Multimodal Co-learning: A Domain Adaptation Method for Building Extraction from Optical Remote Sensing Imagery