Roof type selection based on patch-based classification using deep learning for high resolution satellite imagery

T. Partovi, F. Fraundorfer, S. Azimi, D. Marmanis, P. Reinartz

Research output: Contribution to journalArticleResearchpeer-review

Abstract

3D building reconstruction from remote sensing image data from satellites is still an active research topic and very valuable for 3D city modelling. The roof model is the most important component to reconstruct the Level of Details 2 (LoD2) for a building in 3D modelling. While the general solution for roof modelling relies on the detailed cues (such as lines, corners and planes) extracted from a Digital Surface Model (DSM), the correct detection of the roof type and its modelling can fail due to low quality of the DSM generated by dense stereo matching. To reduce dependencies of roof modelling on DSMs, the pansharpened satellite images as a rich resource of information are used in addition. In this paper, two strategies are employed for roof type classification. In the first one, building roof types are classified in a state-of-the-art supervised pre-trained convolutional neural network (CNN) framework. In the second strategy, deep features from deep layers of different pre-trained CNN model are extracted and then an RBF kernel using SVM is employed to classify the building roof type. Based on roof complexity of the scene, a roof library including seven types of roofs is defined. A new semi-automatic method is proposed to generate training and test patches of each roof type in the library. Using the pre-trained CNN model does not only decrease the computation time for training significantly but also increases the classification accuracy.

LanguageEnglish
Pages653-657
Number of pages5
JournalInternational Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives
Volume42
Issue number1W1
DOIs
StatusPublished - 30 May 2017

Fingerprint

Satellite imagery
satellite imagery
Roofs
roof
learning
neural network
modeling
Neural networks
reconstruction
Deep learning
Satellites
resources
Remote sensing
remote sensing

Keywords

  • Convolutional neural networks
  • Deep learning method
  • High resolution satellite imagery
  • Roof reconstruction

ASJC Scopus subject areas

  • Information Systems
  • Geography, Planning and Development

Cite this

Roof type selection based on patch-based classification using deep learning for high resolution satellite imagery. / Partovi, T.; Fraundorfer, F.; Azimi, S.; Marmanis, D.; Reinartz, P.

In: International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives, Vol. 42, No. 1W1, 30.05.2017, p. 653-657.

Research output: Contribution to journalArticleResearchpeer-review

@article{abf0be606b6548a98257877dc5e95000,
title = "Roof type selection based on patch-based classification using deep learning for high resolution satellite imagery",
abstract = "3D building reconstruction from remote sensing image data from satellites is still an active research topic and very valuable for 3D city modelling. The roof model is the most important component to reconstruct the Level of Details 2 (LoD2) for a building in 3D modelling. While the general solution for roof modelling relies on the detailed cues (such as lines, corners and planes) extracted from a Digital Surface Model (DSM), the correct detection of the roof type and its modelling can fail due to low quality of the DSM generated by dense stereo matching. To reduce dependencies of roof modelling on DSMs, the pansharpened satellite images as a rich resource of information are used in addition. In this paper, two strategies are employed for roof type classification. In the first one, building roof types are classified in a state-of-the-art supervised pre-trained convolutional neural network (CNN) framework. In the second strategy, deep features from deep layers of different pre-trained CNN model are extracted and then an RBF kernel using SVM is employed to classify the building roof type. Based on roof complexity of the scene, a roof library including seven types of roofs is defined. A new semi-automatic method is proposed to generate training and test patches of each roof type in the library. Using the pre-trained CNN model does not only decrease the computation time for training significantly but also increases the classification accuracy.",
keywords = "Convolutional neural networks, Deep learning method, High resolution satellite imagery, Roof reconstruction",
author = "T. Partovi and F. Fraundorfer and S. Azimi and D. Marmanis and P. Reinartz",
year = "2017",
month = "5",
day = "30",
doi = "10.5194/isprs-archives-XLII-1-W1-653-2017",
language = "English",
volume = "42",
pages = "653--657",
journal = "The international archives of photogrammetry, remote sensing and spatial information sciences",
issn = "1682-1750",
publisher = "International Society for Photogrammetry and Remote Sensing",
number = "1W1",

}

TY - JOUR

T1 - Roof type selection based on patch-based classification using deep learning for high resolution satellite imagery

AU - Partovi, T.

AU - Fraundorfer, F.

AU - Azimi, S.

AU - Marmanis, D.

AU - Reinartz, P.

PY - 2017/5/30

Y1 - 2017/5/30

N2 - 3D building reconstruction from remote sensing image data from satellites is still an active research topic and very valuable for 3D city modelling. The roof model is the most important component to reconstruct the Level of Details 2 (LoD2) for a building in 3D modelling. While the general solution for roof modelling relies on the detailed cues (such as lines, corners and planes) extracted from a Digital Surface Model (DSM), the correct detection of the roof type and its modelling can fail due to low quality of the DSM generated by dense stereo matching. To reduce dependencies of roof modelling on DSMs, the pansharpened satellite images as a rich resource of information are used in addition. In this paper, two strategies are employed for roof type classification. In the first one, building roof types are classified in a state-of-the-art supervised pre-trained convolutional neural network (CNN) framework. In the second strategy, deep features from deep layers of different pre-trained CNN model are extracted and then an RBF kernel using SVM is employed to classify the building roof type. Based on roof complexity of the scene, a roof library including seven types of roofs is defined. A new semi-automatic method is proposed to generate training and test patches of each roof type in the library. Using the pre-trained CNN model does not only decrease the computation time for training significantly but also increases the classification accuracy.

AB - 3D building reconstruction from remote sensing image data from satellites is still an active research topic and very valuable for 3D city modelling. The roof model is the most important component to reconstruct the Level of Details 2 (LoD2) for a building in 3D modelling. While the general solution for roof modelling relies on the detailed cues (such as lines, corners and planes) extracted from a Digital Surface Model (DSM), the correct detection of the roof type and its modelling can fail due to low quality of the DSM generated by dense stereo matching. To reduce dependencies of roof modelling on DSMs, the pansharpened satellite images as a rich resource of information are used in addition. In this paper, two strategies are employed for roof type classification. In the first one, building roof types are classified in a state-of-the-art supervised pre-trained convolutional neural network (CNN) framework. In the second strategy, deep features from deep layers of different pre-trained CNN model are extracted and then an RBF kernel using SVM is employed to classify the building roof type. Based on roof complexity of the scene, a roof library including seven types of roofs is defined. A new semi-automatic method is proposed to generate training and test patches of each roof type in the library. Using the pre-trained CNN model does not only decrease the computation time for training significantly but also increases the classification accuracy.

KW - Convolutional neural networks

KW - Deep learning method

KW - High resolution satellite imagery

KW - Roof reconstruction

UR - http://www.scopus.com/inward/record.url?scp=85021082463&partnerID=8YFLogxK

U2 - 10.5194/isprs-archives-XLII-1-W1-653-2017

DO - 10.5194/isprs-archives-XLII-1-W1-653-2017

M3 - Article

VL - 42

SP - 653

EP - 657

JO - The international archives of photogrammetry, remote sensing and spatial information sciences

T2 - The international archives of photogrammetry, remote sensing and spatial information sciences

JF - The international archives of photogrammetry, remote sensing and spatial information sciences

SN - 1682-1750

IS - 1W1

ER -