Improving the accuracy of random forest-based land-use classification using fused images and digital surface models produced via different interpolation methods


Akar A.

Concurrency and Computation: Practice and Experience, vol.34, no.6, 2022 (SCI-Expanded) identifier identifier

  • Publication Type: Article / Article
  • Volume: 34 Issue: 6
  • Publication Date: 2022
  • Doi Number: 10.1002/cpe.6787
  • Journal Name: Concurrency and Computation: Practice and Experience
  • Journal Indexes: Science Citation Index Expanded (SCI-EXPANDED)
  • Keywords: digital surface model, image classification, kriging, land-use, random forest, unmanned aerial vehicle
  • Erzincan Binali Yildirim University Affiliated: Yes

Abstract

© 2021 John Wiley & Sons, Ltd.Land-use maps produced with high accuracy are extremely important as a basis for better use of the land as they are widely utilized in many areas such as agricultural policies, natural resources management, and environmental operations. This study aimed to produce a high accuracy land-use map using digital surface models (DSMs) produced using different interpolation methods and a random forest (RF) classifier. High spatial resolution Triplesat-2 images, Worldview-2 (WV-2) images, and unmanned aerial vehicle (UAV) images were used in this study. First, DSMs were produced from the point cloud using different interpolation methods. The image fusion process was applied and these fused images were classified using an RF classifier together with the DSMs. Overall results showed that the DSM obtained by the kriging method yielded better results than the other methods by increasing the classification accuracy by 7%–13%. In addition, the McNemar test was applied to the images resulting from the classification of fused images with and without using DSMs to investigate the statistical significance of the differences. McNemar test demonstrated that the (Formula presented.) values were greater than 3.84, which revealed that the DSM produced via kriging interpolation had significantly increased the classification accuracy at a 95% confidence interval.