EVALUATION OF SEVERAL FULLY CONVOLUTIONAL NETWORKS IN SAR IMAGE CHANGE DETECTION

In recent years, the world is suffering from frequent natural disasters. Change detection (CD) technology can quickly identify the change information on the ground and has developed into an important means of disaster monitoring and assessment. Synthetic aperture radar (SAR) has the characteristics of periodic observation and wide coverage. Moreover, SAR has the advantages of penetrating, all-day and all-weather observation, which plays an important role in disaster monitoring. Due to the rapid development of satellite sensors, the available CD data has been greatly enriched. This situation provides an opportunity for deep learning change detection (DLCD) techniques. However, SAR data are affected by speckle noise and lack of available labeled samples, it remains challenging to precisely locate the change information with high efficiency. This paper focuses on several commonly used and outstanding networks in the DLCD field to evaluate their performance and develop them to SAR data. In addition, Transfer learning experiments are designed to evaluate the generalization performance of each network for the CD task. The experimental results show that the Siamese CD network encoding multi-temporal data separately has the best ability to detect changes and generalization performance. In addition, adding high quality explicit difference guidance information to the network is more specific for the CD task, which can further improve network performance and refine the boundaries of changed ground objects on change map.