Demolished building detection from aerial imagery using deep learning

Shu Su,Takahiko Nawata

doi:10.5194/ica-proc-2-122-2019

Abstract

Abstract. In this paper, we present a novel approach for demolished building detection using bi-temporal aerial images and building boundary polygon data. The building boundary polygon data can enable the proposed method to distinguish buildings from non-buildings. Moreover, it can enable the exclusion of non-building changes such as those caused by changes in tree cover, roads, and vegetation. The results of demolished building detection can be achieved by using the building-base. The proposed method classifies each building as demolished or undemolished. The architectures, which based on U-Net and VGG19, are implemented for realizing automatic demolished building detection. The result suggested that U-Net is a useful architecture for image classification problems as well as for semantic segmentation tasks. In order to verify the effectiveness of proposed method, the detection performance is evaluated using images of an entire city. The results suggest that the proposed method can accurately detect demolished buildings with a low mis-detection rate and low over-detection rate.

Highlights

Over the last decade, with advances in computer vision techniques, building change detection has emerged as an active research area in the domain of photogrammetry and remote sensing
In this paper, we present a novel approach for demolished building detection that uses bi-temporal aerial images and building polygon data
We proposed a novel demolished building detection method that used bi-temporal aerial images and building polygon data based on the U-Net and VGG19 architectures

Summary

Introduction

With advances in computer vision techniques, building change detection has emerged as an active research area in the domain of photogrammetry and remote sensing. Previous studies on building change detection have mainly focused on the spectral information and color variations in bi-temporal images (Bourdis et al, 2011). These methods yield too many errors, including mis-detection (i.e., a demolished building going undetected), over-detection (i.e., an undemolished building being detected as demolished, or a lot of changes in non-building area.), especially over-detection

Results

Discussion

Conclusion