Adversarial Shape Learning for Building Extraction in VHR Remote Sensing Images.

Area of Science:

• Computer Vision
• Remote Sensing
• Artificial Intelligence

Background:

• Building extraction from very high-resolution remote sensing images (VHR RSIs) is challenging due to occlusion and boundary ambiguity.
• Conventional convolutional neural networks (CNNs) struggle to capture essential building shape patterns crucial for accurate recognition.

Purpose of the Study:

• To propose an adversarial shape learning network (ASLNet) for improved building segmentation in VHR RSIs.
• To explicitly model building shape patterns to overcome limitations of existing methods.

Main Methods:

• Introduction of an adversarial learning strategy within the ASLNet to enforce shape constraints.
• Integration of a CNN shape regularizer to enhance the representation of shape features.
• Utilizing object-based quality assessment metrics to evaluate geometric accuracy.

Main Results:

• The proposed ASLNet significantly improves pixel-based accuracy in building segmentation.
• ASLNet demonstrates substantial enhancements in object-based quality measurements compared to baseline methods.
• Experimental validation on two benchmark datasets confirms the effectiveness of the proposed approach.

Conclusions:

• The ASLNet effectively addresses shape modeling challenges in building extraction from VHR RSIs.
• The integration of adversarial learning and shape regularization leads to superior segmentation performance.
• The method offers a promising advancement for accurate and geometrically precise building segmentation in remote sensing applications.