Applying a Deep Learning Approach for Building Extraction From High-Resolution Remote Sensing Imagery

Applying a Deep Learning Approach for Building Extraction From High-Resolution Remote Sensing Imagery

Dolonchapa Prabhakar (Indian Institute of Technology, Roorkee, India) and Pradeep Kumar Garg (Indian Institute of Technology, Roorkee, India)
DOI: 10.4018/978-1-6684-7319-1.ch008
OnDemand:
(Individual Chapters)
Available
$37.50
No Current Special Offers
TOTAL SAVINGS: $37.50

Abstract

As data science applies to mapping buildings, great attention has been given to the potential of using deep learning and new data sources. However, given that convolutional neural networks (CNNs) dominate image classification tasks, automating the building extraction process is becoming more and more common. Increased access to unstructured data (such as imagery and text) and developments in deep learning and computer vision algorithms have improved the possibility of automating the extraction of building attributes from satellite images in a cost-effective and large-scale manner. By applying intelligent software-based solutions to satellite imageries, the manual process of acquiring features such as building footprints can be expedited. Manual feature acquisition is time-consuming and expensive. The buildings may be recovered from RGB photos and are extremely properly identified. This chapter offers suggestions to quicken the development of DL-centred building extraction techniques using remotely sensed images.
Chapter Preview
Top

Introduction

Automatic building detection, a fundamental task in the field of remote sensing, is one of the most crucial tasks in population density estimation, the creation and updating of thematic maps (Yuan, 2018 and Wang & Miao, 2022), change detection, disaster management (Wei et al., 2020), and urban planning (Chen et al., 2020 and Temenos et al., 2022), which results from the fact that people spend the majority of their time residing inside the buildings and interacting with one another. As a result, it's essential to map every single building's placement precisely in the course of the urban planning process because doing so is quicker and more accurate than using more conventional techniques. Due to the extensive building coverage, the manual method is quite a time- and money-consuming. The rapid urbanization and need for current maps for contemporary applications, which necessitate more frequent updates to building information, present additional difficulties. This prompts ongoing studies into automating building extraction, which is advantageous for geomatics and environmental science applications (Y. Wu et al., 2022).

This has driven the research to utilise satellite images, which can depict the majority of the metropolitan environment. Although building extraction has recently garnered a lot of attention, it is still a difficult operation because of the occlusion, complexity and noise of the backdrop elements in the actual remote sensing imageries. In recent years, a variety of different approaches to extracting buildings from images acquired using remote sensing have been developed, but the majority of these techniques concentrate on the pixel (Sirmaçek & Ünsalan, 2008), spectrum (Zhang, 1999), length, edge (Ferraioli, 2010); (Li & Wu, 2008), shape (Lin et al., 2014), texture (Awrangjeb et al., 2013), and shadow (Silva et al., 2018). Having precise and efficient building information is necessary for carrying out these applications. The following are some of the special characteristics and difficulties that come with building extraction:

  • Boundary concerns are particularly important in remote sensing images because of the more intricate and varied backgrounds and settings, as well as how buildings have more regular and well-defined outlines than natural items.

  • There is typically a lot of variation in building kinds. Their tones, textures, and range of spatial dimensions are all different. They may also differ from building to building in terms of their colours and shapes.

  • Due to a number of intricate aspects, including manmade non architectural elements, shadows, and the complexity of building facades, there is a substantial cause for concern over the remote correlation that exists between structures and the material in their surroundings.

  • Due to their proximity to elements made of analogous resources, including roads, buildings are sometimes mistaken for these characteristics. Particular importance should be placed on the quality of segmentation of boundary contours.

  • Processing of large-scale dataset. The VHR imagery provides detailed building information for segmentation at the cost of a huge volume of data to be processed.

High-resolution remote sensing images are now widely accessible thanks to the quick development of imaging technology. Consequently, study in the area of remote sensing has exploded, and automatic building segmentation from high-resolution images has drawn a lot of interest (Luo et al., 2021). The method of extracting buildings from remote sensing imagery is depicted in Figure 1. This process may be characterised as a semantic segmentation problem and entails pixel-level categorization of images to produce binary images with contents of building or non-building.

Complete Chapter List

Search this Book:
Reset