An Improved Segformer for Semantic Segmentation of UAV-Based Mine Restoration Scenes

Area of Science:

• Environmental Science
• Remote Sensing
• Computer Vision

Background:

• Mine ecological restoration is vital for sustainable development in resource-dependent regions.
• Current UAV-based remote sensing monitoring methods lack accuracy and adaptability.
• Challenges include small-object recognition, multi-scale feature fusion, and blurred boundaries.

Purpose of the Study:

• To develop an enhanced semantic segmentation model for improved mine ecological restoration monitoring.
• To address limitations in existing UAV imagery analysis for mine sites.

Main Methods:

• Proposed an enhanced Segformer model incorporating a multi-scale feature-enhanced feature pyramid network (MSFE-FPN).
• Integrated a selective feature aggregation pyramid pooling module (SFA-PPM) for global semantic perception.
• Embedded an efficient local attention (ELA) module to improve edge and small-target sensitivity.
• Utilized the HUNAN Mine UAV Dataset (HNMUD) and the Aeroscapes dataset for evaluation.

Main Results:

• The proposed model demonstrated superior segmentation accuracy and generalization ability.
• Effectively improved recognition of small objects and enhanced boundary definition.
• Outperformed existing methods in analyzing mine restoration scenes.

Conclusions:

• The enhanced Segformer model provides a robust solution for mine ecological restoration monitoring.
• Offers significant advancements in image analysis for environmental management.
• Supports sustainable development through accurate remote sensing applications.