Unsupervised Domain Adaptive Transfer Learning for Urban Built-Up Area Extraction

Built-up areas are the main gathering place for human activities. The widespread availability of various satellite sensors provides a rich data source for mapping built-up areas. Deep learning can automatically learn multi-level features of targets from sample data in an end-to-end manner, overcoming the limitations of traditional methods based on handcrafted features. However, existing deep-learning-based methods rely on the quantity and distribution of sample data, and the trained models often exhibit limited generalization ability when faced with image data from novel scenarios. To effectively tackle this issue, this study proposes an unsupervised domain adaptive transfer learning method based on adversarial machine learning. This method aims to utilize the feature information of the source domain to train a classifier suitable for target domain feature discrimination without requiring a target domain label, and achieve built-up area extraction of different sensor images. The model comprises a feature extraction module, a label classification module, and a domain discrimination module. Through adversarial training, the feature knowledge from the source domain is transferred to the target domain, achieving feature alignment and efficient discrimination of built-up areas. The Gaofen-2 (GF-2) and Sentinel-2 datasets were employed for experimental evaluation. The results show that the proposed method, trained on the GF-2 image dataset (source domain), can be transferred unsupervised to the Sentinel-2 image dataset (target domain), demonstrating robust detection performance. Further comparative experiments have also demonstrated the superiority of our method in extracting built-up areas through transfer learning.