Scholarlink

<p>Building footprint extraction is a fundamental task in the field of remote sensing and computer vision and has numerous applications such as urban planning, disaster management, and environmental monitoring. In this study, we present an approach for building footprint extraction using the U-net architecture with ResNet as backbone. The study utilizes publicly available datasets for both training and testing purposes. Various architectures are compared for building extraction, and the most promising architecture is identified for further investigation. Through experimentation, U-net with ResNet-34 as backbone is found to exhibit superior performance compared to other backbones. To further enhance the performance, an SE (Squeeze-and-Excitation) module is integrated with the ResNet-34 in U-net architecture. The SE module is designed to recalibrate the channel-wise feature responses, enabling the network to focus on more relevant and informative features. The experimental results demonstrate that the U-net architecture with ResNet-34 backbone and with an integrated SE module achieves an overall improvement of 3.46% in accuracy in comparison to the model without the SE module. Further, when compared to other well-known architectures, the improvement in accuracy varies between 6.154% and 16.923%. Our findings highlight the effectiveness of the ResNet-34 architecture for building footprint extraction. Moreover, the incorporation of the SE module improves the model’s performance, leading to more precise and reliable building extraction results. The presented approach contributes to the advancement of building extraction techniques and provides valuable insights for researchers and practitioners working in the field of remote sensing, computer vision and related applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Building Extraction for Urban Infrastructure Mapping Using Deep Neural Networks and High-Resolution Remote Sensing Data

  • 期刊论文

  • 发表日期：

    2025年1月31日

查看全文
  • Surendra Kumar Sharma,
  • Vivek Sharma,
  • Vishal Mishra

摘要

Building footprint extraction is a fundamental task in the field of remote sensing and computer vision and has numerous applications such as urban planning, disaster management, and environmental monitoring. In this study, we present an approach for building footprint extraction using the U-net architecture with ResNet as backbone. The study utilizes publicly available datasets for both training and testing purposes. Various architectures are compared for building extraction, and the most promising architecture is identified for further investigation. Through experimentation, U-net with ResNet-34 as backbone is found to exhibit superior performance compared to other backbones. To further enhance the performance, an SE (Squeeze-and-Excitation) module is integrated with the ResNet-34 in U-net architecture. The SE module is designed to recalibrate the channel-wise feature responses, enabling the network to focus on more relevant and informative features. The experimental results demonstrate that the U-net architecture with ResNet-34 backbone and with an integrated SE module achieves an overall improvement of 3.46% in accuracy in comparison to the model without the SE module. Further, when compared to other well-known architectures, the improvement in accuracy varies between 6.154% and 16.923%. Our findings highlight the effectiveness of the ResNet-34 architecture for building footprint extraction. Moreover, the incorporation of the SE module improves the model’s performance, leading to more precise and reliable building extraction results. The presented approach contributes to the advancement of building extraction techniques and provides valuable insights for researchers and practitioners working in the field of remote sensing, computer vision and related applications.