models. With the ever-increasing of observational data, data-driven machine learning methods can effectively describe many complex non-linear phenomena. Nevertheless, pure data-driven models still have shortcomings in representation, interpretation, generalization capabilities, and sample efficiency. Conventional machine learning methods are confronted with challenges brought by spatiotemporal heterogeneity and sample sparsity. Recently, Physics-Informed Machine Learning (PIML) can effectively leverage observation data to describe and analyze dynamical systems when physical principles are uncertain. PIML has gain wide attention and been extensively applied in physics, computer science, biology, medical science, and geosciences. In recent years, artificial intelligence and machine learning technologies have been widely applied in geography, especially in GIScience and remote sensing, attracting wide research interests of geographers. This line of research is termed GeoAI and has become a cutting-edge research frontier in geography. PIML methods integrate the ideas of model-driven and data-driven methods, introducing new research paradigms for GeoAI and improving the description and prediction of complex geographical phenomena. This survey first summarizes recent progress in this domain from the perspectives of the representation of physical priors and the integration of physical priors in machine learning methods. Physical prior refers to existing independent knowledge that is already available before building machine learning models. This survey reviews the representation of physical priors from the aspects of augmented data and customized features, physical laws and constraints, governing equations as well as geometric properties. We also review how physical priors are integrated into various machine learning models, including constraint modeling, auxiliary task design as well as model training and inference. Based on the PIML survey framework, we explore the relationships between spatiotemporal priors and other physical priors, before briefly reviewing and summarizing typical case studies of spatiotemporal prior-informed GeoAI research. We also discuss the research agenda and future prospects of spatiotemporal prior representation and the spatiotemporal prior-informed GeoAI in the context of geo-machine learning and GeoAI frontiers. In light of fast progress of PIML, we contend that GeoAI studies that are well informed by spatiotemporal priors can gradually establish a generic geographical representation, analysis, prediction, and interpretation framework, which not only helps handle many classical problems in GIScience but also addresses future profound challenges of human being by encouraging geographers to explore more research opportunities when collaborating with researchers from other disciplines.

Key words: PIML, physical priors, machine learning, deep learning, spatiotemporal priors, spatio-temporal representation, Geospatial Intelligence, Geographic Information Science