Self-Supervised Learning and Multi-Sensor Fusion for Alpine Wetland Vegetation Mapping: Bayinbuluke, China

Area of Science:

• Ecology
• Remote Sensing
• Machine Learning

Background:

• Accurate wetland vegetation mapping is crucial for ecological monitoring and conservation.
• Challenges include spatial heterogeneity, limited ground-truth data, and invasive species spread.
• Early detection of invasive plants is vital for preserving ecosystem integrity.

Purpose of the Study:

• To develop a novel framework for enhanced vegetation classification in alpine wetlands.
• To integrate self-supervised learning (SSL), supervised segmentation, and multi-sensor data fusion.
• To improve the early detection of invasive species and mapping of rare vegetation.

Main Methods:

• Utilized high-resolution satellite imagery (PlanetScope-3, Jilin-1) and fused multi-sensor data.
• Employed self-supervised learning (SSL) methods (BYOL, DINO, MoCo v3) for feature representation learning.
• Integrated SSL with supervised segmentation for vegetation classification.

Main Results:

• Self-supervised learning (SSL) methods showed varied classification performance.
• Multi-sensor data fusion significantly improved the detection of rare and fragmented vegetation.
• The framework enabled early identification of invasive species in the Bayinbuluke Alpine Wetland.

Conclusions:

• The proposed SSL-fusion strategy offers a scalable, high-precision solution for wetland monitoring.
• This approach reduces reliance on labor-intensive field data collection.
• It effectively supports invasive species management and ecological conservation efforts.