3D-MODELING INFRASTRUCTURE FACILITIES USING DEEP LEARNING BASED ON HIGH RESOLUTION SATELLITE IMAGES

Gvozdev, O.; Kosheleva, N.; Kozub, V.; Murynin, A.; Richter, A.
Abstract:
The work describes a method for restoring three-dimensional models of rigid objects from one satellite image, based on the usage of convolutional neural network models. General requirements for reference and test satellite images are: seasonal shooting (lack of snow cover); images’ geo-linking to railway infrastructure objects and to the general coordinate projection; presence of image metadata on survey and lighting conditions (geo-reference data and time of survey, photography and position of the Sun); visibility conditions - shooting in daylight, absence of survey artifacts, complete or partial visibility of objects for construction; pictures’ ortotransformation. The proposed method includes the following steps of processing satellite images. At the first stage, an integral analysis of the image is performed, on which an object of a certain physical class is detected. It is based on the usage of pre-trained networks, rough datasets of objects unloaded from geographic information portals. At the second stage, processing is performed in the locations of areas on the image which were detected in the first step. This network is trained in classes of areas suitable for evaluating three-dimensional shapes and sizes of objects. Such classes include, in particular, roofs (to restore the shape of the base), shadows (to restore the shape of the roof), guide walls and shadows (to restore heights), orts of the buildings (the vicinity of that points in which orthogonal guide walls and eaves of the building’s roof converge). At the third stage, the shape and dimensions are evaluated by the waste of informative areas. It is based on linearization and polarization of the raster contour, as well as the usage of scaling coefficients based on reference objects. During linearization, for each point of the contour line, a segment of the regression line of this and several subsequent points is found. Then there is a line at the points of which the direction of the segments changes. During polarization, the segments are oriented along given guides, for example, in one of two orthogonal directions corresponding to the semi-axes of an equivalent ellipse. At the fourth stage, the geometric model is built. The objects of different physical classes can be given by algebraic formulas, which operands are the simple figures and coefficients that establish dimensions, position and slope in multivariate Euclidean space. For the typical objects, algebraic formulas are considered to be known and the restoration of their models is performed automatically. The models of atypical objects are displayed interactively using a geometric calculator (based on algebraic operations on objects and coefficients). The research was carried out with the support of the Ministry of Education and Science of Russia (unique project identifier RFMEFI60719X0312).
SGEM Research areas:
Year:
2020
Type of Publication:
In Proceedings
Keywords:
satellite image; 3d-modeling; artificial neural network; machine learning; objects of economic infrastructure
Volume:
20
SGEM Book title:
20th International Multidisciplinary Scientific GeoConference SGEM 2020
Book number:
2.2
SGEM Series:
International Multidisciplinary Scientific GeoConference-SGEM
Pages:
149-156
Publisher address:
51 Al. Malinov blvd, Sofia, 1712, Bulgaria
SGEM supporters:
SWS Scholarly Society; Acad Sci Czech Republ; Latvian Acad Sci; Polish Acad Sci; Russian Acad Sci; Serbian Acad Sci & Arts; Natl Acad Sci Ukraine; Natl Acad Sci Armenia; Sci Council Japan; European Acad Sci, Arts & Letters; Acad Fine Arts Zagreb Croatia; C
Period:
18 - 24 August, 2020
ISBN:
978-619-7603-07-1
ISSN:
1314-2704
Conference:
20th International Multidisciplinary Scientific GeoConference SGEM 2020, 18 - 24 August, 2020
DOI:
10.5593/sgem2020/2.2/s10.018
Hits: 1114