Dropout-based Adversarial Training Networks for Remote Sensing Scene Classification

Scene  classification  in  remote  sensing  (RS)  images  is  a  challenging  task  due  to  the  lack  of  well labeled  data.  Recently,  deep transfer  learning  (DTL)  has been  proposed to  handle  this  task.  However, most DTL methods cannot effectively deal with ambiguous features  on  class  boundaries and multi-modal  structures of  RS  data, so their performance is unsatisfactory. To handle the challenges, this letter  presents  a  novel  dropout-based  adversarial  training  network  for   RS   scene classification.   Specifically,   a   dropout-based   labelclassifier  module is  designed  to reduce  the  selection  of  ambiguous features.  Then,  a  dropout-based  domain discriminator  module  is constructed to capture  multi-modal  structures  of  RS  images so  as to achieve  fine-grained  alignment  between  cross-domain  distributions.  Third,  a  joint  distribution  of  features  and  labels  is  built  to  further  enhance  the  performance. Experiments  on  seven public  RS data sets show  that  our  model  outperforms    several  state-of-the-arts    under different conditions.