Detecting Buildings and Nonbuildings from Satellite Images Using U-Net

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Automatic building detection from satellite imagery is crucial for urban planning, disaster management, and population tracking.
• Traditional image processing methods are often slow and inaccurate.
• Convolutional Neural Networks (CNNs), particularly U-Net architectures, show promise for improved building extraction.

Purpose of the Study:

• To evaluate the effectiveness of U-Net based CNNs for automatic building detection in high-resolution satellite images.
• To compare the performance of different transfer learning models (VGGNet, ResNet) integrated with U-Net.
• To develop an accurate and efficient building detection system using limited training data.

Main Methods:

• Utilized a U-Net architecture, originally designed for medical image segmentation, for building detection in RGB satellite images.
• Implemented transfer learning by adapting the encoder of the U-Net with pre-trained models: VGGNet and ResNet.
• Compared the performance of the custom U-Net, VGGNet-backboned U-Net, and ResNet-backboned U-Net.

Main Results:

• The U-Net architecture demonstrated excellent accuracy even with a limited dataset.
• VGGNet as a backbone for the U-Net achieved the highest accuracy of 84.9% in feature extraction.
• The proposed VGGNet-backboned U-Net model outperformed existing state-of-the-art models for similar tasks.

Conclusions:

• Transfer learning, specifically using VGGNet with U-Net, significantly enhances building detection accuracy from satellite imagery.
• The developed method offers a superior and efficient solution for automatic building extraction compared to current approaches.
• This technology has strong implications for improving civic planning, disaster response, and map updates.