Building segmentation in agricultural land using high resolution satellite imagery based on deep learning approach

L. Y. Liu, C. K. Wang

Research output: Contribution to journalConference articlepeer-review

2 Citations (Scopus)


Understanding building area in agricultural land is important since arable land area in Taiwan is limited. One of the practical ways is manual digitization on high resolution satellite imagery, which can avoid field investigation and achieve satisfying results. However, such practice is tedious and labor intensive. Past researches have shown that deep learning methods are useful to segment buildings in different cities using satellite imagery. In this study, ENVINet5 model was trained and used to segment buildings from high resolution Pleiades pansharpened imagery. The training images (with the size of 2500 pixels × 2500 pixels) were randomly selected from 9 counties/cities to increase diversity since each county/city has different building patterns. The performance of ENVINet5 model reached 0.977, 0.814, 0.847, and 0.829 respectively on accuracy, precision, recall, and F1 score. Since evaluation by pixels can be difficult to show geometry of buildings, we evaluated the model by counting the number of inference building segments, which was post-processed from inference result of ENVINet5 trained model. Further analysis by counting the inference building segments is discussed in this study.

Original languageEnglish
Pages (from-to)587-594
Number of pages8
JournalInternational Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives
Issue numberB3-2021
Publication statusPublished - 2021 Jun 28
Event2021 24th ISPRS Congress Commission III: Imaging Today, Foreseeing Tomorrow - Nice, France
Duration: 2021 Jul 52021 Jul 9

All Science Journal Classification (ASJC) codes

  • Information Systems
  • Geography, Planning and Development


Dive into the research topics of 'Building segmentation in agricultural land using high resolution satellite imagery based on deep learning approach'. Together they form a unique fingerprint.

Cite this