Automatic recognition of loess landforms using Random Forest method

• Published in: Journal of mountain science Vol. 14; no. 5; pp. 885 - 897
• Main Authors: Zhao, Wu-fan, Xiong, Li-yang, Ding, Hu, Tang, Guo-an
• Format: Journal Article
• Language: English
• Published: Heidelberg Science Press 01.05.2017; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Classification; Datasets; Derivatives; Digital Elevation Models; Earth; Earth and Environmental Science; Earth Sciences; Ecology; Elevation; Environment; Feature extraction; Feature recognition; Forests; Geography; Geomorphology; Landforms; Loess; Methods; Morphology; Procedures; Terraces; Terrain; Texture; Variance analysis; Watersheds; Landform recognition; Feature fusion; Loess landform; DEM; Random Forest
• ISSN: 1672-6316; 1993-0321; 1008-2786
• DOI: 10.1007/s11629-016-4320-9

The automatic recognition of landforms is regarded as one of the most important procedures to classify landforms and deepen the understanding on the morphology of the earth. However, landform types are rather complex and gradual changes often occur in these landforms, thus increasing the difficulty in automatically recognizing and classifying landforms. In this study, small-scale watersheds, which are regarded as natural geomorphological elements, were extracted and selected as basic analysis and recognition units based on the data of SRTM DEM. In addition, datasets integrated with terrain derivatives（e.g., average slope gradient, and elevation range） and texture derivatives（e.g., slope gradient contrast and elevation variance） were constructed to quantify the topographical characteristics of watersheds. Finally, Random Forest（RF） method was employed to automatically select features and classify landforms based on their topographical characteristics. The proposed method was applied and validated in seven case areas in the Northern Shaanxi Loess Plateau for its complex andgradual changed landforms. Experimental results show that the highest recognition accuracy based on the selected derivations is 92.06%. During the recognition procedure, the contributions of terrain derivations were higher than that of texture derivations within selected derivative datasets. Loess terrace and loess mid-mountain obtained the highest accuracy among the seven typical loess landforms. However, the recognition precision of loess hill, loess hill–ridge, and loess sloping ridge is relatively low. The experiment also shows that watershed-based strategy could achieve better results than object-based strategy, and the method of RF could effectively extract and recognize the feature of landforms.