摘要
地貌识别已经成为地貌学研究的关键一环,而基于集水区的地貌单元研究成为了地貌识别领域的研究热点。前人的研究衍生出以下新的问题,能否从基于局部集水区地貌的特征对大尺度地貌类型进行识别,对何种地形描述方法适应性更强等。由此本文选取岩溶地貌、黄土地貌、冰缘地貌、风成地貌、流水地貌这五种代表中国主要地貌类型的样区,引入复杂网络理论方法,以复杂网络指标、地形指标为基本数据源,使用LightGBM、XGBoost、RF 3种典型机器学习方法对中国主要地貌类型进行自动识别的研究。实验结论如下:集水区复杂网络结构和地形特征都对地貌具有一定的解释力和识别效果,总体识别精度分别为77.5%和72.5%,在本文选取的地貌类型中,LightGBM、XGBoost、RF 3种机器学习方法对冰缘地貌的识别精度最高,最高可达100%;2种地貌描述方法结合的地貌识别效果相较单一地貌描述方法具有显著提高,总体精度比单一复杂网络指标和单一地形指标,分别提高了5%和10%;同时LightGBM对于复杂网络量化因子和地形指标特征集的结合具有更好的适应性,总体精度可达82.5%。总体而言,本研究工作拓展了基于集水区地貌识别方法的应用区域和应用范畴,为基于集水区的地貌识别研究提供了新的思路。
Landform recognition has become a key part of geomorphological research,which has been widely concerned by scholars.The research of geomorphic units based on catchment has become a hotspot in the field of landform recognition.Previous studies have generated a series of new questions,such as whether large-scale landform types can be identified based on local catchment landform features,which landform description methods are more adaptable,and what is the knowledge bottleneck of current landform recognition methods based on the catchment.So,in this paper,we selected sample areas representing five major landform types in China,including karst,loess,periglacial,aeolian,and fluvial.Based on the complex network theory,we took the complex network indicators and the topographic metrics as the basic data sources.Three typical machine learning methods,i.e.,LightGBM,XGBoost,and RF,were used to automatically identify the main geomorphic types in China.Results show that both the complex network structure and the terrain features of the catchment have certain explanatory power and recognition effect on landforms,and the overall recognition accuracy is 77.5%and 72.5%,respectively.Among the five geomorphologic types selected,LightGBM,XGBoost,and RF machine learning methods have the highest recognition accuracy(up to 100%)on periglacic geomorphology.Compared to a single geomorphic description data source,the geomorphic recognition effect that combines the two data sources is significantly improved.The overall accuracy using two data sources is 5%and 10%higher than that using the single complex network dataset and the single topographic dataset,respectively.Moreover,LightGBM has better adaptability to the combination of complex network and terrain factor feature sets,and the overall accuracy can reach 82.5%.In general,this study expands the application area and scope of catchment landform recognition methods,and provides a new idea for the research of catchment landform recognition.
作者
戚梦
陈楠
林偲蔚
周千千
QI Meng;CHEN Nan;LIN Siwei;ZHOU Qianqian(Key Lab for Spatial Data Mining and Information Sharing of Education Ministry,Fuzhou University,Fuzhou 350108,China;The Academy of Digital China,Fuzhou University,Fuzhou 350108,China;School of Geographic and Oceanographic Sciences,Nanjing University,Nanjing 210023,China;College of computer and data science,Fuzhou University,Fuzhou 350108,China)
出处
《地球信息科学学报》
CSCD
北大核心
2023年第5期909-923,共15页
Journal of Geo-information Science
基金
国家自然科学基金项目(41771423)。
关键词
中国
地貌识别
集水区
DEM
复杂网络
地形指标
机器学习
China
landform recognition
watershed
DEM
complex network
terrain feature
machine learning
