Scholarlink

<p>The study of urban antispaces has brought interesting conclusions in urban dynamics. Still, the number of investigations is scarce since identifying and localizing these antispaces is laborious and complicated. Therefore, it is essential to have a tool that facilitates decision-makers in detecting these antispaces. This is why the present work seeks to study the feasibility of using deep learning techniques using object detection and semantic segmentation models to detect these antispaces automatically. For this, a dataset named AntisMx2020 was created using satellite images in the first instance. Object detection models (YOLOv4 and YOLOv5) and a semantic segmentation model (Deeplabv3+) were trained on this dataset. Subsequently, the performance of these models in antispace detection was compared, and YOLOv5 was selected for further experimentation to increase its performance. Its hyperparameters were adjusted using an evolutionary method, resulting in substantial performance improvement of 4.6%, that is, a final average precision of 75.3%. The recall and F1-score achieved considering this final configuration were 69.2% and 70.8%, respectively. The proposed methodology allowed the acquisition of models with good performance in antispace detection, demonstrating its feasibility, which opened the door to future research.</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

Detection of urban antispaces using deep learning in northern border cities of Mexico

  • 期刊论文

  • 发表日期：

    2026年1月22日

查看全文
  • Jorge Adrián Martínez-López,
  • Héctor De-la-Torre-Gutiérrez,
  • Francisco J. Hernandez-Lopez

摘要

The study of urban antispaces has brought interesting conclusions in urban dynamics. Still, the number of investigations is scarce since identifying and localizing these antispaces is laborious and complicated. Therefore, it is essential to have a tool that facilitates decision-makers in detecting these antispaces. This is why the present work seeks to study the feasibility of using deep learning techniques using object detection and semantic segmentation models to detect these antispaces automatically. For this, a dataset named AntisMx2020 was created using satellite images in the first instance. Object detection models (YOLOv4 and YOLOv5) and a semantic segmentation model (Deeplabv3+) were trained on this dataset. Subsequently, the performance of these models in antispace detection was compared, and YOLOv5 was selected for further experimentation to increase its performance. Its hyperparameters were adjusted using an evolutionary method, resulting in substantial performance improvement of 4.6%, that is, a final average precision of 75.3%. The recall and F1-score achieved considering this final configuration were 69.2% and 70.8%, respectively. The proposed methodology allowed the acquisition of models with good performance in antispace detection, demonstrating its feasibility, which opened the door to future research.