Details
Original language | English |
---|---|
Pages (from-to) | 2335-2347 |
Number of pages | 13 |
Journal | Earth science informatics |
Volume | 15 |
Issue number | 4 |
Publication status | Published - Dec 2022 |
Externally published | Yes |
Abstract
Forest changes caused by fires, clear-cuts, and Land Use/Land Cover (LULC) changes have negatively affected the climate, wildlife, and global ecosystem. By monitoring the forest changes, managers and planners can make appropriate decisions to preserve these natural areas. In this regard, this paper presents a novel deep learning-based forest change detection method including two main steps: (1) producing a new difference image enabling a more efficient distinction of changed and unchanged areas, and (2) generating a reliable forest change map by applying the recurrent residual-based U-Net deep neural network (R2U-Net) on the difference image. The recently introduced forest fused difference image (FFDI) is first improved by modifying its weighted angular operator as well as applying the fast local laplacian filter (FLLF) to generate an enhanced forest fused difference image (EFFDI). R2U-Net is subsequently used to segment the EFFDI into the changed and unchanged regions because it preserves their geometric shapes more effectively than other U-net variants. To assess the efficacy of the presented method, experimental results were conducted on four bi-temporal images acquired by the Sentinel 2 and Landsat 8 satellite sensors. The qualitative and quantitative results demonstrated the effectiveness of the proposed EFFDI in reflecting the true forest changes from the background. Moreover, compared with the other conventional U-Net-based models, including U-Net, ResU-Net, and U-Net ++, forest changes and their geometrical details were better preserved by R2U-Net. Furthermore, the proposed approach outperformed the other state-of-the-art supervised and unsupervised change detection methods in terms of quantitative and qualitative results, demonstrating its high potential for forest change detection applications.
Keywords
- Deep learning, Forest change detection, Recurrent residual-based U-net (R2U-net), Remote sensing, U-net
ASJC Scopus subject areas
- Earth and Planetary Sciences(all)
- General Earth and Planetary Sciences
Sustainable Development Goals
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: Earth science informatics, Vol. 15, No. 4, 12.2022, p. 2335-2347.
Research output: Contribution to journal › Article › Research › peer review
}
TY - JOUR
T1 - FCD-R2U-net
T2 - Forest change detection in bi-temporal satellite images using the recurrent residual-based U-net
AU - Khankeshizadeh, Ehsan
AU - Mohammadzadeh, Ali
AU - Moghimi, Armin
AU - Mohsenifar, Amin
N1 - Publisher Copyright: © 2022, The Author(s), under exclusive licence to Springer-Verlag GmbH Germany, part of Springer Nature.
PY - 2022/12
Y1 - 2022/12
N2 - Forest changes caused by fires, clear-cuts, and Land Use/Land Cover (LULC) changes have negatively affected the climate, wildlife, and global ecosystem. By monitoring the forest changes, managers and planners can make appropriate decisions to preserve these natural areas. In this regard, this paper presents a novel deep learning-based forest change detection method including two main steps: (1) producing a new difference image enabling a more efficient distinction of changed and unchanged areas, and (2) generating a reliable forest change map by applying the recurrent residual-based U-Net deep neural network (R2U-Net) on the difference image. The recently introduced forest fused difference image (FFDI) is first improved by modifying its weighted angular operator as well as applying the fast local laplacian filter (FLLF) to generate an enhanced forest fused difference image (EFFDI). R2U-Net is subsequently used to segment the EFFDI into the changed and unchanged regions because it preserves their geometric shapes more effectively than other U-net variants. To assess the efficacy of the presented method, experimental results were conducted on four bi-temporal images acquired by the Sentinel 2 and Landsat 8 satellite sensors. The qualitative and quantitative results demonstrated the effectiveness of the proposed EFFDI in reflecting the true forest changes from the background. Moreover, compared with the other conventional U-Net-based models, including U-Net, ResU-Net, and U-Net ++, forest changes and their geometrical details were better preserved by R2U-Net. Furthermore, the proposed approach outperformed the other state-of-the-art supervised and unsupervised change detection methods in terms of quantitative and qualitative results, demonstrating its high potential for forest change detection applications.
AB - Forest changes caused by fires, clear-cuts, and Land Use/Land Cover (LULC) changes have negatively affected the climate, wildlife, and global ecosystem. By monitoring the forest changes, managers and planners can make appropriate decisions to preserve these natural areas. In this regard, this paper presents a novel deep learning-based forest change detection method including two main steps: (1) producing a new difference image enabling a more efficient distinction of changed and unchanged areas, and (2) generating a reliable forest change map by applying the recurrent residual-based U-Net deep neural network (R2U-Net) on the difference image. The recently introduced forest fused difference image (FFDI) is first improved by modifying its weighted angular operator as well as applying the fast local laplacian filter (FLLF) to generate an enhanced forest fused difference image (EFFDI). R2U-Net is subsequently used to segment the EFFDI into the changed and unchanged regions because it preserves their geometric shapes more effectively than other U-net variants. To assess the efficacy of the presented method, experimental results were conducted on four bi-temporal images acquired by the Sentinel 2 and Landsat 8 satellite sensors. The qualitative and quantitative results demonstrated the effectiveness of the proposed EFFDI in reflecting the true forest changes from the background. Moreover, compared with the other conventional U-Net-based models, including U-Net, ResU-Net, and U-Net ++, forest changes and their geometrical details were better preserved by R2U-Net. Furthermore, the proposed approach outperformed the other state-of-the-art supervised and unsupervised change detection methods in terms of quantitative and qualitative results, demonstrating its high potential for forest change detection applications.
KW - Deep learning
KW - Forest change detection
KW - Recurrent residual-based U-net (R2U-net)
KW - Remote sensing
KW - U-net
UR - http://www.scopus.com/inward/record.url?scp=85141148567&partnerID=8YFLogxK
U2 - 10.1007/s12145-022-00885-6
DO - 10.1007/s12145-022-00885-6
M3 - Article
AN - SCOPUS:85141148567
VL - 15
SP - 2335
EP - 2347
JO - Earth science informatics
JF - Earth science informatics
SN - 1865-0473
IS - 4
ER -