Constructing 3D Models of Rigid Objects from Satellite Images with High Spatial Resolution Using Convolutional Neural Networks

Abstract

A way of constructing 3D models of rigid objects from one satellite image is described. It is based on the use of two convolution neural networks which sequentially process high-resolution satellite images. The first neural network performs integral image analysis for segmentation and identification of objects of specified physical classes. The second neural network performs local image analysis and works with images segmented by the first neural network in areas of the image that presumably contain objects of specified classes. An algorithm for reconstructing a 3D model of an object from raster domains of a segmented image obtained from local analysis is described. It is based on regression analysis, the assessing of equivalent figures, and the linearization and polarization of contours. Results from the algorithm’s operation are given using the example of railway infrastructure facilities. The results from constructing 3D models of three objects of the railway infrastructure, identified via the operation of neural networks for four informative classes of areas are presented, e.g. roofs, walls, railroad tracks, contanc lines (poles). Standard dimensions (e.g., the railway gauge (1.52 m) and the height of railway support poles (11.35 m)) are used to estimate scaling coefficients that allow determination of base dimensions and object heights. The possibility of constructing 3D models of objects of areas from 210 to 4200 m2 is shown.