Revealing the competing mechanisms between land use development and restoration in wetland evolution : an interpretable deep learning approach
Zhang, Zixia ; Zhou, Shenbei ; Duan, Yeqing LU
; Hou, Jiaping
and Ning, Jing
(2026)
In Landscape Ecology
41(8).
- Abstract
Context: The spatial competitive relationship between wetland restoration and productive land expansion represents a core challenge in regional land use sustainability. Objectives: This study aims to develop a deep learning framework capable of jointly capturing the spatiotemporal characteristics of land competition. It further seeks to understand changes in the competitive relationship between wetlands and productive land by combining predicted probability distributions with differences in feature contributions. Methods: Using multi-source spatial data for the Yellow River Delta during 2001–2020, this study developed a CNN-LSTM spatiotemporal deep learning framework coupled with a CA-based neighborhood mechanism. Gini impurity and... (More)
Context: The spatial competitive relationship between wetland restoration and productive land expansion represents a core challenge in regional land use sustainability. Objectives: This study aims to develop a deep learning framework capable of jointly capturing the spatiotemporal characteristics of land competition. It further seeks to understand changes in the competitive relationship between wetlands and productive land by combining predicted probability distributions with differences in feature contributions. Methods: Using multi-source spatial data for the Yellow River Delta during 2001–2020, this study developed a CNN-LSTM spatiotemporal deep learning framework coupled with a CA-based neighborhood mechanism. Gini impurity and Integrated Gradients (IG) were introduced to analyze potential competition zones and differences in feature contributions to prediction results. Results: The CNN-LSTM model achieved the best overall performance, with an FoM of 0.1534. The case results showed an overall trade-off pattern between wetlands and productive land. Higher Gini impurity values tended to occur in areas with more interspersed land use and more complex spatial patterns. IG results indicated different patterns of feature dependence in model discrimination. Wetlands were related to natural conditions and historical states, whereas productive land depended more on neighborhood structure and spatial agglomeration, with temporal variation in feature contributions. Conclusions: Competition between wetlands and productive land is neither spatially uniform nor temporally static. It is more concentrated in complex transitional zones and shows different characteristics across stages. The framework developed in this study provides a methodological pathway for research on complex land systems in landscape ecology.
(Less)
- author
- Zhang, Zixia
; Zhou, Shenbei
; Duan, Yeqing
LU
; Hou, Jiaping
and Ning, Jing
- organization
- publishing date
- 2026-08
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Cellular Automata, Gini impurity, Land competition, Model interpretability, Spatiotemporal deep learning, Wetland restoration
- in
- Landscape Ecology
- volume
- 41
- issue
- 8
- article number
- 143
- publisher
- Springer
- external identifiers
-
- scopus:105046754211
- ISSN
- 0921-2973
- DOI
- 10.1007/s10980-026-02384-1
- language
- English
- LU publication?
- yes
- additional info
- Publisher Copyright: © The Author(s) 2026.
- id
- 7999376d-0b63-45b6-8b05-ee2fe5cd4dab
- date added to LUP
- 2026-09-14 14:40:07
- date last changed
- 2026-09-14 14:41:01
@article{7999376d-0b63-45b6-8b05-ee2fe5cd4dab,
abstract = {{<p>Context: The spatial competitive relationship between wetland restoration and productive land expansion represents a core challenge in regional land use sustainability. Objectives: This study aims to develop a deep learning framework capable of jointly capturing the spatiotemporal characteristics of land competition. It further seeks to understand changes in the competitive relationship between wetlands and productive land by combining predicted probability distributions with differences in feature contributions. Methods: Using multi-source spatial data for the Yellow River Delta during 2001–2020, this study developed a CNN-LSTM spatiotemporal deep learning framework coupled with a CA-based neighborhood mechanism. Gini impurity and Integrated Gradients (IG) were introduced to analyze potential competition zones and differences in feature contributions to prediction results. Results: The CNN-LSTM model achieved the best overall performance, with an FoM of 0.1534. The case results showed an overall trade-off pattern between wetlands and productive land. Higher Gini impurity values tended to occur in areas with more interspersed land use and more complex spatial patterns. IG results indicated different patterns of feature dependence in model discrimination. Wetlands were related to natural conditions and historical states, whereas productive land depended more on neighborhood structure and spatial agglomeration, with temporal variation in feature contributions. Conclusions: Competition between wetlands and productive land is neither spatially uniform nor temporally static. It is more concentrated in complex transitional zones and shows different characteristics across stages. The framework developed in this study provides a methodological pathway for research on complex land systems in landscape ecology.</p>}},
author = {{Zhang, Zixia and Zhou, Shenbei and Duan, Yeqing and Hou, Jiaping and Ning, Jing}},
issn = {{0921-2973}},
keywords = {{Cellular Automata; Gini impurity; Land competition; Model interpretability; Spatiotemporal deep learning; Wetland restoration}},
language = {{eng}},
number = {{8}},
publisher = {{Springer}},
series = {{Landscape Ecology}},
title = {{Revealing the competing mechanisms between land use development and restoration in wetland evolution : an interpretable deep learning approach}},
url = {{http://dx.doi.org/10.1007/s10980-026-02384-1}},
doi = {{10.1007/s10980-026-02384-1}},
volume = {{41}},
year = {{2026}},
}