A method for extracting buildings from remote sensing images based on 3DJA-UNet3

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Building extraction from remote sensing imagery is crucial for urban planning and monitoring.
• Existing UNet3+ models face challenges with segmentation accuracy and boundary precision.

Purpose of the Study:

• To enhance building extraction accuracy and efficiency using a novel module for the UNet3+ model.
• To improve feature representation and reduce network complexity.

Main Methods:

• A 3D Joint Attention (3DJA) module was developed to improve feature correlation (local/global, channel).
• A bottleneck module was incorporated to reduce parameters and enhance training efficiency.
• Experiments were conducted on WHU, INRIA, and Massachusetts building datasets.

Main Results:

• The proposed 3DJA-UNet3+ model demonstrated competitive performance.
• The model achieved strong results in overall accuracy, mean intersection over union, and F1-score.
• The 3DJA module effectively enhanced semantic information and feature representation.

Conclusions:

• The 3DJA-UNet3+ model offers an effective solution for building extraction from remote sensing data.
• The proposed module improves segmentation accuracy and model efficiency.
• The approach provides a valuable contribution to urban planning and monitoring applications.