Research On Extraction Of Multi-channel Fusion Terrain Feature Line Based On Deep Learning

Ridge and valley lines are the most important morphological features in terrain.The accurate extraction of ridge and valley line information from terrain data is of great significance to terrain analysis,hydrological analysis,and three-dimensional reconstruction of terrain.It is an important problem in the study of topography and geomorphology,even in the study of computer graphics.After years of research,people have so far proposed a variety of methods to extract terrain feature lines,but the existing extraction methods still have many problems: the spatial positioning error of the feature line extraction is large and the accuracy is not good;there are omissions,false mentions,and Pseudo-feature line problems;the accuracy of extraction depends on manually selected elements to identify features;more manual postprocessing is required,and the extraction automation of the algorithm need to be improved.In order to improve the accuracy of the extraction of terrain feature lines,based on the in-depth analysis of the problems in terrain feature extraction,this paper proposes an idea of fusion and extraction of terrain feature lines from multiple data sources: extraction based on multiple data sources for different terrain features,the multiple feature extraction results are complemented,so that they can be merged to achieve a better extraction performance.We have built a multi-channel fusion terrain feature extraction network based on deep learning,comprehensively using digital elevation model,remote sensing images,slope information and aspect information and other data sources to learn and extract terrain features.The network structure is constructed based on the U-net network.In response to the loss of feature details caused by the down-sampling operation in the U-net network,the edge of the extraction result is smooth,etc.,by deepening the network level,optimizing the activation function,and constructing a multi-channel fusion of terrain features Extract the network structure and other three aspects to optimize the network,so that it can extract the physical and visual features of the terrain based on multiple types of data input,and make the two features complementary and merge to construct a joint feature;On this basis,image semantic segmentation is performed based on joint features,and the final terrain feature line is obtained.Experiments show that the extraction accuracy and efficiency of this method are significantly higher than the traditional single-channel terrain feature extraction method.The main contributions of this paper include:(1)The construction of multi-channel fusion terrain feature extraction data set.The data set includes four combined data sources: DEM,remote sensing image,slope map,and aspect map.In the process of extracting the terrain feature lines of the training data set,in order to improve the accuracy of extraction,the following methods are used: firstly,the optimal threshold value of confluence is extracted by the mean change point method,and then extracted based on the physical simulation method of surface water flow.The accurate natural ridge line and valley line are used as label data.(2)The construction of a network of multi-channel terrain feature fusion and extraction structure.The network is constructed based on the U-net network structure.Through the deepening of the network hierarchy,the optimization of activation functions,and the construction of a multi-channel fusion network framework,based on a variety of data combinations,the physical and visual features of the landform can be extracted and complemented and merged.The joint feature is obtained,and then the feature line is extracted.This method effectively integrates the physical and visual features of the terrain,retains the detailed information in the terrain,and makes the feature extraction result significantly improved compared to the traditional single-channel extraction method.