SegFormer-based segmentation approach for landscape planning and overhead remote sensing image analysis.

Area of Science:

• Computer Vision
• Geospatial Analysis
• Remote Sensing

Background:

• Accurate landscape element analysis is crucial for spatial information extraction.
• Existing semantic segmentation models face challenges in precision, structural stability, and computational efficiency.

Purpose of the Study:

• To develop a semantic segmentation framework for landscape element analysis.
• To achieve high precision, structural stability, and computational efficiency in spatial information extraction.

Main Methods:

• Utilized a Transformer architecture with SegFormer as the core.
• Incorporated a Mix Transformer (MiT) encoder for multi-scale features and an MLP decoder for semantic fusion.
• Introduced Edge-aware Auxiliary Supervision and a Unified Label Mapping strategy for improved structural learning and generalization.

Main Results:

• Achieved a Mean Intersection over Union (mIoU) of 76.5% on the LoveDA dataset, a 5.3% improvement over DeepLabV3+.
• Demonstrated superior performance in Boundary F1-score for structurally sensitive categories.
• Attained an inference speed of 42.1 Frames Per Second (FPS) without significant efficiency loss.

Conclusions:

• The proposed framework offers a reproducible and engineering-feasible solution for structured spatial representation.
• The framework effectively improves landscape element analysis for remote sensing and geospatial information analysis.
• Performance gains are attributed to structure-adaptive workflows and planning-oriented metrics, not just network scaling.