Robust Building Extraction for High Spatial Resolution Remote Sensing Images with Self-Attention Network.

Area of Science:

• Computer Vision
• Remote Sensing
• Deep Learning
• Semantic Segmentation

Background:

• Building extraction from high-resolution remote sensing imagery is crucial for urban planning and geographic information systems.
• Existing methods often struggle with capturing both local and global contextual information for accurate building delineation.

Purpose of the Study:

• To introduce a novel semantic segmentation model, the Pyramid Self-Attention Network (PISANet), for precise building extraction.
• To develop an efficient and simple end-to-end network capable of learning comprehensive building features.

Main Methods:

• PISANet utilizes a two-part structure: a backbone for local feature extraction and a pyramid self-attention module for global and comprehensive feature learning.
• The model employs a supervised, end-to-end learning approach, taking remote sensing images and labels as input to generate building probability maps.

Main Results:

• PISANet achieved high performance on two datasets, with overall accuracy reaching up to 96.15%.
• The model demonstrated superior intersection-over-union (IoU) scores (up to 87.97%) and F1 indices (up to 93.55%).
• Experiments confirmed PISANet's ability to maintain high accuracy, reduce errors, and improve the integrity of extracted buildings.

Conclusions:

• The proposed PISANet effectively extracts buildings from high-resolution remote sensing images by integrating local and global contextual information.
• The network's simplified yet powerful architecture facilitates implementation and achieves state-of-the-art results in building segmentation.