Scholarlink

<p>The accurate extraction of building footprints, especially in regions with varying building densities, poses significant challenges due to the complexity and variation in building structures. This study explores the use of Indian high-resolution satellite imagery from Cartosat datasets and Convolutional Neural Networks (CNN), to address these challenges while improving building extraction in low, medium, and high-density urban pockets. High-resolution PAN and Multispectral imagery (Cartosat-2E (0.6 and 2&#xa0;m resolution) and Cartosat-3 (0.3 and 1&#xa0;m resolution) datasets) was pre-processed using geometric corrections, followed by data augmentation and structured training and testing data split for different spatial locations for unbiased model calibration and testing. The U-Net model was trained over 100 epochs, exhibiting improved accuracy with each epoch as it adjusted its parameters. Cartosat-3 outperformed CartoSat-2E for automated building footprint extraction as it has shown 89.56%, 88.05%, and 85.67% accuracies, as compared to 68.48%, 63.78%, and 45.67% accuracies displayed by Cartosat-2E for low, medium, and high-density areas, respectively. The implementation of trained model on a larger area (~ 1875 sq km area) with reasonable accuracies shows the robustness of developed model. The U-Net model showed promising performance with both datasets, but the CartoSat-3 dataset shows enhanced performance in capturing building structures especially in dense urban environments. This approach highlights the potential of using Cartosat-3 in combination with advanced deep learning models for precise urban mapping, offering valuable insights for urban planning and management.</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

A Deep Learning Approach for Automated Building Footprint Extraction from Cartosat Imagery

  • 期刊论文

  • 发表日期：

    2026年2月16日

查看全文
  • Kakali Deka,
  • Kshama Gupta,
  • Surendra Kumar Sharma,
  • Ashutosh Bhardwaj,
  • Bhoomika Ghale,
  • Sandeep Maithani

摘要

The accurate extraction of building footprints, especially in regions with varying building densities, poses significant challenges due to the complexity and variation in building structures. This study explores the use of Indian high-resolution satellite imagery from Cartosat datasets and Convolutional Neural Networks (CNN), to address these challenges while improving building extraction in low, medium, and high-density urban pockets. High-resolution PAN and Multispectral imagery (Cartosat-2E (0.6 and 2 m resolution) and Cartosat-3 (0.3 and 1 m resolution) datasets) was pre-processed using geometric corrections, followed by data augmentation and structured training and testing data split for different spatial locations for unbiased model calibration and testing. The U-Net model was trained over 100 epochs, exhibiting improved accuracy with each epoch as it adjusted its parameters. Cartosat-3 outperformed CartoSat-2E for automated building footprint extraction as it has shown 89.56%, 88.05%, and 85.67% accuracies, as compared to 68.48%, 63.78%, and 45.67% accuracies displayed by Cartosat-2E for low, medium, and high-density areas, respectively. The implementation of trained model on a larger area (~ 1875 sq km area) with reasonable accuracies shows the robustness of developed model. The U-Net model showed promising performance with both datasets, but the CartoSat-3 dataset shows enhanced performance in capturing building structures especially in dense urban environments. This approach highlights the potential of using Cartosat-3 in combination with advanced deep learning models for precise urban mapping, offering valuable insights for urban planning and management.