Building rooftop extraction from high resolution aerial images using multiscale global perceptron with spatial...

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Building rooftop extraction is crucial for cartography, urban planning, and intelligent city development.
• High-resolution aerial imagery aids automatic building detection, but challenges like scale variation and complex shapes persist.
• Existing deep learning methods (CNNs, Transformers) struggle with fragmentation and detail loss in rooftop extraction.

Purpose of the Study:

• To develop an advanced deep learning model for precise building rooftop extraction.
• To enhance contextual representation and restore spatial details in building extraction from aerial images.
• To overcome limitations of current methods in handling scale variation and complex building geometries.

Main Methods:

• Developed a multi-scale global perceptron network integrating Transformer and Convolutional Neural Networks (CNNs).
• Employed an improved multi-head-attention encoder with multi-scale tokens for enhanced global semantic correlations.
• Designed a context refinement decoder utilizing high-level semantics and shallow features to restore spatial details.

Main Results:

• The proposed model achieved a 95.18% F1 score on the WHU dataset and 93.29% on the Massub dataset.
• Quantitative analysis and visual experiments demonstrated superior performance compared to state-of-the-art methods.
• The network effectively addressed issues of fragmentation and lack of spatial detail in building rooftop extraction.

Conclusions:

• The multi-scale global perceptron network offers an efficient and superior solution for building rooftop extraction.
• The novel encoder-decoder architecture enhances feature extraction and detail restoration.
• This method significantly advances automated building detection in high-resolution aerial imagery.