Saddle And Peak Elements Recognition Technology Based On Cnn

The elements such as saddles and peaks in DEM are important control points in the natural geomorphology.Their accurate extraction and automatic labeling have important theoretical and practical significance for geoscience analysis,map mapping and geographic information application.the subject take the landform elements of peaks and saddles in DEM as research objects.It intends to analyze the spatial shape and extended structure of landform point,the similarities and differences between the point group constraint relationship between the elements in different landform types.Studying with Convolutional Neural Network and its extension network to find the deep learning mechanisms of point topographical features from grid features to landform element features.To design and construct a depth neural network model that can extract the landform points with global features,improve the possible missing and misleading of feature points in the manual selection feature.Therefore,the main work of the thesis is divided into the following two aspects:Starting from the geologic definition of the saddle point,combining deep learning technology with digital terrain analysis,A saddle extraction method combining shallow features and deep features is designed.Using the improved Convolutional Neural Network to automatically extract the shallow and deep semantic features of the saddle region and perform feature fusion in DEM.and the probability distribution of the candidate saddle is obtained through the Softmax classification layer.The coordinates of the candidate saddle points are then corrected by the MLP network to identify the final saddle point location.In this idea,a deep learning network for saddle points is constructed.The Lenet-5 network model is pre-trained by means of ImageNet dataset.The network training and saddle point identification test are carried out by self-built dataset,and the validity of the reconstructed network model is verified.Based on the spatial structure and morphological characteristics of the topographical area in the DEM data,the shared convolutional layer of Faster R-CNN is used to mine the characteristics of the top-level attributes implicit in the DEM.Using the RPN candidate region generation network to extract the candidate regions of the hilltop,and obtaining the approximate location of the hilltop region through the "end-end" pattern recognition method,and the accurate coordinates of the hilltop points are combined with the classification and position regression network.Constructing the hilltop region identification method of Faster R-CNN combined with RPN.Initialize the network model by ImageNet dataset,and complete the network training and hilltop region identification test on the self-built dataset,which verified the validity of the proposed network model.