Identifying and counting of buildings using artificial neural network and reduced representation in high-resolution images

Abstract

In this study, Hölder Exponent (HE), Variance (var) and densely populated range are measured to extract meaningful features from the high resolution (HR) images. These extracted features are considered here as the reduced representation of high resolution (HR) images. Five-layer simplified artificial neural network (ANN) architecture configured and trained using the reduced representation for building detection and counting in HR images. The scale and orientation of the movable window are changed to find an optimum bounding box of the buildings. The method is validated by applying on World View-2 pan-sharpened multispectral images having spatial resolution 0.46 m. In comparison with CNN and ResNet-18, the accuracy assessment result is quite promising (92%) with proposed method for detecting buildings in HR images. The proposed method, distinctly identified scale variant buildings and detected individual buildings even if they are in close proximity.

Keywords:

• High resolution image
• building detection
• ANN
• deep learning
• reduced representation
• bounding
• box

Acknowledgments

The authors are grateful to the Director, NRSC, Hyderabad, India and CGM, RCs, NRSC, Hyderabad, India for their encouragement and guidance on carrying out this research. Authors are thankful to the General Manager, RRSC-East, NRSC, Kolkata, India for his support during the course of this study. Authors sincerely thank the anonymous reviewer for contributing insightful remarks and useful suggestions that has substantially improved the quality of the manuscript.