Research On Peak Recognition And Terrain Generation Method Of Deep Learning

The extraction of terrain elements has always been a research hotspot in the field of digital terrain analysis.As the important control points of geomorphology,peak elements’ accurate extraction and automatic annotation are of great significance to environmental changes,vegetation distribution,and scenic spot planning.In order to solve the problems of traditional methods which were limited by the qualitative description,physical characteristics and spatial correlation of peak elements,the thesis is based on the deep neural network in deep learning,to study the influence mechanism of different landforms and DEM scales on peak elements,the learning mechanism of terrain regeneration network model,build a deep network model for intelligent identification of peak elements and a method for augmenting terrain samples.The main research contents are as follows:(1)Analyze the spatial structure and morphological characteristics of the mountain top area in the DEM data,combine the information entropy theory to study the characteristics of DEM data under different landforms and different scales,and explore suitable analysis window scales that can express the complete mountain top shape.On this basis,the Faster R-CNN network is used as the basic deep network,and the feature pyramid network(FPN)is used to learn the shape and spatial characteristics of the mountain top,and the candidate region generation network RPN is used to extract the location of the mountain top area.A new method of peak elements machine learning is proposed,comparing with the traditional feature extraction method and the extraction performance of the basic Faster R-CNN network,verified the effectiveness of the new method.(2)The research of digital terrain analysis based on deep learning often relies on a large number of data samples.However,due to the limited actual terrain,the sample size is scarce,so it is necessary to study the method of increasing terrain samples.In this paper,a depth neural network was designed to generate virtual terrain based on the structure of depth convolution generation countermeasure network and the real DEM data of mountain area.The feasibility of this method was verified by experiments.It not only improved the scarcity of sample data of typical terrain elements,but also expanded the application of depth learning method in the field of Geosciences.