Water Body Extraction from Multi-temporal Polarimetric SAR Images Based on Transfer Learning

  Objectives  Machine learning classifier-based water body extraction methods for polarimetric synthetic aperture radar (PolSAR) images have high reliability but typically require a great number of training samples. Consequently, it is very difficult and time-consuming to manually collect enough training samples when extracting water body from multi-temporal PolSAR images. To this problem, transfer learning is used to reduce the labor cost of querying new samples and improve the timeliness of water body extraction of multi-temporal PolSAR images.

  Methods  Firstly, an optimal source domain image from multi-temporal images is automatically selected according to the distribution difference between images, and the other images are taken as target domain images. Secondly, a group of training samples are queried in the source domain image as source sample set, and the same number of unlabeled samples are randomly sampled from each target domain image as their target domain sample set. And the knowledge of source domain samples is transferred to target domain samples via the transfer learning method. Finally, a random forest classifier-based water body extraction model is trained using the target domain sample set, and is used for the water body extraction of target domain images.

  Results  We have conducted experiments using six PolSAR images and two kinds of transfer learning methods, the results show that: (1) The label transfer accuracy and the water body extraction accuracy are positively correlated. (2) Inductive transfer learning methods achieve higher label transfer accuracy and lower standard deviation. (3) A smaller distribution difference between source and target domain images indicate a greater transferability, and thus a better water body extraction accuracy. (4) The water body extraction results of inductive transfer learning methods have a higher rate of missing detection, while the results of transductive transfer learning methods have a higher rate of false detection.

  Conclusions  In the water body extraction of multi-temporal PolSAR images, the use of transfer learning methods can significantly reduce the number of manually labeled samples needed to construct high-performance classifiers, while maintaining the water body extraction accuracy at a high level. It has great application potentiality in the emergency response of flood disaster.