CVMFusion: ConvNeXtV2 and Visual Mamba Fusion for Remote Sensing Segmentation.

Area of Science:

• Remote Sensing
• Computer Vision
• Geographic Information Systems

Background:

• Extracting coastlines from high-resolution remote sensing imagery presents challenges due to complex details and target variability.
• Existing Convolutional Neural Networks (CNNs) struggle with long-range dependencies, while Transformers are computationally expensive.
• There is a need for efficient and accurate land-sea segmentation methods that can handle intricate details and global context.

Purpose of the Study:

• To propose CVMFusion, a novel land-sea segmentation network designed to overcome limitations of current methods.
• To integrate the strengths of CNNs for local feature extraction and Mamba for global context modeling.
• To achieve high accuracy in coastline extraction, particularly in challenging scenarios involving small objects and complex boundaries.

Main Methods:

• Developed CVMFusion, a U-shaped encoder-decoder network with hierarchical organization.
• Employed parallel ConvNeXtV2 and VMamba branches in the encoder for local and global feature capture.
• Integrated Dynamic Multi-Scale Attention (DyMSA) and Dynamic Weighted Cross-Attention (DyWCA) modules for adaptive feature fusion via skip connections.

Main Results:

• CVMFusion achieved Mean Intersection over Union (MIoU) accuracies of 98.05% and 96.28% on two public datasets.
• The proposed network outperformed existing land-sea segmentation methods.
• Demonstrated superior performance in segmenting small objects and intricate boundary regions.

Conclusions:

• CVMFusion effectively fuses local and global features for accurate coastline extraction.
• The adaptive fusion mechanisms and hierarchical structure contribute to improved segmentation performance.
• CVMFusion represents a significant advancement in land-sea segmentation for high-resolution remote sensing imagery.