Scholarlink

Buildings are dynamic base map elements as they relate to population growth. Building object detection is generally performed on orthophoto data manually or digitized on screen. A rapidly growing method to automatically detect buildings is the deep learning approach using the Mask Region-based Convolutional Neural Network (Mask R-CNN) architecture. This research aims to identify the ability of Mask R-CNN in detecting buildings and determine the accuracy value of the detection results. The research location is Tanjung Karang Village, Mataram City, West Nusa Tenggara. The deep learning model used shows correctfitting conditions characterized by a plot of training loss and validation loss values that decrease until stable. Detection was carried out in two areas: ROI 1 resulted in a precision value of 90.97%, recall 95.27%, F1 Score 93.07% and accuracy 87.05%; ROI 2 resulted in a precision value of 85.94%, recall 88.38%, F1 Score 87.24% and accuracy 77.22%. The model is more optimal in areas with low and regular building density criteria such as ROI 1 and still requires manual improvement by the operator.

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 Detection on Orthophoto Data Using Deep Learning Mask R-CNN Approach in Tanjung Karang Village, Mataram City

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Diva Amevia Mahendradani,
  • Maulana Yudinugroho,
  • Mochamad Irwan Hariyono,
  • Naufal Setiawan,
  • Dessy Apriyanti

摘要

Buildings are dynamic base map elements as they relate to population growth. Building object detection is generally performed on orthophoto data manually or digitized on screen. A rapidly growing method to automatically detect buildings is the deep learning approach using the Mask Region-based Convolutional Neural Network (Mask R-CNN) architecture. This research aims to identify the ability of Mask R-CNN in detecting buildings and determine the accuracy value of the detection results. The research location is Tanjung Karang Village, Mataram City, West Nusa Tenggara. The deep learning model used shows correctfitting conditions characterized by a plot of training loss and validation loss values that decrease until stable. Detection was carried out in two areas: ROI 1 resulted in a precision value of 90.97%, recall 95.27%, F1 Score 93.07% and accuracy 87.05%; ROI 2 resulted in a precision value of 85.94%, recall 88.38%, F1 Score 87.24% and accuracy 77.22%. The model is more optimal in areas with low and regular building density criteria such as ROI 1 and still requires manual improvement by the operator.