Deep Layered Network Based on Rotation Operation and Residual Transform for Building Segmentation from Remote Sensing...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• High-resolution remote sensing (HRS) images present challenges in building segmentation due to limitations in capturing both detailed and global information.
• Existing deep learning models often struggle with insufficient feature representation for accurate building extraction from HRS data.

Purpose of the Study:

• To propose a novel deep learning model, C_ASegformer, for enhanced building segmentation in HRS images.
• To improve the integration of hierarchical and contextual information for more robust feature representation.

Main Methods:

• Developed a Deep Layered Enhanced Fusion (DLEF) module to integrate hierarchical information from diverse receptive fields.
• Introduced a Triplet Attention (TA) Module to capture inter-dependencies between buildings and their environment.
• Designed a Multi-Level Dilated Connection (MDC) Module for efficient multi-scale contextual relationship capture.

Main Results:

• C_ASegformer achieved high performance on benchmark datasets, including the Massachusetts dataset.
• Achieved 95.42% Overall Accuracy (OA), 85.69% F1-score, and 75.46% mean Intersection over Union (mIoU) on the Massachusetts dataset.
• Demonstrated superior accuracy compared to state-of-the-art models across multiple datasets.

Conclusions:

• The proposed C_ASegformer model effectively addresses the limitations of current methods for building segmentation in HRS images.
• The novel modules (DLEF, TA, MDC) significantly enhance feature representation and contextual understanding, leading to improved segmentation accuracy.
• The model's performance validates its sophistication and effectiveness for detailed building extraction from remote sensing data.