RSSFormer: Foreground Saliency Enhancement for Remote Sensing Land-Cover Segmentation.

Area of Science:

• Computer Science
• Remote Sensing
• Geospatial Analysis

Background:

• High spatial resolution (HSR) remote sensing images present unique challenges for semantic segmentation due to complex foreground-background relationships.
• Issues include large-scale variations, intricate background samples, and imbalanced foreground-background distribution, hindering current context modeling approaches.

Purpose of the Study:

• To develop an advanced remote sensing segmentation framework (RSSFormer) that effectively models foreground saliency.
• To improve land cover segmentation accuracy in HSR remote sensing imagery by addressing existing challenges.

Main Methods:

• Proposed RSSFormer framework incorporating an Adaptive Transformer Fusion Module for adaptive feature fusion and background noise suppression.
• Implemented a Detail-aware Attention Layer to extract foreground-specific details using spatial and channel attention mechanisms.
• Introduced a Foreground Saliency Guided Loss function to focus optimization on challenging samples with low foreground saliency.

Main Results:

• RSSFormer demonstrated superior performance compared to existing general semantic and remote sensing segmentation methods on LoveDA, Vaihingen, Potsdam, and iSAID datasets.
• The method achieved a favorable balance between computational efficiency and segmentation accuracy.
• Experimental validation confirmed the effectiveness of the proposed modules in enhancing foreground saliency.

Conclusions:

• RSSFormer offers a robust solution for remote sensing land cover segmentation, particularly for HSR imagery.
• The framework's foreground saliency modeling capabilities significantly improve segmentation outcomes.
• The proposed approach provides a valuable advancement in geospatial analysis and remote sensing applications.