SwinClustering: a new paradigm for landscape character assessment through visual segmentation
The research value of Landscape Character Assessment (LCA) lies in gaining a deeper understanding of the inherent attributes and interrelationships of various landscapes, thereby providing scientific basis for landscape planning, design, conservation, and sustainable utilization. The traditional LCA...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Frontiers Media S.A.
2025-02-01
|
| Series: | Frontiers in Environmental Science |
| Subjects: | |
| Online Access: | https://www.frontiersin.org/articles/10.3389/fenvs.2025.1509113/full |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | The research value of Landscape Character Assessment (LCA) lies in gaining a deeper understanding of the inherent attributes and interrelationships of various landscapes, thereby providing scientific basis for landscape planning, design, conservation, and sustainable utilization. The traditional LCA methods often overlook the inherent connections between various landscape attributes and geographical spatial relationships among data points, which restricts their application in sustainable multi-scale landscape element assessments. Accordingly, this paper proposes a new paradigm for LCA, SwinClustering, built upon the cutting-edge Swin Transformer architecture. This approach employs a visual segmentation method to achieve multi-scale clustering, utilizing nine key attributes of landscape elements: altitude, aspect, geology, landcover, landform, relief, slope, soil, and vegetation. By extracting semantic features through the GIS-aware Swin Transformer backbone network and leveraging the Feature Pyramid Decoder for segmentation clustering, SwinClustering offers a comprehensive analysis of landscape characteristics. Furthermore, we design a specific training strategy that enables coarseness and fineness control of the clustering results. SwinClustering is tested across three distinct scales: the national scale of China, the municipal scale of Beijing Municipality and the district scale of Wuyishan National Park. These experiments yield promising results, validating the method’s effectiveness across diverse geographic scales. Crucially, the proposed SwinClustering paradigm establishes a unified clustering framework to deeply learn the intrinsic connection between various landscape attributes and the spatial relationship between different geographic locations. Furthermore, its strong generalization capabilities enable its seamless application to LCA tasks at arbitrary scales, marking a sustainable development in the field of LCA. |
|---|---|
| ISSN: | 2296-665X |