测绘学报 ›› 2021, Vol. 50 ›› Issue (9): 1170-1182.doi: 10.11947/j.AGCS.2021.20210091

• 智能化测绘 • 上一篇    下一篇

深度学习赋能地图制图的若干思考

艾廷华   

  1. 武汉大学资源与环境科学学院, 湖北 武汉 430079
  • 收稿日期:2021-02-21 修回日期:2021-04-04 发布日期:2021-10-09
  • 作者简介:艾廷华(1969-),男,教授,博士,研究方向为地图综合、空间数据挖掘。E-mail:tinghuaai@whu.edu.cn
  • 基金资助:
    国家重点研发计划(2017YFB0503500)

Some thoughts on deep learning enabling cartography

AI Tinghua   

  1. School of resource and environment sciences, Wuhan university, Wuhan 430079, China
  • Received:2021-02-21 Revised:2021-04-04 Published:2021-10-09
  • Supported by:
    The National Key Research and Development Program of China (No. 2017YFB0503500)

摘要: 地图制图学包含地图制作与地图应用两大任务,分别与人工智能技术有不解之缘。经历了符号主义智能表达的地图制图专家系统、行为主义智能表达的空间优化决策后,地图制图面临与连接主义下的深度学习的结合,以提升地图制图的智能化水平。本文针对“深度学习+地图制图”命题讨论了3个问题。一是从深度学习方法与地图空间问题解答策略思想的一致性,基于梯度下降、局部相关、特征降维和非线性化性质,回答了两者结合的可行性;二是从地图学独特的学科特点和技术环境分析了两者结合面临的挑战,涉及地图数据组织的非规范性、样本建立的专业需求、几何与地理特征的融合,以及地图固有的空间尺度性;三是分别讨论了地图制作与地图应用融入深度学习的切入点和具体方法。

关键词: 地图制图, 人工智能, 深度学习, 图卷积学习模型

Abstract: The cartography discipline includes issues of map making and map applications. Both tasks have deep associations with artificial intelligence. Among different intelligence representation methods, the symbolism intelligence approach used to apply with cartography generating mapping expert system technology, the activism intelligence applied with map analysis resulting in optimization decision technology. Nowadays the combination of cartography and connectionism intelligence deep learning faces challenging problems to improve the intelligence level. This study focuses on the issue “deep learning+cartography” discussing three questions. First from the perspective of the consistent ideas in deep learning and map space settlement argues the combination is possible, because both methods have the similar ideas of gradient descent, local spatial association, dimension reduction and non-linear processing. Secondly, by analyzing the mapping characteristics and technology contexts discusses the challenges from the combination, including the irregular data structure in map organization, sample establishment requiring geo-domain knowledge, the integration of geometric and geographic properties and the spatial scale issues in cartography. Thirdly, from the viewpoints of map making and map application respectively examines the practical methods to combine deep learning and cartography.

Key words: cartography, artificial intelligence, deep learning, graph convolution neural network

中图分类号: