CLASS-INCREMENTAL LEARNING FOR REMOTE SENSING IMAGES BASED ON KNOWLEDGE DISTILLATION

In real-world recognition of remote sensing (RS) targets, large amount of RS data is hard to be acquired at once, but arrives in batches, which means the constantly adaption of models for new data and new classes. However, training old and new data together from scratch has certain requirements on data storage space and retraining time, so incremental learning come to be desirable for future RS recognition systems. In this article, a class incremental learning method based on knowledge distillation is proposed for RS image classification. In order to better retain the knowledge of old tasks, a relation-based loss is used as a new matching manner in distillation loss, which frees the student model from the burden of matching the exact output of the teacher model. Experiment results based on UC Merced 21, NWPU-RESISC45 and plane objects of FAIR1M demonstrate the advantages of the proposed method.