SENSOR ADAPTATION NETWORK FOR BUILDING FOOTPRINT EXTRACTION USING SAR IMAGES

Building footprint extraction is critical for urban applications. The effectiveness of optical images depends heavily on cloud cover and daylight, which limits their utility in time-sensitive tasks. Synthetic aperture radar (SAR) data has the advantage of all-weather observation, but accurately interpreting them is challenging. Recently, several studies have attempted to transfer knowledge from optical to SAR data to enhance all-weather observation ability. However, existing knowledge distillation (KD)-based methods may fail to effectively extract shared and specific information. In this study, a multitask learning-based sensor adaptation network has been proposed for SAR building footprint extraction to efficiently encode shared and specific information, thereby enhancing the representational capacity of multisensor features. The sensor adaptation network comprises a sensor adaptation encoder to construct optical and SAR encoders focusing on shared and specific information. Then, the online KD scheme is employed to align the distribution of the SAR branch with the optical branch, thereby updating SAR-specific operations and decoders. Therefore, the proposed sensor adaptation network can efficiently encode shared and specific information to enhance the representational capacity of multisensor features. Comparative experiments demonstrate the effectiveness of the proposed sensor adaptation network for SAR building footprint extraction.